Google Earth Engine and Artificial Intelligence (AI): A Comprehensive Review

Abstract

Remote sensing (RS) plays an important role gathering data in many critical domains (e.g., global climate change, risk assessment and vulnerability reduction of natural hazards, resilience of ecosystems, and urban planning). Retrieving, managing, and analyzing large amounts of RS imagery poses substantial challenges. Google Earth Engine (GEE) provides a scalable, cloud-based, geospatial retrieval and processing platform. GEE also provides access to the vast majority of freely available, public, multi-temporal RS data and offers free cloud-based computational power for geospatial data analysis. Artificial intelligence (AI) methods are a critical enabling technology to automating the interpretation of RS imagery, particularly on object-based domains, so the integration of AI methods into GEE represents a promising path towards operationalizing automated RS-based monitoring programs. In this article, we provide a systematic review of relevant literature to identify recent research that incorporates AI methods in GEE. We then discuss some of the major challenges of integrating GEE and AI and identify several priorities for future research. We developed an interactive web application designed to allow readers to intuitively and dynamically review the publications included in this literature review.

1. Introduction and Motivation

Big data approaches have been making substantial changes in science and in society at large [1,2]. Geospatial big data, which are collected with ubiquitous location-aware sensors that are inherently geospatial [3], are a significant portion of big data. The size of such data is growing rapidly, by at least 20% per year [4]. The United Nations Initiative on Global Geospatial Information Management (UN-GGIM) estimated that 2.5 quintillion bytes of data (one quintillion bytes = 1000 petabytes (PB); 1 PB = 1000 Terabytes (TB)) are being generated every single day, a large portion of which is location-aware. About 25 PB of data are being generated per day at Google, a significant portion of which is spatio-temporal data [4]. This trend will accelerate even faster as the world becomes more mobile and as unoccupied aircraft systems (UAS) and satellite imagery are acquired more often and at higher resolutions [5]. Along with this exponential increase in geospatial big data, the need for cloud computing and high-performance computing for modeling, analyzing, and simulating geospatial contents is also rapidly increasing [4]. Geospatial big data have recently gained attention from researchers and practitioners in geographic information science (GIScience) and remote sensing (RS) [6]. Efficient collection, management, storage, analysis, and visualization of big data have become critical for the development of intelligent decision systems and provide unprecedented opportunities for business, science, and engineering [7]. Handling the 5 “Vs” (volume, variety, velocity, veracity, and value [8]) of big data is still a very challenging task. This is even more challenging for RS imagery due to its large volume (i.e., high resolution and multiple bands) and long timespan; geospatial big data pose significant challenges to conventional geographic information systems (GIS) as well as RS approaches and platforms [9,10,11,12,13].

Geospatial big data, especially RS big data, have posed substantial challenges due to their large volume, high spatial-temporal resolution, and complexity. One of the very promising and practical solutions for analyzing RS big data is Google Earth Engine (GEE). GEE is a scalable, cloud-based geospatial retrieval and processing platform. It also provides access to the vast majority of freely available, public, multi-temporal RS data and offers free cloud-based computational power for geospatial data analysis [14,15,16]. More specifically, GEE provides free access to a multi-PB archive of geospatial datasets spanning over 40 years of historical and current Earth observation (EO) imagery, including satellite imagery (e.g., Sentinel from the European Space Agency (ESA), Landsat from the United States Geological Survey (USGS), Moderate Resolution Imaging Spectroradiometer (MODIS) from the National Aeronautics and Space Administration (NASA), the Cropland Data Layer (CDL) from the United States Department of Agriculture’s (USDA) and National Agricultural Statistics Service (NASS), and the National Agriculture Imagery Program (NAIP), also from the USDA), airborne imagery, weather and climate datasets, as well as digital elevation models (DEMs) [14,16]. Those RS data can be efficiently imported and processed on the cloud platform, avoiding the need to download data to local computers for processing [17]. Along with computing and storage resources, GEE also supports many RS algorithms (e.g., image enhancement, image classification, and cloud masking), which are readily accessible and customizable and allow data processing and visualization at different scales through JavaScript or Python Application Program Interfaces (APIs) [14,16,18,19]. These capabilities reduce most of the time-consuming preprocessing steps needed in traditional RS approaches. The computational power of GEE along with its comprehensive data catalog and data processing methods make GEE an ideal platform for solving geospatial big data problems. GEE allows researchers and practitioners to focus on developing and solving their domain problems by making it easier to retrieve data and algorithms and to compute all in one place. For example, the Landsat archive on GEE is already preprocessed for atmospheric and topographic effects—this saves researchers and practitioners a substantial amount of time and effort in terms of downloading and preprocessing data [16]. GEE, with free planetary-scale geospatial big data (solved the data availability, data storage, and data preprocessing challenges) and free computing resources, facilitates computationally cumbersome geospatial big data analysis for researchers and practitioners with minimal local computing and storage resources. GEE, in the parlance of the RS Communication Model, reduces the number of channels required to construct an RS system, and therefore the time required to go from query to result [20]. Researchers from a wide range of fields are able to generate multiscale (local, national, regional, continental, and global scale) insights that would have been nearly impossible without the geospatial big data and computing capacity available in GEE [21].

GEE provides the free cloud-computing platform to tackle geospatial big data challenges, and recent substantial advances in artificial intelligence (AI) can and will further elevate the power of GEE. We cover three of AI’s main subdisciplines in this paper: computer vision (CV), machine learning (ML) and its subdomain, deep learning (DL). These technologies are central to leveraging big data for applications in many domains and have achieved significant advances in a wide range of applications that have a high social impact, such as damage assessment and prediction of natural disasters (e.g., automatic flooding damage assessment [1] and wildfire prediction [22]) and healthcare [23,24,25]. Geospatial artificial intelligence (GeoAI) combines methods in spatial science (e.g., GIScience and RS), AI, data mining, and high-performance computing to extract meaningful knowledge from geospatial big data [26]. GeoAI stems from GIScience methods applied to RS data but has advanced the field of AI to solve geospatial-specific big data challenges and problems.

There are substantial separate bodies of research covering AI (especially CV, ML and DL) and GEE. However, much less research directly combines AI and GEE. Allowing researchers and practitioners to harness the power of both GEE and AI for their research and real-world problems is the core motivation for us to investigate a range of recent developments that combine GEE and AI. Thus, our paper can serve as an academic bridge for researchers and practitioners in GEE and AI, highlighting how scientists are using GEE and AI and in which domain areas. Researchers and practitioners in GEE and AI will gain strength from each other and thus make the science move forward more effectively and efficiently, making it possible to tackle global challenges such as those relevant to climate change.

1.1. Selection Criterion for Reviewed Papers and Brief Graphic Summary

There is a substantial body of work on GEE (e.g., see recent reviews [14,18,27,28,29]) and AI for RS (especially DL, ML and CV used in a RS setting, see recent reviews in [30,31]), respectively. However, much less research has gone into detailing the integration of GEE with AI. In the literature review process, we initially identified 500+ papers relevant to GEE. We then performed a systematic search based on the following strategies: (1) keyword search on Google Scholar: the keywords used for our literature search are “Google Earth Engine” AND “machine learning” OR “deep learning” OR “computer vision”; (2) references tracking: we went through the papers cited in recent GEE reviews ([14,18,27,28,29]) (i.e., the “References” list of the papers) and also tracked the last two years’ worth of new papers citing the existing GEE review papers on their Google Scholar page. Note that our search was restricted to research articles published in English and in peer-reviewed journals or conference proceedings. A total number of 200 highly relevant articles were identified by excluding the papers that purely use GEE for RS data download or those that do not use AI (including its branches CV, ML, DL). Figure 1 shows the spatial distribution and statistics summary of the papers covered in this review. The number of published papers by year (2015 to 2022) has dramatically increased since 2019. “Remote Sensing” and “Remote Sensing of Environment’’ are the leading journals where most GEE and AI papers are published. In addition, most first authors institutions are based in China and the United States. (Note that a freely accessible interactive version of the map and all charts throughout the paper can be accessed via our web app tool; the web app tool URL and its brief demo video are provided in Appendix A).

1.2. Roadmap

Here, we provide a roadmap to the rest of the paper. Section 2 outlines the scope of this review and our intended audience. Section 3 is the core of the paper, focused on identifying important and recent developments and their implications in terms of applications (Section 3.2) and novel methods (Section 3.3) that leverage GEE and AI. Section 3 covers a wide array of recent research combining GEE and AI from multiple domains with many cross-connections. The paper concludes in Section 4 with a discussion of key challenges and opportunities, from both application (Section 4.2) and technical (Section 4.3) perspectives. Specifically, we focus on the main challenges preventing GEE and AI integration, as well as some possible future research directions. To make the substantial number of papers we reviewed (200 total) more transparent and easier to retrieve and understand, we developed an interactive web tool (see Appendix A for details). As evaluation metrics are essential for measuring the performance of AI/ML/DL/CV models, we provide a set of commonly used evaluation metrics in Appendix B. To make the main text of the paper concise, each application area detailed in Section 3.2 contains a table and brief textual summary of the papers in that field. However, a more detailed and comprehensive summary for each section can be found in Appendix C for those that are interested. Lastly, as there are plenty of acronyms in this paper, we provide a full list of abbreviations right before the Appendix A, Appendix B and Appendix C.

2. Scope and Intended Audience

It is very challenging to repeatedly produce up-to-date and accurate maps and obtain up-to-date and accurate information, especially at large scales, for many important applications and monitoring systems due to time, effort, and cost. As larger volumes of geospatial data become available, an ever-increasing number of techniques for analyzing them have increased the number and scope of monitoring applications (e.g., global water mapping [32], forest and deforest monitoring [33], and global climate change research [34]).

Downloading, analyzing, and managing a multi-decadal time series of satellite imagery over large areas is not practical using desktop computing resources [35]. Complementing huge volumes of open-access satellite data are new technologies and services (e.g., cloud computing, AI) that are shifting the manner in which RS data are used for environmental monitoring. The technologies and platforms present opportunities for new advances in data collection for monitoring climate change mitigation, particularly where traditional means of data exploration and analysis, such as government-led statistical census efforts, are costly and time consuming [36]. GEE [37], launched on 2 December 2010, receives significant attention in the earth science community because it provides free and dedicated geospatial data resources and services. These services include RS imagery storage, preprocessing routines, and hosting AI algorithms all in one place. Similar cloud computing platforms and services that support geospatial data hosting and/or computation include the NASA Earth Exchange (NEX) [38] (2013) and Geostationary-NEX (GeoNEX) [39] (2020), Earth on Amazon Web Services (AWS) [40] (launched in September 2016), Microsoft’s Azure services (Geospatial Analytics in Azure Stream Analytics) [41] (launched in 2017) and Microsoft Planetary Computer [42] (launched in December 2020). GEE has a bigger community, more data, and more algorithms all in one place whereas these other systems came later and are not as robust. Building on the emerging needs of GEE and AI to process and analyze big RS data via cloud computing for many domain problems, like any review paper, one of our major goals is to survey recent work on GEE with AI to provide suggestions for new directions built upon these evolving methods. Another important goal of this paper is to provide a bridge between GEE and AI researchers and practitioners, especially those who have interdisciplinary backgrounds and expertise. It is our hope that this paper helps move towards a smoother and deeper integration of GEE with AI. As our comprehensive investigation shows, there are still several open challenges preventing researchers from using GEE and AI for their research (for example, the lack of options for those interested in using DL models on the platform, detailed further in Section 4.1.2).

This comprehensive review is relevant to any research, practice, and education domains that could take advantage of RS imagery coupled with AI and cloud computing, including, but not limited to RS, GIScience, earth science, geosciences, computer science, data science, information science, hydroinformatics, and image analysis. This paper does not attempt to review publications that use GEE without utilizing AI methods (for recent reviews of GEE, see [14,18,28,29]), nor to review DL with RS (recent reviews on DL with RS can be found in [30,31,43,44]). This review focuses on investigating recent GEE work that has integrated with AI, including its branches ML, DL, and CV, for a wide and various range of applications (e.g., crop, wetland, and water mapping, detailed in Section 3.2). From our review, only a small subset of papers contribute significantly towards implementing novel AI architecture or methods within GEE (detailed in Section 3.3).

To our knowledge, no review solely focuses on the combination of GEE and AI. Our review has a narrowed scope of GEE integrated with AI, but a wider and deeper scope in terms of AI methods and metrics applied to many domains on the GEE platform. Other GEE review papers have sections for ML models (“X papers use random forest (RF) models and Y use Support Vector Machines (SVMs) …”) as part of a general GEE review, but do not explicitly focus on the implementation of AI models in GEE. In addition, what significantly distinguishes our review from other GEE survey papers is an interactive web app, named iLit4GEE-AI (see Appendix A). We developed this app to allow our readers to intuitively retrieve relevant GEE with AI literature that fits their needs. For example, a user can very quickly filter out those published articles that used RF models and F1-score for wetland mapping. Most importantly, iLit4GEE-AI will serve as a live and interactive literature repository for integrating GEE with AI, as we will continue to update the data in the web app. In the future, we hope our web app will serve as an important and up-to-date resource for the GEE and AI research and practice community. Through our deep, thorough, and interactive investigation (see Appendix A for a visual, interactive investigation using our web app iLit4GEE-AI), we hope to develop a basis for a smoother and deeper integration of GEE and AI, which will help move many domains forward. Further, many of the domains presented in this paper (Section 3.2) are highly related, as different aspects of our environment are inherently linked. By aggregating research across domains and making it searchable and filterable, we hope to spur innovation, collaboration, and code sharing between researchers in the pursuit of tackling cross-disciplinary, complex issues such as those related to global warming. For example, water body identification, deforestation monitoring, and wildfire detection are all separate domains, but researchers and practitioners in different domains may use common data sources, processing methods, and algorithms in their final results. As we continue to compile papers written at the intersection of GEE and AI via our web app tool iLit4GEE-AI, it will become easier for researchers to find relevant literature and code resources even if they are from different areas of study.

3. The State of the Art: GEE with AI

In this section, we firstly provide an overview of the reviewed 200 studies using GEE and AI (Section 3.1). Then we investigate studies leveraging GEE with AI from the perspective of applications (Section 3.2) and highlight those with novel methods in Section 3.3.

3.1. Overview of the Reviewed Studies

In this paper, we have reviewed 200 papers. A word-cloud visualization of the titles and keywords of the reviewed 200 papers are provided in Figure 2. The word clouds provide an informative (general and specific) focus of those reviewed papers. The most frequently used words are “Google Earth Engine”, “classification”, “imagery”, “machine learning”, “mapping”, “remote sensing”, and “detection”. We can also see that there are specific keywords, such as “cloud”, “water”, “forest”, “crop”, “soil”, “fire” and “urban” reflecting many of the categories we identified for this review paper. Additionally, “Landsat”, “Sentinel-1”, and “Sentinel-2”, “SAR” (or synthetic aperture radar) as well as “China”, “Brazil”, and “Asia” detail some of the many study areas from these 200 papers. Note that we only included and reviewed the papers that integrate GEE and AI (including its branches ML/DL/CV).

Figure 3 shows that most published work leveraging the power of GEE integrated with AI is still at the application stage and that there is room to develop novel methods to advance earth observation in relevant fields. To break this down further, in (b) we can see that ML is the dominant method, and in (c) the most-used tasks are classification. In Figure 4, the primary applications that have applied GEE integrated with AI are crop, LULC, vegetation, wetland, water, and forest, and that primary study areas are China, Brazil, and the United States. The most-used RS data types are Landsat 8 OLI, and Sentinel-2. From Figure 5, we see that the most-used ML models are RF, SVM, and CART, while the top evaluation metrics used are: overall accuracy (OA), producer’s accuracy (PA), user’s accuracy (UA), and Kappa. (Note that a freely accessible interactive version of the map and charts can be accessed via our web app tool; the web app tool URL and its brief demo video are provided in Appendix A; also, in Appendix B, we provide our resources for an introduction of a list of commonly used evaluation metrics).

3.2. Advances in Applications

We organized the following subsections according to total citation count. Thus, readers will start in the thematic research area with the highest number of citations that use ML/DL/CV on GEE. As readers move through Section 3.2, they will then be covering topics with a less developed presence on GEE (that the authors are aware of) that also utilize ML/DL/CV. For each subsection, a table with information such as study area, RS data type, and what sort of ML/DL model or CV algorithm the authors used will accompany each reference. Note that each table in this section is ordered chronologically to show trends in data type and model usage. Each table will be accompanied by a word cloud showing terms from paper titles and keywords given by the authors. For Section 3.2.8, Section 3.2.9, Section 3.2.10, Section 3.2.11, Section 3.2.12, Section 3.2.13, Section 3.2.14, Section 3.2.15, Section 3.2.16, Section 3.2.17 and Section 3.2.18, there are not enough publications to make the word clouds informative, so in addition to titles and keywords we also include the abstract text. Below each table are accompanying summaries for each reference in the table. References with an “*” next to them denote that a specific paper used novel methods or went beyond a straight-forward application of data and methods on GEE. In this paper, we define a very narrow view of what a novel method is: using ML/DL/CV models and algorithms in new ways on GEE (see Section 3.3 for more details on novel GEE methods). This means that even if a paper combined data in a new way or formed a new data preprocessing method, their paper was deemed an application since our focus is on ML/DL/CV methods. Each paper we reviewed in Section 3.2 is grouped into a specific subsection, ranging from 2 total citations (Bathymetric mapping) to 37 citations (Crop mapping).

Note that papers could be divided up into several different sections. Thus, there is some subjectivity in this assignment of categories to different papers, so readers should be aware that there is overlap. Generally, most papers could be classified as LULC or “land use and land cover”. Agriculture, vegetation, water, and forests are all classes commonly classified in LULC analyses. Our intention was to represent the focus of a given paper. For example, if a paper predicted for several classes in their analysis but their focus was on producing deforestation maps, even for papers that had “LULC” in their title, then their paper will be under “Forest and deforestation monitoring”. Similarly, if a paper was creating vegetation maps of tidal flats, then this paper is under “Vegetation mapping” and not “Wetland mapping”. As another example, if authors were monitoring vegetation or water indices in RS imagery but their goal was to monitor reclamation progress or pollution levels at mining sites, their papers would be found under “Heavy industry and pollution monitoring”. Only in the case where the goal was to expressly create a general LULC map would that paper go under “Land cover classification”.

3.2.1. Crop Mapping

Crop mapping is the most-well-developed application using GEE and AI (37 studies).

Table 1. Studies targeting crop mapping from RS imagery using AI (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Lobell et al. (2015) [45]	regression	multiple linear regression	Cropland Data Layer, Landsat 5, Landsat 7	United States
Shelestov et al. (2017) [46]	classification	CART, ensemble NN, IKPamir, MLP, NB, RF, SVM	Landsat 8 OLI, Landsat 8 TOA	Ukraine
Xiong et al. (2017) [47]	classification, segmentation	RF, RHSeg, SVM	Landsat 8 OLI TOA, Sentinel-2 MSI TOA, SRTM DEM	Africa (continent)
Xiong et al. (2017) [48]	classification	DT, k-means	Africover, CUI, FROMGC, GCEV1, GLC 2000, Global30, Globcover, GRIPC, IKONOS, LULC 2000, MCD12Q1, MYD13, QuickBird, WorldView 2	Africa (continent)
Deines et al. (2017) [49]	classification, regression	CART, RF	AIM-RRB, Cropland Data Layer, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, NLCD, USGS MIrAD-US	United States
Teluguntla et al. (2018) [16]	classification	RF	Google Earth Pro, Landsat 7 TM, Landsat 8 OLI, National Geospatial Agency	Australia, China
Kelley et al. (2018) [50]	classification	RF	Landsat 8 TOA, SRTM DEM	Nicaragua
Ragettli et al. (2018) [51]	classification	k-means, region growing clustering algorithm, RF	Landsat 7, Landsat 8, MOD09A1, MOD09Q1, MYD09A1, MYD09Q1, Sentinel 2, SRTM DEM	Kazakhstan, Kyrgyzstan
Ghazaryan et al. (2018) [52]	classification	decision fusion, RF, SVM	Landsat 8 OLI, Sentinel-1	Ukraine
Mandal et al. (2018) [53]	classification	k-means	Sentinel-1	India
Oliphant et al. (2019) [54]	classification	RF	GeoEye, Landsat 7 ETM+, Landsat 8 OLI, National Geospatial Agency, Quickbird, SRTM DEM, WorldView	Brunei, Cambodia, Indonesia, Japan, Laos, Malaysia, Myanmar, North Korean, Philippines, South Korean, Thailand, Vietnam
Sun et al. (2019) [55]	regression	CNN, CNN-LSTM hybrid, LSTM	Cropland Data Layer, MOD11A2, MOD09A1	United States
Wang et al. (2019) [56]	classification	ANN, CART, RF, SVM	Sentinel 2	China
Tian et al. (2019) [57]	classification	RF	Sentinel-1, Sentinel-2 MSI, SRTM DEM	China
Xie et al. (2019) [58]	classification	RF	Cropland Data Layer, GIAM, GMIA, LANID, Landsat 5, Landsat 7, Landsat 8, MIrAD-US, MOD11A2, MOD13Q1, NAIP, NLCD	United States
Jin et al. (2019) [59]	classification, regression	linear regression, RF	Sentinel-1, Sentinel-2 MSI, MYD11A2	Kenya, Tanzania
Rudiyanto et al. (2019) [60]	classification	ANN, k-means, RF, SVM	Google Earth, Sentinel-1	Indonesia, Malaysia
Wang et al. (2019) [61]	classification	GMM, k-means, RF	Cropland Data Layer, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	United States
Liang et al. (2019) [62]	classification	CART	Cropland Data Layer, Landsat 5 TM, Landsat 8 OLI	United States
Tian et al. (2019) [63]	classification	DT	Landsat 7, Landsat 8, MOD13Q1, Sentinel-2, SRTM DEM	China
Neetu et al. (2019) [64]	classification	CART, RF, SVM	Sentinel-2 MSI	India
Gumma et al. (2020) [65]	classification	RF	GeoEye, GDEM, IKONOS, Landsat 7, Landsat 8, Quickbird, WorldView	Bangladesh, Bhutan, India, Nepal, Pakistan, Sri Lanka
Han et al. (2020) [66]	regression	BGT, BST, DT, GPR, KNN, NN, RF, SVM	MODIS Terra	China
Phalke et al. (2020) [67]	classification	CART, RF, SVM	Google Earth, Landsat 7 ETM+, Landsat 8 OLI, National Geospatial Agency, SRTM DEM	Asia, Europe, Middle East (multiple countries)
Samasse et al. (2020) [19]	classification	RF	Landsat 8 OLI, Rapid Land Cover Mapper	Burkina Faso, Mali, Mauritania, Niger, Senegal
Chen et al. (2020) [68]	classification	RF	JRC Global Surface Water, Sentinel-1, Sentinel-2 MSI	China
Amani et al. (2020) [69] *	classification, segmentation	ANN, SNIC	Canada’s Annual Crop Inventory, MCD12Q1, Sentinel-1, Sentinel-2	Canada
You and Dong (2020) [70]	classification	RF	Google Earth, Sentinel-1, Sentinel-2 MSI	China
Poortinga et al. (2021) [71]	segmentation	U-Net	NICFI Planet	Thailand
Adrian et al. (2021) [72] *	classification, segmentation	DnCNN, RF, SegNet, U-Net, 3D U-Net	Sentinel-1, Sentinel-2, WorldView 3	United States
Cao et al. (2021) [73]	regression	CNN, DNN, LSTM, RF	MOD13A2, SRTM DEM	China
Luo et al. (2021) [74]	classification	RF, SNIC	Sentinel-1	China
Ni et al. (2021) [75]	classification	SVM	Sentinel-2	China
Sun et al. (2022) [76]	regression	RF	FROM-GLC10, MOD15A3, Sentinel-2	China
Li et al. (2022) [77]	classification, segmentation	RF, SNIC	Google Earth, Sentinel-2	China
Han et al. (2022) [78]	classification	RF	Landsat 5 TM, Landsat 8 OLI, Sentinel-2	China
Hedayati et al. (2022) [79]	classification	fuzzy rules, Maximum Likelihood	ALOS PALSAR, Google Earth, Landsat 8, MYD11A2	Iran

below summarizes those studies and a word cloud generated from the titles and keywords of the 37 studies is provided in Figure 6. The most frequently used words are “Google Earth Engine”, “mapping”, and “classification”. However, the term “yield” also reflects that regression tasks such as yield prediction are almost as common as classification tasks such as producing maps. Additionally, specific words such as “30-m” and “Asia” indicate the spatial resolution and coverage of the RS imagery used in the reviewed papers. From our interactive web app (see Appendix A) and Table 1, Landsat 8 Operational Land Imager (OLI), Shuttle Radar Topography Mission (SRTM) DEM, and Sentinel-2 are the most-used RS data. The most popular AI models are RF, SVM, Classification and Regression Trees (CART), and k-means models; and the most-used evaluation metrics are user’s accuracy (UA), producer’s accuracy (PA), Kappa, and R². A brief summary of those studies is provided right below Table 1. More detailed textual summaries for most of the reviewed crop mapping studies are provided in Appendix C.1.

Creating country-level, crop-specific maps using RS data can be difficult because of the large amount of data involved. GEE provides data storage and online processing capabilities, greatly ameliorating the issues with downloading data and managing computing resources. Several algorithms were compared in [46] on GEE, CART, IKPamir, logistic regression, a MLP, NB, RF, and an SVM, for crop-type classification in Ukraine. The authors also used an ensemble NN but had to move off the GEE platform since NNs were not currently supported. It is often difficult to map croplands on a large scale using RS imagery because of a lack of ground-truth validation data. However, there are also problems relating to differing cultivation techniques and definitions of what makes up cropland. To address these issues, the authors in [67] collected large amounts of training points from Google Earth imagery and analyzed Landsat and DEM data to create a cropland data layer across Europe, the Middle East, and Russia. Accurate classification and mapping of crops is essential for supporting sustainable land management. A two-step approach for crop identification in the central region of Ukraine was developed in [52] by exploiting intra-annual variation of temporal signatures of remotely sensed observations (Sentinel-1 and Landsat images) and prior knowledge of crop calendars. Crop maps are often created using vegetation indices and field observation data. The authors in [73] argued that this may lead to datasets and ML models that can only predict in specific areas and not generalize up to larger areas (i.e., regions or countries) or to other time periods in the same area. They further argued that what is needed is a more generalized method that can take in information like weather and climate data or DEM data and scale up to field-level predictions or larger.

Agricultural expansion can cause harmful effects to ecosystems and their levels of biodiversity. Producing crop-type maps using RS imagery and ML is one way to help monitor agricultural expansion over large areas, and these maps in turn can help policymakers and land-use managers make more informed decisions about current and future land use. However, creating the maps themselves normally requires a lot of data and it is not a straightforward task to pick an ML model that will perform well with that data. There is also the concern that the predictions from that ML model will be uninterpretable, given that many ML and DL models are so-called “black boxes”. To get around this issue, the authors in [79] trained a maximum likelihood model and a fuzzy-rules classifier to determine paddy rice distribution in Iran. Plants look very different in RS imagery depending on the type of imagery that you use, but also over the course of a plant’s lifetime. This is especially true of crops like rice, so it is important to incorporate phenological information in order to be able to monitor it over time. Over a three-year time period, the authors in [75] were able to map paddy rice using Sentinel imagery by utilizing several different spectral indices and creating composites of different paddy rice growth periods. Continued agricultural expansion threatens many ecosystems around the globe with high levels of biodiversity. Being able to monitor agricultural expansion is one part in being able to make timely decisions related to water and soil health in addition to pollution levels caused by fertilizer use. Mapping croplands over a large scale with NNs and high-resolution RS imagery has resulted in highly accurate maps, but NNs are computationally expensive to train. A U-Net was used in [71] to map sugarcane in Thailand but used a lightweight NN as an encoder for the DL model to reduce computing costs. Sugarcane grows in rainy conditions in complex landscapes, making mapping it difficult. However, using phenology information can help identify sugarcane in high-resolution RS imagery as shown in [56]. The performance of ANN to CART, RF, and SVM models on GEE was compared for sugarcane mapping in China using Sentinel-2 imagery. Shade-grown coffee landscapes are critical to biodiversity in the forested tropics, but mapping it is difficult because of mountainous terrain, cloud cover, and spectral similarities to more traditional forested landscapes. Landsat, precipitation, and DEM data were used in [50] to map shade-grown coffee in Nicaragua using an RF model. Accuracy scores across different land class types (including shade-grown coffee) were high; a relative variable importance was also analyzed on what data contributed most to the RF model’s performance. It is difficult to know beforehand the effect different datasets will have on producing LULC maps. It is therefore useful to compare the performance of a ML classifier on different datasets like Landsat and Sentinel imagery, so that future researchers know which datasets fit their application. The differences between Landsat and Sentinel imagery were explored in [78] for identifying cotton in China over the course of the plant’s life cycle.

Crop maps are increasingly being produced at the national and global levels, but this process requires a lot of compute resources. Cloud computing offers free access to data and computing, yet many studies producing crop maps and crop yield estimates do not take advantage of these resources. In the United States, crop yield estimates for soybeans start very late in the season, but early estimates are needed to inform management decisions like when to harvest. The authors in [55] used a CNN–LSTM hybrid model to predict soybean yield in the contiguous United States using RS imagery alongside weather data and showed that the hybrid approach works better than either CNN or LSTM alone, although the results were better in some states than others. Additionally, the authors created combinations of input data to determine which variables were most important in training their NN. Still, the authors had to move their DL training off the GEE platform because it did not currently support NN architectures. Many variables, including climate/weather, fertilizer, soil, economic, and hydrological data, can be incorporated into crop yield prediction simulation models. However, this amount of data, needed to make the crop models accurate, is often not available in specific countries or are too time-consuming and cost-intensive to collect and maintain. RS imagery can help fill this need by providing open data over long temporal scales and global coverage, regardless of country. The authors in [66] demonstrated that by using climate and soil data with RS imagery on the GEE platform, it was possible to predict winter wheat yields 1–2 months ahead of harvesting in China. Producing crop type maps is often a useful first step in predicting crop yield. However, crop type maps that are derived from lower-resolution RS data suffer from uncertainties in areas where soil, crops, and plants are heavily mixed. Current cropland products only focus on a subset of staple crops. Optical and SAR Sentinel data were combined in [72] to create higher-resolution maps capable of displaying information on less commonly mapped non-staple crops in the US.

It is challenging to map cropland extent over large countries or regions in a rapid, repeatable, and accurate manner. This is in part due to the large amount of RS imagery that are usually required to make these maps, in addition to needing to access validation datasets in comparable formats across geo-political boundaries. Even when this is possible, crop maps are created using coarse RS imagery, limiting the utility of the output crop maps. In [16], the authors feed RS imagery with elevation and government data in Australia and China into an RF model to produce crop extent maps at 30 m, 250 m, and 1 km resolutions. It is difficult to achieve continuous, cloud-free imagery in Australia and China over time, so this analysis depends on creating bi-monthly composites. The authors noted that this analysis could have benefitted from a larger dataset in addition to comparing more classification algorithms to help reduce uncertainties from the RF model. LAI and fraction of photosynthetically active radiation (FPAR) are two important features while trying to produce crop extent maps and crop yield estimates. However, most current products for producing crop extent maps and crop yield estimates are derived from low-resolution RS imagery. In order to produce these maps and estimates at a higher resolution, the authors in [76] utilized GEE, Sentinel-2 and field data to train an RF to first estimate LAI and FPAR at a much finer spatial scale.

Global crop maps often fail to capture small farms because the resolution of the RS imagery used to create the maps is too coarse. Additionally, agricultural areas change over time and so the underlying validation (which is hard to acquire in the first place) often changes. Thus, producing high-resolution maps that track agricultural areas that are able to track crop production over time in an accurate fashion has proved difficult. Landsat-8 and Sentinel-2 imagery were combined in [47] with elevation data to produce a crop map across continental Africa on the GEE platform. Crop maps that are produced to cover a large area are often created from coarse RS imagery. This poses problems with identifying small or fragmented farms, as well as farms that are mixed-use or have several crop types over the same small area. Several attempts have been made to map land-use classes over large areas, but these maps do not focus specifically on crops and so their utility to food production studies is limited. To address these issues, [54] used RS imagery from several different platforms (GeoEye, Landsat, NGA, Quickbird, WorldView) to produce a 30-m resolution crop map for Southeast and Northeast Asia. Using an RF model, the authors achieved high accuracy rates across several crop type classes and made the resulting data layer public. However, to create cloud-free scenes from optical imagery across countries, the authors had to rely on multi-year composites. The authors noted that in the future, a harmonized Landsat–Sentinel dataset would be useful to expand spatial and temporal data coverage.

Sustainable management of agricultural water resources requires improved understanding of irrigation patterns in space and time. Annual irrigation maps (1999–2016) in the US Northern High Plains were produced in [49] by combining all available Landsat satellite imagery with climate and soil covariables in an RF classification workflow. In [51], the authors implemented an automatic irrigation mapping procedure in GEE that uses surface reflectance satellite imagery from different sensors (Landsat 7/8, Sentinel-2, MODIS Terra and Aqua imagery, SRTM DEM). A rapid method was developed to map Landsat-scale (30 m) irrigated croplands in [58] across the conterminous United States (CONUS). The method was based upon an automatic generation of training samples for most areas based on the assumptions that irrigated crops appear greener than non-irrigated crops and had limited water stress.

Cropland classification is highly dependent on RS imagery resolution, the scale of a given analysis, the processing steps, and the input training data. Coarse resolution cropland data products have been found to contain large errors, but even higher resolution maps tend to have low accuracy rates and overestimate overall crop area. An open-source map was created in [19] for several West African countries using an RF model trained on Landsat data. The amount of RS data collected is increasing every day. This poses a problem for how best to analyze RS imagery and extract useful information from it, regardless of the EO domain. The authors in [77] implemented a dynamic feature importance tool that automatically finds the most important subset of input features for identifying crop types in China. They then fed these features to the SNIC algorithm and then to an RF on GEE and combined the output predictions with growth period information to produce crop-type maps that incorporate plant phenology. By incorporating growth stage information as an input feature to the ML model, the authors achieved a 6–7% boost in OA, precision, and recall across different crops like rice, maize, and soybeans. In their paper, the authors showed that red edge, NDVI, red, SWIR2, and aerosol information contributed the most to their analysis. However, the authors themselves stated that their method was unstable due to the nature of their feature importance algorithm. Depending on what data was chosen with their feature importance algorithm, the accuracy of the method fluctuated. Thus, their method was good for reducing data size and should be used when compute is limited, though using all of the data in a given time series was shown to work better.

3.2.2. Land Cover Classification

Land cover classification is the 2nd-most-developed domain area using GEE and AI (27 studies total).

Table 2. Studies targeting LULC from RS imagery using AI (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Azzari and Lobell (2017) [80]	classification	RF	Landsat 8 OLI TOA, Sentinel-2 MSI TOA, SRTM DEM	Zambia
Midekisa et al. (2017) [81]	classification	RF	DMSP NTL, Globeland30, Hansen Global Forest Change, Landsat 7 ETM+	Africa (continent)
Hu et al. (2018) [82]	classification	CART	GlobCover, Landsat 5 TM, Landsat 8 OLI	China
Ge et al. (2019) [83]	classification	Bayesian hierarchical model, RF	ALOS DSM, Landsat 8 OLI, VIIRS NTL	China
Lee et al. (2018) [84] *	classification	BULC-U	GlobCover, Landsat 5	Brazil
Zurqani et al. (2018) [85]	classification	RF	Landsat 5 SR, Landsat 5 TOA, Landsat 8 SR, Landsat 8 TOA, NAIP, NLCD, SRTM DEM, USGS Watershed Boundary Dataset	United States
Murray et al. (2018) [86] *	classification	RF	Landsat 7 ETM+, Landsat 8 OLI, Landsat 8 SR, SRTM DEM	Global
Mardani et al. (2019) [87]	classification	BT, SVM	FAO land cover, Sentinel-2	Lesotho
Gong et al. (2019) [88]	classification	RF	FROM-GLC, Landsat 8, Sentinel-2, SRTM DEM	Global
Hao et al. (2019) [89]	classification	CART	GLDAS, GlobeLand30, Landsat 8 OLI, MOD11A2, MOD13A2	China
Miettinen et al. (2019) [90]	classification	DT, maximum likelihood, RF	ALOS PALSAR-2, Landsat 7 ETM+, Landsat 8 OLI, Sentinel-1, SRTM DEM	Brunei, Indonesia, Malaysia, Singapore, Timor-Leste
Xie et al. (2019) [91]	classification	RF	Global Field Photo Library, Google Earth, Landsat 5 TM, Landsat 7 ETM+, MCD12Q1, MCD43A4	China
Adepoju and Adelabu (2020) [92]	classification	CART, gmoMaxEnt, RF, SVM	Google Earth, Landsat 8 OLI, SRTM DEM	South Africa
Ghorbanian et al. (2020) [93]	classification	RF	Google Earth, Sentinel-1, Sentinel-2	Iran
Liang et al. (2020) [94] *	classification	CART, MD, RF	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, SRTM DEM	China
Zeng et al. (2020) [95]	classification	RF	GFSAD, GHSL, Hansen Global Forest Change, Landsat 8 OLI, Sentinel-1 GRD, SRTM DEM	South Africa
Naboureh et al. (2020) [96]	classification	RF	Landsat 8 OLI, SRTM DEM	Iran
Naboureh et al. (2020) [97]	classification	SVM, RUESVM	Google Earth, Sentinel-2	China, Iran
Li et al. (2020) [98]	classification	RF	FROM-GLC10, GHSL, Landsat 8 OLI, MYD11A2, Sentinel-2 MSI, SRTM DEM, Suomi-NPP NTL	Africa (continent)
Huang et al. (2020) [99]	classification	RF	CCI-LC, FROM-GLC, Google Earth, Landsat 5 TM	Global
Tassi and Vizzari (2020) [100]	classification, segmentation	RF, SNIC, SVM	Landsat 8, PlanetScope, Sentinel-2	Italy
Shetty et al. (2021) [101]	classification	CART, RF, RVM, SVM	BCLL, GlobCover, Google Earth, Landsat 8 OLI	India
Feizizadeh et al. (2021) [102]	classification	CART, RF, SVM	aerial photography, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Iran
Shafizadeh-Moghadam et al. (2021) [103]	classification, segmentation	RF, SNIC	Google Earth, Landsat 8	Iran, Iraq, Kuwait, Saudi Arabia, Syria, Turkey
Pan et al. (2021) [104]	classification	CART, RF	Landsat 5 TM, MCD12Q1, SRTM DEM	Australia, United States
Becker et al. (2021) [105]	classification	RF	Landsat 8	Brazil
Jin et al. (2022) [106]	classification, segmentation	RF, SNIC	ALOS PALSAR, CCI-LC, CGLS-LC, FROM-GLC, GFSAD30, GHSL, JRC Global Surface Water, Landsat 7 ETM+, Landsat 8 OLI, MCD12Q1	Asia (multiple countries)

below summarizes those studies and a word cloud generated from the titles and keywords of those papers is provided in Figure 7. The most frequently used words are “Google Earth Engine”, “land”, “cover”, and “classification”. “Landsat” RS imagery features heavily in LULC research, though the trend is moving towards higher-resolution data and creating maps over much larger areas. The words “Sentinel” and “Africa” illustrate this point well.

From our interactive web app (see Appendix A) and Table 2, Landsat 8 OLI, SRTM DEM, and Google Earth are mostly used. The most popular AI models are RF, CART, and SVM, and the mostly used evaluation metrics are overall accuracy (OA), PA, UA, and Kappa. A brief summary of those studies is provided right below Table 2. More detailed textual summaries for most of the reviewed land cover classification studies are provided in Appendix C.2.

LULC maps can help decision-makers and land managers make more informed decisions about the environment. Still, producing LULC maps with ML and RS data requires a lot of compute and labeled input training data. GEE currently offers free compute, so researchers can use the data that they are interested in without having to worry about hardware setup or compute time. The authors in [102] took advantage of this to create an LULC map in Northern Iran, predicting for water, rangelands, built-up areas, orchards, and other LULC classes. They used Landsat RS imagery, field observations, and historical datasets to train CART, RF, and SVM models. The SVM performed better than the CART and RF models, but perhaps more importantly the authors also ran a spatial uncertainty analysis to show each model’s confidence level on the output maps. More research should include uncertainty incorporated into reporting metrics or on maps produced with ML to better convey a model’s certainty to both citizens and decision-makers.

There are currently high data and computational costs of having to store RS data across different machines using different ML algorithms. There is an additional challenge in that most RS analyses depend on optical data, which is often obscured by clouds and shadows. In addition, most land cover maps have coarse resolution and do not often describe the same things as other maps (making them not directly comparable). These static maps need to be more accurate and updated frequently to be of real use, and cloud computing alongside data and algorithms being in one place has allowed both of these to become a reality. An RF model was used in [80] to determine land-use classes such as vegetation, croplands, and urban areas from Landsat imagery in Zambia. An approach was presented in [81] to quantify continental land cover and impervious surface changes over continental Africa for 2000–2015 using Landsat images and an RF classifier on GEE. Simple change detection based on Landsat images from two different years with two different phenophases yields unsatisfactory results and may induce many misclassifications and pseudo-change identifications because of the phenological differences between RS images. A land-use/land-cover type discrimination method based on a CART was proposed in [82], which applied change-vector analysis in posterior probability space (CVAPS) and the best histogram maximum entropy method for change detection, and further improved the accuracy of the land-updating results in combination with NDVI timing analysis. The last land-cover map of Iran was produced with MODIS imagery in 2016. Now, there are much higher resolution satellite data products, but it is difficult to collect more ground-truth validation data. Cloud computing and ML can help produce newer land cover classification maps that are easy to reuse. Such a workflow was designed in [93] on GEE for Iran using Sentinel-1 and -2 data and an RF model and SNIC. With the ground-truth training samples available, the authors used SNIC to segment land-use classes into objects while the RF model classifies them on the pixel level.

Numerous efforts have been made to end poverty around the globe. Mapping land-use changes in poverty areas can provide insights into the poverty reduction progress. Landsat images available on GEE were utilized in [83] to map annual land-use changes in China’s poverty-stricken areas. An open-source land cover mapping processing pipeline was created in [87] using GEE. The authors argued that land cover maps specifically can help countries properly plan for sustainable levels of food production, but that many developing countries did not have the financial or compute resources to monitor land classes in real time. Using SVM and bagged trees (BT) models, the authors predicted urban, agriculture, tree, vegetation, water, and barren land-use types in Lesotho.

In RS imagery, many different land-use types have similar spectral signatures or are very complex, making them difficult to be properly identified. Several different ML models available on GEE were trained in [92] with different combinations of input data to determine which were the most important in determining land-use types in Golden Gate Highland Park in China. Although RS and ML have allowed LULC analysis to become ever more accurate for general LULC classes, it is still challenging to correctly identify land subtypes. For example, while classifying vegetation to a high degree of accuracy has become more commonplace, identifying vegetation subtypes like shrubs or grassland is not as straightforward, especially in mixed-use areas. In addition, as is the case for many RS applications, it is challenging to know which types of input data will contribute to a given ML model’s ability to learn these subtypes. Therefore, the authors in [95] set out to compare the contribution of SAR data and different indices (NDVI, EVI, SAVI, NDWI) derived from optical data on overall classifier performance. A land cover map of the whole African continent at 10 m resolution was generated in [98], using multiple data sources including Sentinel-2, Landsat-8, Global Human Settlement Layer (GHSL), Night Time Light (NTL) Data, SRTM, and MODIS Land Surface Temperature (LST). Different combinations of data sources were tried to determine the best data input configurations. Pixel-based classification methods often suffer from “salt-and-pepper” noise in their end predictions. Object-based classifiers can help alleviate this problem but are not commonly used because of their high compute overhead. While GEE does not have many object-based classifiers, it does provide free compute. To take advantage of this while comparing the performance of pixel-based and object-based classification methods, [100] produced LULC maps in Italy using Landsat, Planet, and Sentinel RS imagery. The authors compared the performance of RF and SVM models alone with that of the same models used in conjunction with the SNIC and gray-level co-occurrence matrix (GLCM) texture data. Their results showed that pixel-based methods worked better at lower resolutions (i.e., using Landsat data), whereas object-based methods worked better for higher-resolution RS imagery. The best classifier was the RF model trained with SNIC and incorporating GLCM data. Still, the authors noted that ML models were heavily influenced by input data, feature engineering, the classes that you were trying to predict for, and the place being studied. Many studies evaluate ML methods and the effect that input data sources have on their performance. Not as much research is done into determining how data sampling strategies affect ML classifiers. The authors in [101] compared different data sampling strategies and their effects on how different ML classifiers performed on LULC tasks. A multi-seasonal sample set was collected in [88] for global land cover mapping in 2015 from Landsat 8 images. The concept of “stable classification” was used to approximately determine how much reduction in training sample and how much land cover change or image interpretation errors can be acceptable.

Mountain Land Cover (MLC) classification can be relatively challenging due to high spatial heterogeneity and the cloud contamination in optical satellite imagery over the mountainous areas. Distribution of Land Cover (LC) classes in these areas is mostly imbalanced. To date, three approaches have been proposed to address the class imbalance problem: (1) applying specific classification methods by focusing on the learning of minority classes, (2) assigning higher weights on minority classes by adjusting classifiers, and (3) rebalancing training datasets (e.g., oversampling and under-sampling techniques). A hybrid data-balancing method, called the Partial Random Over-Sampling and Random Under-Sampling (PROSRUS), was proposed in [96] to resolve the class imbalance issue. The class imbalance problem reduces classification accuracy for infrequent and rare LC classes. A new method was proposed in [97] by integrating random under-sampling of majority classes and an ensemble of Support Vector Machines, namely Random Under-sampling Ensemble of Support Vector Machines (RUESVMs). Rapid urban expansion puts pressure on local ecosystems and human well-being, so urban sustainability studies are increasingly turning to applications that process large amounts of geospatial data and model ecosystem services. Currently, it is not straightforward for urban or ecology scientists to use cloud-based platforms like GEE as their processing routines are more complicated than the many common mapping applications (i.e., classification) available on GEE. While determining ecosystem service values is complicated (many disciplines, many opinions, etc.), GEE was used in [94] to illustrate a processing workflow for how LULC classes can be used to compute more complex ecosystem service values.

Watersheds around the world are under stress, both due to climate change and human disturbance. LULC maps can help with planning and conservation decisions, but they are often difficult to make because they are compute-intensive to make. GEE has helped many researchers by providing freely available data, methods, and compute, but researchers often find that they run into compute limits on the platform before they can complete their analyses. To overcome these compute limits in GEE, the authors in [103] used feature reduction techniques and designed their own parallel processing algorithms to produce an LULC map across several Middle Eastern countries. To get a better idea of how water resources were being affected by LULC classes, the authors combined topographic data, spectral data, RS image composites, and texture information to train a combined SNIC-RF model. They achieved high accuracy across several LULC classes and showed feature importances for each class in their analysis. However, the authors noted that other than SNIC, advanced object-based classification and segmentation algorithms were not available on GEE.

3.2.3. Forest and Deforestation Monitoring

Forest and deforestation monitoring is the 3rd-most-developed application using GEE and AI (20 studies total).

Table 3. Studies targeting forest change and deforestation from RS imagery using AI.

References	Method	Model Comparison	RS Data Type	Study Area
Lee et al. (2016) [107]	classification	CART, MD, RF	Landsat 8	Indonesia
Wang et al. (2019) [15]	classification	RF	ALOS PALSAR, GlobeLand30-2010, Hansen Global Forest Change dataset, JRC Yearly Water Classification History, Landsat 5 TM, Landsat 7 ETM+, RapidEye, TerraClass-2010, USGS Global Tree Cover 2010	Brazil
Voight et al. (2019) [108]	classification	CART, Markov Chain model, MLP	Google Earth, Landsat MSS, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Belize
Koskinen et al. (2019) [109]	classification	CART, RF, SVM	ALOS PALSAR, Google Earth, Landsat 8 OLI, NAFORMA, Sentinel-1, Sentinel-2 MSI, SRTM DEM	Tanzania
Duan et al. (2019) [110]	classification	RF	Google Earth, Sentinel-2	China
Poortinga et al. (2019) [111]	classification	DT, Monte Carlo, RF	ALOS GDSM, Landsat 8, PlanetScope, RapidEye, Sentinel-1, Sentinel-2	Myanmar
Shimizu et al. (2019) [112]	classification	RF	Google Earth, Landsat 8, MCD12Q1, PlanetScope, RapidEye, Sentinel-1	Myanmar
Ramdani (2019) [113]	classification	GMM, KNN, RF, SVM	Sentinel-1, SRTM DEM	Indonesia
Çolak et al. (2019) [114]	classification	SVM	CORINE LULC, Sentinel-1, Sentinel-2	Turkey
Shaharum et al. (2020) [115]	classification	CART, RF, SVM	Google Earth, Landsat 8, SRTM DEM	Malaysia
de Sousa et al. (2020) [116]	classification	RF	ALOS PALSAR, Landsat 8 OLI, SRTM DEM	Gabon, Liberia
Brovelli et al. (2020) [117]	classification	ANN, RF	CBERS 2B, CBERS 4, Landsat 5, Landsat 7, Landsat 8, Sentinel-2	Brazil
Kamal et al. (2020) [118]	classification	SVM	Landsat 8 OLI	Indonesia
Wei et al. (2020) [119]	classification	binomial logistic regression	AW3D30, CHELSA V1.2, GeoEye-1, GMTED2010, Google Earth, Hansen Global Forest Change, Landsat 5, NAIP	United States
Praticò et al. (2021) [120]	classification	CART, k-means, RF, SVM	Sentinel-2	Italy
Xie et al. (2021) [121]	classification	RF	Sentinel-1, Sentinel-2, SRTM DEM	China
Floreano and de Moraes (2021) [122]	classification	Markov-CA, MLP, RF	Google Earth Pro, Landsat 5 TM, Landsat 7 ETM, Landsat 8 OLI	Brazil
Kumar et al. (2022) [123]	classification	RF	Forest Survey of India, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, MCD12Q1	India
Zhao et al. (2022) [124]	classification, segmentation	LandTrendr, RF, U-Net	Google Earth Pro, Hansen Global Forest Change, MTBS, MCD64A1, Planet, Sentinel-1, SRTM DEM	Brazil, United States
Wimberly et al. (2022) [125]	classification, segmentation	LandTrendr, RF	Google Earth, Landsat 7 ETM+, Landsat 8 OLI, WorldView	Ghana

below summarizes those studies and a word cloud generated from the titles and keywords of those studies is provided in Figure 8. The most frequently used words are “Google Earth Engine” and “forest change”. A significant amount of research is done in monitoring deforestation over time and differentiating forest cover from plantations of oil palm, for example. Thus, the words “deforestation”, “plantation”, “palm oil”, and “time series” feature prominently in the word cloud. Additionally, Landsat, Sentinel-1, and Sentinel-2 data are frequently used in forest and deforestation monitoring research in tropical forests in places like the Amazon and Myanmar. From our interactive web app (see Appendix A) and Table 3, the most-used RS datasets are Landsat 8 OLI, SRTM DEM, and Google Earth. The most popular AI models are RF, SVM, and CART, and the most-used evaluation metrics are OA, PA, UA, Kappa. A brief summary of those studies is provided right below Table 3. More detailed textual summaries for most of the reviewed forest and deforestation monitoring studies are provided in Appendix C.3.

Forests provide many ecosystem services, from preventing soil erosion, regulating the hydrological cycle, and providing shelter for many plant and animal species. However, deforestation is occurring at a rate that is making it impossible for individual species to recover. As deforestation accelerates, there are cascading effects for entire ecosystems. In Brazil, agriculture, ranching, and land occupation is causing the vast forest of the Amazon to become fragmented. Still, it is difficult to monitor the changes through time due to cloud cover and the rate that new satellite imagery comes in every day. The authors in [117] showed how GEE can be used to overcome data storage and compute needs and analyze about 20 year’s worth of Landsat data to determine forest cover changes. Land use maps can help inform policymakers and land-use managers but are often static and of coarse resolution. It would be more useful to create these maps in a repeatable manner, one in which code and data could be reused for making decisions based on up-to-date information. Sentinel-2 data were analyzed in [120] and several different ML classifiers were trained to distinguish between four different forest types in Italy during both summer and winter seasons. Monitoring tree species distribution is an important metric in monitoring overall forest health and in determining current carbon storage efforts. However, doing so is difficult without the use of high-resolution RS data, much of which is either private and inaccessible or too expensive to collect (in the case of LiDAR or UAS data). Recent research in environmental mapping applications uses DL and NN to identify tree species across large areas with minimal feature engineering, but NNs currently need large amounts of compute and labeled input data to train on. To classify tree species across a large area in China while fitting within compute restraints, an RF was trained on the GEE platform in [121] using optical and SAR imagery, DEM data, and field observations.

A participatory forest mapping methodology was developed and tested in [109] to map the extent and species composition of forest plantations in the Southern Highlands area of Tanzania. Collecting field observations of plant phenology can be time- and labor-intensive to repeatedly obtain. RS imagery can help continuously monitor phenology information because of its high spatial and temporal resolution. To create a forest type map in India using RS imagery and ML, the authors in [123] predicted for evergreen and deciduous forest types, as well as “non-forest” classes. Collecting, storing, and processing large amounts of RS imagery presents a barrier to doing research in the environmental and earth sciences. GEE provides data storage, compute, data processing, and ML algorithms on its platform. The researchers in [118] used GEE to map mangrove extent in Indonesia.

Global forests face many anthropogenic threats, one of the most prominent being the conversion to agriculture. This is often done through deforestation by fire or clearcutting, but a less studied mechanism related to forest health is the slow degradation caused by continual disturbances. This is more difficult to track using EO methods like vegetative indices derived from RS data, because forest degradation does not always result in canopy loss. The tropical forests of Ghana, many of which are in protected areas, still suffer from logging, mining activities, fires, and expanding agriculture production. All of these disturbances contribute to declines in forest health; a method was developed in [125] for monitoring tropical forest loss and recovery based on Landsat data. Forest logging and forest fires are both dominant drivers of forest loss worldwide. However, deforestation monitoring efforts are often limited by low-resolution RS data and the inability to create forest maps on a continual basis. Using SAR data as input and high-resolution optical data as validation data, a U-Net was trained on Google Cloud in [124] to create monthly forest loss maps.

Forest degradation and deforestation have been occurring around the globe during the past decades, which threaten the biodiversity of ecosystems. As a highly forested landscape, southern Belize has been experiencing deforestation due to agricultural expansion. Landsat 8 imagery on the GEE was utilized in [108] to perform a supervised classification. An MLP model was then built to predict future deforestation patterns and magnitude based on the drivers of past deforestation patterns in the region. Deforestation is accelerating in the tropics in part due to industrial oil palm plantation expansion. Being able to monitor illegal deforestation can aid in conserving the remaining forested landscape and is thus critical to maintaining biodiversity and ecosystem services. A low-cost method was demonstrated in [107] for monitoring industrial oil palm plantations in Indonesia using Landsat 8 imagery that allowed them to distinguish between oil palm (immature oil palm, mature oil palm), forest, clouds, and water classes using the CART, RF, and MD algorithms. Oil palm can play a key role in both ecosystems and local economies but is also a common cause of deforestation. Thus, monitoring and managing the oil palm industry is often necessary to ensure that deforestation does not occur, but traditional ground surveys are time-, effort-, and cost-intensive. RS imagery can help monitor large areas over time. In [115], several ML models were compared to map oil palm using 30 m Landsat 8 imagery in Malaysia. Tropical forests in different Sub-Saharan African countries face high rates of deforestation to illegal logging and cropland expansion. Being able to monitor tropical deforestation is important to monitor ecosystem balances and health. However, separate efforts to produce deforestation maps or products use different data sources, describe slightly different things, or lead to diverging land cover estimates, making them difficult or impossible to directly compare. To explore how GEE could be used to create an open-source processing pipeline for deforestation mapping in Liberia and Gabon, two different RF models were used in [116] to create data masks and then predictions for various land types there. The Amazon Rainforest is home to much of the world’s biodiversity and plays an important role in natural carbon sequestration. However, this region is experiencing high rates of deforestation due to the expansion of agriculture and cattle farming. It remains challenging, though, to monitor such a large area given its size and biological complexity and use that information to produce forest change projections into the future. An RF was used in [122] for initial LULC classification, then used an MLP to simulate possible deforestation scenarios into the future.

Mapping how much carbon forests sequester remains difficult because current techniques rely on mapping forested versus deforested landscapes. However, a major source of uncertainty stems from the fact that degraded forests, ones open to selective logging, are not a separate class but can emit carbon heavily even though they are counted as forested regions. This issue was addressed in [15] by mapping disturbed forest areas in Brazil using 27 years of Landsat surface reflectance imagery.

3.2.4. Vegetation Mapping

Vegetation mapping is a well-developed application domain using GEE and AI (18 studies total).

Table 4. Studies targeting wetland mapping from RS imagery using AI (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Johansen et al. (2015) [126]	classification	CART, RF, NDVI, Foliage Projective Cover	Landsat 5 TM, Landsat 7 ETM+	Australia
Traganos et al. (2018) [127]	classification	CART, RF, SVM	Sentinel-2 LIC TOA	Greece
Tsai et al. (2018) [128]	classification	DT, RF	Landsat 7 TM, Landsat 8 OLI	China
Jansen et al. (2018) [129]	regression	multiple linear regression, polynomial linear regression	Landsat 7 ETM+, Landsat 8 OLI, USGS National Elevation Dataset	United States
Jones et al. (2018) [130]	classification	RF	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, USGS National Elevation Dataset	United States
Campos-Taberner et al. (2018) [131]	regression	RF	BELMANIP2, MCD15A3H, MCD43A4	Global
Xin and Adler (2019) [132]	classification	FCNN, CNN-LSTM hybrid	Sentinel-2 MSI	United States
Parente et al. (2019) [43]	classification, segmentation	LSTM, RF, U-Net	PlanetScope	Brazil
Parente et al. (2019) [133]	classification	RF	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, MOD13Q1	Brazil
Zhang et al. (2019) [134]	classification	RF	Google Earth, Landsat 8 OLI	China
Alencar et al. (2020) [135] *	classification	DT, RF	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Brazil
Zhou et al. (2020) [21]	regression	CART, CNB, MLP, RF, SVM	Landsat 8 OLI, MCD43A1, MCD43A4	United States
Tian et al. (2020) [136]	classification	SAE, SVM	Google Earth, Landsat 5 TM, Landsat 8 OLI, Pleiades 2, QuickBird, SPOT 4, SPOT 6, UAS, WorldView 1, WorldView 3	China
Srinet et al. (2020) [137]	classification	RF	MOD09A1, SRTM DEM, WorldClim V2 Bioclim	India
Long et al. (2021) [138] *	classification, segmentation	CART, LandTrendr, MD, NB, RF, SVM	CGLS-LC100, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, Sentinel-1, Sentinel-2, SRTM DEM	China
Yan et al. (2021) [139]	classification	RF	Gaofen-2, Landsat 4 TM, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, Pléiades A, QuickBird, UAS, WorldView 2	China
Wu et al. (2021) [140]	classification	RF	Gaofen-2, Landsat 8 OLI	China
Pipia et al. (2021) [141] *	regression	GPR	HyMap, Sentinel-2	Europe (multiple countries)

below summarizes those studies and a word cloud generated from the titles and keywords of the 18 papers is provided in Figure 9. The most frequently used words are “Google Earth Engine”, “Landsat”, and “vegetation mapping”. Plant dynamics (i.e., phenology and land-use changes over time) play an important role in differentiating vegetation from forest cover. From our interactive web app (see Appendix A) and Table 4, the most-used RS datasets are Landsat 8 OLI, Landsat 5 Thematic Mapper (TM), and Landsat 7 Enhanced Thematic Mapper Plus (ETM+). The most popular AI models are RF, CART, and SVM, and the mostly used evaluation metrics are PA, UA, and OA. A brief summary of those studies is provided right below Table 4. More detailed textual summaries for most of the reviewed vegetation mapping studies are provided in Appendix C.4.

Accurate near real-time estimates of vegetation cover and biomass are critical to adaptive rangeland management. An approach was developed and tested in [129] to automate the mapping and quantification of vegetation cover and biomass using Landsat 7 and Landsat 8 imagery across the grazing season (i.e., changing phenological conditions). Annual percent land cover maps of plant functional types across rangeland ecosystems were produced to effectively and efficiently respond to pressing challenges facing conservation of biodiversity and ecosystem services. The authors in [130] utilized the historical Landsat satellite record, gridded meteorology, abiotic land surface data, and over 30,000 field plots within an RF model to predict per-pixel percent cover of annual forbs and grasses, perennial forbs and grasses, shrubs, and bare ground over the western United States from 1984 to 2017, at approximately 30 m resolution. Rangelands in the western United States are home to many different animal and plant species. They are ecologically diverse and have been traditionally monitored by taking and analyzing in situ measurements in different areas. However, continually collecting field observations can be time- and labor-intensive and land managers are often asked to make decisions about large areas with sparse field information. RS data can help monitor rangelands with a large spatial scope and a short return time, making them key to informing land management decisions in a timely manner. Using climate and field data alongside Landsat imagery and MODIS land-use maps, ML models used in [21] were able to predict for several important rangeland indicators like plant height, total vegetation and rock cover, as well as bare soil.

Invasive species can degrade ecosystems and harm biodiversity as well as soil and water quality. It is often difficult to monitor invasive species in coastal environments from optical RS imagery, though, because of frequent cloud cover. A specific invasive species in China was used in [136] as a case study for developing an ML pipeline that takes into account both cloud cover and phenological information. Invasive species can have harmful environmental effects as they disrupt ecosystem balances. Long-term datasets, like those of the grass S. alterniflora, are not always available, making them difficult to detect using RS methods. In order to produce a map of this invasive species, field data were collected and processed in [139] in addition to UAS imagery and optical RS data from several different platforms.

It is often difficult to detect changes in savanna landscapes due to their high heterogeneity in vegetation types, which makes it even harder to attribute change to natural or anthropogenic causes. This is especially problematic in areas like the Brazilian Cerrado where agricultural expansion is happening on a large scale. In order to clarify what changes have been happening there, over three decades worth of Landsat imagery was used in [135] to determine which areas have experienced vegetation change. Wetlands provide many ecosystem services and provide important habitats for several different plant and animal species. In order to make informed conservation and policy decisions, it is important not only to be able to map the current state of wetlands vegetation, but how that vegetation is changing over time. However, different sets of input data and ML methods used for change detection of wetland vegetation need to be evaluated more fully as choices made during preprocessing and hyperparameter tuning can affect the end result of an analysis. The authors in [138] used an adaptive stacking algorithm to train an ML classifier on optical, SAR, and DEM data to identify wetland vegetation.

Seagrasses provide many ecosystem services, from carbon storage, providing habitat for many marine species, and preventing coastal erosion. However, they are in decline due to anthropogenic impacts. Mapping their extent is key to being able to conserve them. Bathymetry and RS data were combined in [127] to create a processing and analysis pipeline for large-scale seagrass habitat monitoring in Greece using GEE. Grasslands are often integrated into land-use type or cropland-specific maps, even high-resolution products. However, different grassland species are not identified and thus are classified as a single homogenous land or crop type. This is a problem not just because previous maps have not separated out different grassland types, but it is difficult to recognize them in RS imagery because they look very similar. Some experts are able to recognize such classes, but it is time-consuming to analyze grassland types at scale. Thus, DL techniques that do not rely on expert knowledge are needed so that these identification systems can work over large areas over time. A CNN–LSTM hybrid model was used in [132] to identify grassland types in Sentinel-2 imagery in the United States.

Feature engineering is important in ML, but it is labor-intensive and often requires domain expertise [1]. As one ML branch, DL does not need feature engineering, as deep NN will figure it out from large, annotated data examples, but DL requires much more large training data than ML [1]. The authors in [43] addressed this issue by comparing the performance of an RF model with feature engineering to an LSTM and U-Net NN models without feature engineering for identifying pasturelands in Brazil. Monitoring vegetation on a large spatial scale can be difficult because field data collection takes only snapshots in time and is labor-intensive and expensive. Instead, methods for measuring vegetation need to be done over time so that change detection is possible. Still, novel methods, such as those utilizing RS imagery, need to meet current governmental quality standards. An example of how this can be done is illustrated in [126] in Australia using the GEE platform by comparing how well several ML classifiers compare to index-based methods like NDVI. Although coastal wetland systems are critical habitats for different animal and plant species, it is difficult to monitor them due to cloud cover and difficulty in obtaining RS imagery at high and low tides. Previous studies have used single images or spectral time series to try and identify wetland vegetation, but coastal wetland environments are complex ecosystems. The same species of plant can look different at different stages of its life while also being submerged under water in some RS scenes. The authors in [140] argued that phenology information in RS time series can better capture tidal flat wetland vegetation and so compared phenology information to statistical (min, max, median) and temporal features (quartile ranges). Mapping plant functional types is important because it can give ecosystem modelers and environmental planners a better idea of the spatial distribution of vegetation. This in turn has implications on how resilient areas and ecosystems are and will be to changing climatic factors like heat stress. However, plant function type classification relies on and is often derived directly from current LULC map products that themselves can contain inaccuracies. To explore how plant functional types can be derived directly from RS information, [137] trained an RF model on field, DEM, MODIS, and climate data.

Many methods have been developed to estimate different vegetative properties from RS imagery in response to environmental changes. One such method, Gaussian Process Regression (GPR), is increasingly used to do so because it is a transparent ML model that also outputs model uncertainties. However, as environmental and earth scientists move to GEE for finding and processing data, they may find a lack of GPR models ready to use or train. This is most likely because GPR models become slow and memory-intensive when trained on large RS time series imagery. Such a model was implemented in [141] that has been optimized for green LAI in RS imagery but does so in a way that is also optimized for GEE.

3.2.5. Water Mapping and Water Quality Monitoring

Water mapping and water quality monitoring is another well-developed application domain using GEE and AI (18 studies total).

Table 5. Studies targeting water body detection from RS imagery using AI (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Pekel et al. (2016) [32] *	classification	expert system	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	global
Zou et al. (2017) [142]	regression	multiple linear regression	Global Inland Water, Landsat 5, Landsat 7, NLCD	United States
Chen et al. (2017) [143]	segmentation	non-local active contour algorithm	Gaofen-1, Google Earth, Landsat 8 OLI, SRTM DEM	Tibet
Wang et al. (2018) [144]	classification	RF	JRC Global Surface Water, Landsat 4 TM, Landsat 5 TM, Landsat 8 OLI	China
Lin et al. (2018) [145]	regression	BRT, multiple linear regression, nonlinear general additive models, RF	Landsat 5 TM, Landsat 7 ETM+	United States
Griffin et al. (2018) [146]	regression	multiple linear regression	Landsat 5 TM, Landsat 7 ETM+, NASA GSFC Ozone Monitoring Instrument	Canada, Russia, United States
Isikdogan et al. (2019) [147] *	segmentation	DeepWaterMapv, DeepWaterMap, MNDWI, MLP	Landsat 8	Global
Fang et al. (2019) [148]	regression	linear regression, polynomial regression	China Lake Dataset, China’s Ecosystem Assessment and Ecological Security Pattern Database, Global Lakes and Wetlands, Global Reservoir and Dam Database, HydroLakes, Hydroweb, JRC Global Surface Water, SRTM DEM	China
Fuentes et al. (2019) [149]	regression	CART	JRC Global Surface Water, Landsat 5, LiDAR DTM, USGS National Elevation Dataset	Australia, United States
Markert et al. (2020) [150]	segmentation	Bmax Otsu thresholding, Edge Otsu thresholding	MERIT DEM, PlanetScope, Sentinel-1 GRD	Myanmar, Cambodia
Wang et al. (2020) [151] *	classification	MNDWI, MSCNN, RF	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	China
Peterson et al. (2020) [152]	regression	DNN, ELR, MLR, SVR	GREON, Landsat 8, Sentinel-2	United States
Wang et al. (2020) [153]	regression	CART	JRC Global Surface Water, Landsat 5, LiDAR DTM, USGS National Elevation Dataset	Australia, United States
Boothroyd et al. (2021) [154]	classification	RivaMap	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Philippines
Weber et al. (2020) [155]	regression	maximum likelihood, multiple linear regression, RF, SVM	NAIP, National Hydrography Dataset, NLCD, National Wetland Inventory, Sentinel-2	United States
Mayer et al. (2021) [156] *	segmentation	U-Net	JRC Global Surface Water datasets, PlanetScope, Sentinel-1	Cambodia
Li et al. (2021) [157]	classification	NDWI, MNDWI, MuWI-R, Otsu thresholding, SVM	Sentinel-2	Sri Lanka
Li and Niu (2022) [158]	classification	RF	ALOS DSM, China Lake Dataset, China Wetlands Map, Google Earth, Global Reservoir and Dam Database, Global Surface Water, Sentinel-1, Sentinel-2, SRTM DEM	China

below summarizes those studies and a word cloud generated from the titles and keywords of the 18 papers is provided in Figure 10. The most frequently used keywords are “Google Earth Engine”, “surface water”, and “change”. Change detection is very important in water mapping research, though it requires the ability to first map the water. Sentinel-1 data are used almost on par with Landsat data in the water mapping literature, thus the size of these two terms is almost the same. From our interactive web app (see Appendix A) and Table 5, the most-used RS datasets are Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI. The most popular models are RF, multiple linear regression, Modified Normalized Difference Water Index (MNDWI), and SVM, and the most-used evaluation metrics are R², Kappa, and OA. A brief summary of those studies is provided right below Table 5. More detailed textual summaries for most of the reviewed water mapping studies are provided in Appendix C.5.

Static surface water maps are often produced at the regional or national level, but do not show long-term trends resulting from seasonality or global warming’s effects. In [32], the authors created a web portal using GEE as a backend alongside an expert system to identify bodies of water in Landsat imagery. RS has been widely used to map and monitor surface water. In [142], the authors used all available Landsat images to study surface water dynamics in Oklahoma from 1984 to 2015. The authors in [142] found significant inter-annual variations in the number of surface water bodies and surface water areas. They also found that both the number of surface water bodies and surface water areas had a positive relationship with precipitation and a negative relationship with temperature.

Floods and heavy precipitation events often occur at times of heavy cloud cover, making optical imagery not well-suited to water mapping or flood monitoring during those times. Traditionally, ground-based gauges are used to monitor water level and stream flow, but only work at specific points, limiting their utility during large-scale flood events. SAR imagery, however, is often used in water mapping or flood monitoring analyses because of its ability to see through clouds and work over large spatial scales. This is especially important for monsoonal regions like Southeast Asia where intense rains can lead to flood conditions. However, SAR imagery is also susceptible to classification errors when flooding occurs under tree cover or looks like concrete/pavement in urban areas, so preprocessing steps should be carefully considered. The authors in [150] analyzed to what degree different preprocessing steps affect the output water maps using both SAR and DEM data and two variations of Otsu’s thresholding algorithm. Glacial lake outburst floods (GLOF) are one of the serious natural hazards in the Himalayan region. To reduce the potential risks of GLOF, the information about the location and spatial distribution of glacial lakes is critical. In [143], the authors used Landsat 8 images available on GEE to map glacial lakes in the Tibet Plateau region. Their results revealed that climate warming played a major role in glacial lake changes.

Categorizing urban water resources faces two main challenges. First, it is often difficult to distinguish between water and things like asphalt or shadows in urban settings using RS imagery. Second, the distribution of water resources has changed alongside the accelerating impacts of climate change, making up-to-date, temporally aware water monitoring difficult. GEE provides free data storage, datasets, and compute, but as of yet high-accuracy DL models like NNs are not available on the platform. In [151], the authors compared the performance of MNDWI and an RF to that of a multi-scale CNN (MSCNN) and showed that the DL method was the most accurate (with less false classifications) for identifying urban water resources in several Chinese cities. While DL receives a lot of attention in water mapping research, these models still require a lot of input data and large amounts of compute to train them. However, as compute becomes publicly available in cloud-based platforms like GEE, obtaining large amounts of labeled training data remains a key bottleneck to using DL models. One way to make the data labeling process less time- and resource-intensive was illustrated in [156], where the authors used current water maps and a segmentation algorithm to automatically collect data labels from Sentinel-1 imagery.

Optical imagery used in surface water mapping analyses is often occluded by clouds, and many common methods used to map surface water confuse snow, ice, rock, and shadows as water. DeepWaterMapv2 was released in [147] and aimed to address these false positive misclassifications.

ML models have achieved high levels of accuracy in identifying water bodies in RS imagery. However, the models often misclassify soil, rock, clouds, ice, and shadow as water and often rely on cloud-free, optical RS imagery, which is not always available. The authors in [157] used masking, filtering, and segmentation algorithms to identify bodies of water in Sri Lanka in complex, mountainous environments. It is challenging to repeatedly produce up-to-date, accurate surface water maps over large areas. Water bodies change their shape and overall distribution through time, and humans use water in ways that look dissimilar to natural water bodies in RS imagery. Most studies to date focus on one type of water body (lakes, rivers, etc.) or create a binary classification mask giving little to no detail on various water body classification types. To explore the potential to distinguish between surface water body subtypes, [158] used slope, shape, phenology, and flooding information as input to an RF model to predict for lakes, reservoirs, rivers, wetlands, rice fields, and agricultural ponds.

The authors in [144] proposed a new method for quickly mapping yearly minimal and maximal surface water extents. In [148], the authors integrated global surface water (GSW) dataset and SRTM-DEM to determine the spatiotemporal patterns of water storage changes in China’s lakes and reservoirs. Multitemporal, multispectral satellite observations from the Landsat program and Sentinel constellation are particularly useful in fluvial geomorphology, in which river channel mapping and the analysis of planimetric change have long been a focus. The authors in [154] demonstrated a workflow showing how GEE can be used to extract active river channel masks from a section of the Cagayan River (Luzon, Philippines).

Satellite RS can be used to estimate chromophoric dissolved organic matter (CDOM) as a riverine constituent that influences optical properties in surface waters. CDOM absorption is a common proxy for dissolved organic carbon (DOC) concentrations in inland waters, including Arctic rivers. The authors in [146] stated that this was the first study using GEE for RS of water quality parameters in inland waters. Collecting field data for monitoring water quality can be costly in terms of money, time, and effort. Additionally, traditional monitoring techniques do not extend over a large area and are often difficult to repeat over time. Satellite RS imagery can help monitor water quality at frequent intervals over large areas. To estimate water quality parameters like chlorophyll-a (Chl-a) concentrations, turbidity, and dissolved organic matter, [152] used ML and DL models to analyze RS imagery. Harmful algal blooms (HABs) have become a serious issue in freshwater ecosystems. RS has proven to be a cost-effective means for monitoring HABs. The authors in [153] developed a methodological framework for mapping Chl-a concentrations with multi-sensor satellite observations and in-situ water quality samples.

3.2.6. Wetland Mapping

Wetland mapping is one of most well-developed applications using GEE and AI (16 studies total).

Table 6. Studies targeting wetland mapping from RS imagery using AI.

References	Method	Model Comparison	RS Data Type	Study Area
Hird et al. (2017) [35]	classification	BRT	LiDAR DTM, Sentinel-1, Sentinel-2	Canada
Farda (2017) [159]	classification	CART, Fast NB, GMO Max Entropy, IKPamir, MLP, Margin SVM, Pegasos, RF, Voting SVM, Winnow	Landsat 3 MMS, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, ASTER GDEM	Indonesia
Amani et al. (2019) [160]	classification, segmentation	RF, SNIC	Landsat 8	Canada
Mahdianpari et al. (2018) [161]	classification	RF	Sentinel-1, Sentinel-2	Canada
DeLancey et al. (2019) [162]	classification	BRT	LiDAR DEM, Sentinel-1, Sentinel-2, SRTM DEM	Canada
Wu et al. (2019) [163]	classification	k-means	NAIP, JRC Global Surface Water datasets, LiDAR DEMs, National Wetlands Inventory (NWI)	Canada, United States
Amani et al. (2019) [17]	classification	RF	Canadian DEM, Landsat 8, Sentinel-1	Canada
Zhang et al. (2019) [164]	classification	RF	Google Earth Pro, Landsat 8 OLI	China
Mahdianpari et al. (2020) [165]	classification, segmentation	RF, SNIC	aerial photography, Google Earth, Sentinel-1, Sentinel-2	Canada
Hakdaoui et al. (2020) [166]	classification	RF	ASTER DEM, Landsat 5 TM, Sentinel-1 GRD, Sentinel-2 MSI	Morocco
DeLancey et al. (2019) [167]	classification	U-Net, XGBoost	ALOS DEM, Sentinel-1, Sentinel-2	Canada
Mahdianpari et al. (2020) [168]	classification	RF	Canada’s Annual Crop Inventory, Google Earth, Pleiades, Sentinel-1, Sentinel-2, WorldView 2	Canada
Wang et al. (2020) [169]	classification	DT	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	China
Mahdianpari et al. (2020 [170]	classification	CART, MD, RF	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Canada
Sahour et al. (2021) [171]	classification	RF, SVM	aerial photography, Google Earth, JRC Global Surface Water, Sentinel-1, Sentinel-2	United States
Jia et al. (2021) [172]	segmentation	Otsu’s thresholding algorithm	DJI Phantom 4 pro, Gaofen-2, Google Earth, Sentinel 2	China

below summarizes those studies and a word cloud generated from the titles and keywords of the 20 papers is provided in Figure 11. The most frequently used words are “Google Earth Engine”. Wetlands have many different subtypes and occur in both inland and coastal environments, so we can see “wetlands” alongside terms like “tidal flats” and “coastal”. Most of the wetland mapping studies we reviewed take place in Canada and combine high-resolution RS imagery like Sentinel-1 and Sentinel-2 imagery to better distinguish between water and aquatic vegetation. From our interactive web app (see Appendix A) and Table 6, the most-used RS datasets in those studies are Sentinel-1, Sentinel-2, and Google Earth. The most popular models used are RF, boosted regression trees (BRT), CART, and Simple Non-Iterative Clustering (SNIC), and the mostly used evaluation metrics are OA, Kappa, PA, and UA. A brief summary of those studies is provided right below Table 6. More detailed textual summaries for almost all reviewed wetland mapping studies are provided in Appendix C.6.

Wetland serves as the globally biggest carbon pool, and thus has important ecological service functions (e.g., water conservation, regulation, and maintenance of species diversity) [173,174,175]. Global climate change and human activities have posed dramatic challenges in the past few decades to wetland ecosystems, and wetland mapping is essential to conserve and manage terrestrial ecosystems [176]. RS makes investigating large wetland systems and monitoring their change over time possible [177].

Wetlands are highly dynamic landscapes, often making past efforts to map them out-of-date. This is especially true at the regional or national level, where it is often difficult to monitor wetlands at scale due to their remote location and large spatial scale. While there are efforts to monitor wetlands in Canada at the sub-regional and -province level, this is mostly through governmental efforts to produce static maps. Cloud computing on GEE was utilized in [35] to create an open-source, reproducible map of wetland occurrence probability using LiDAR and RS data for the entire area of Alberta. Mapping subtypes of wetlands is difficult because while they look similar in RS imagery, they are diverse environments that cover a wide area. The same is true for classifying peatlands, a subtype of wetlands, which cover large geographic areas in complex patterns. This is problematic because peatlands, like wetlands, provide critical habitats that promote biodiversity while also being a global carbon sink. Past studies have shown that while optical data are useful for peatland mapping, it is often occluded by clouds or other atmospheric conditions. SAR data, on the other hand, can detect bodies of water and vegetation at any time of day or night, but are prone to being noisy due to surface moisture content and roughness. The authors in [162] demonstrated that by combining SAR, optical, and LiDAR data on the GEE platform, a BRT model was able to predict peatland occurrence across Alberta province with relatively high accuracy at high resolution.

Due to the difficulties in producing wetland inventory maps, either from lack of field data or the challenge of recognizing wetlands because of their heterogeneous and fragmented nature, these maps are often only produced at a local level. Furthermore, because of the many local efforts to produce these maps, wetland inventories are often produced with different datasets and different methods, limiting the ability of interested parties/stakeholders to compare or combine maps. Anthropogenic activities are meanwhile converting these wetlands into agricultural or urban landscapes, in addition to natural rain and flooding events changing their spatial makeup. Thus, it is more important than ever to be able to produce wetland inventory class maps in order to monitor and protect existing wetlands. The authors in [161] used optical and SAR RS imagery to produce a 10 m resolution wetland map for the entire province of Newfoundland, Canada, using both an RF model and SNIC. Mapping environmental features like wetlands is the first step in being able to make informed decisions about conservation and restoration projects. However, more relevant to policymakers is how environments change over time. This information would allow them to isolate how human activity has changed wetlands during different periods. The authors in [170] classified wetlands in Newfoundland during three different periods to show the spatial dynamics of these ecosystems. There have been several attempts to produce wetland inventory maps in Canada on a large scale, although they often lack high spatial resolution and the ability to distinguish between wetland sub-types. There is also the issue of a lack of ground-truth field data, a common problem in ML applications in EO (there is overwhelmingly more unlabeled data than labeled data). Using field data collected from one Canadian province was proposed in [17] to create wetland inventory maps for several others using a mix of optical, SAR, and digital elevation data.

Across Canada, wetland mapping is a well-studied phenomenon. However, different local and regional agency wetland inventories use different techniques for monitoring wetlands or have altogether different definitions of what constitutes a wetland. Thus, even though several large-scale wetland maps have been produced, they are often not directly comparable. Additionally, these maps are often static and do not continually monitor wetlands through time. However, these are not the only barriers to mapping wetlands using RS imagery [165]. Others include obtaining sufficient and recent field data to verify wetland monitoring products, but also the difficulty of monitoring such dynamic landscapes. Wetlands do not have clear-cut boundaries, are extremely diverse landscapes and ecosystems, and are often in flux throughout seasons and years due to flooding and drying. The authors use optical and SAR Sentinel data in addition to field samples over the entirety of Canada and show that almost one-fifth of Canada is covered in wetlands. The study in [165] produced a high-resolution (10-m) wetland inventory map of Canada (an approximate area of one billion hectares), using multi-year, multi-source (Sentinel-1 and Sentinel-2) RS data on the GEE platform. Wetlands provide a variety of ecological services and are a key habitat for many species. Human activity has significantly disturbed wetlands as they are drained for urban or agricultural development. However, monitoring their health is challenging because it would require taking repeated field measurements over wide areas. Researchers have used ML and RS data to do so, but the large amount of compute needed to map wetlands is often prohibitive. The authors in [160] analyzed a large number of field samples alongside Landsat imagery with an RF model to produce a wetland map for all of Canada. Wetland mapping and monitoring have been a challenging issue for the RS community during the past decades. Compared with the United States with the National Wetlands Inventory, Canada has been lacking a national wetland inventory until recently. The authors in [168] proposed an object-based classification method to classify Sentinel-1 and Sentinel-2 data on the GEE cloud-computer platform, which resulted in the 10-m Canadian Wetland Inventory.

Large, inundated wetlands can be effectively mapped using RS imagery. Small wetlands or wetlands that are inundated only part of the time are much more difficult to identify. Yet, it is more important to do so now than ever given that wetlands are rapidly being converted for agricultural use or are drying up due to climate-induced drying. Monitoring wetlands at large scales is possible, however, with the help of automated techniques like ML. For example, NAIP imagery and LiDAR derived DEM data were used in [163] to detect wetlands across the northern United States using unsupervised classification on the GEE platform. Being able to identify wetlands in RS imagery is the first step towards monitoring their health or decline in a new climate regime, and to make policy choices based on this information. To this end, spatially high-resolution sensors like LiDAR or data products like NAIP can help researchers identify wetlands in RS imagery but are not collected often enough to map wetlands at a fine temporal resolution. This is problematic because wetlands are dynamic ecosystems; they can be both wet and dry over the course of the same season. To get around this limitation, Sentinel-1 and 2 imagery were combined in [171] with aerial photographs and field data to map the spatial variation of wetlands in portions of the United States over time. Environmental problems are often associated with land-use changes, but these changes are not solely linked to urban expansion. Land use change also negatively affects areas like coastal wetlands, which are not monitored as regularly. The possibility of using GEE to map coastal wetlands in Indonesia was explored in [159] by comparing all of the different classifiers on the platform and how they perform with Landsat, digital elevation, and Haralick texture data. The authors showed that in all cases, ML models did much better at binary than multi-class classification.

Tidal flats, often referred to as coastal non-vegetated areas, are dynamic ecosystems, both due to their natural rhythms of water advance and retreat, but also due to anthropogenic change and rising sea levels. It is difficult to monitor tidal flats without the use of multi-temporal, high-resolution RS imagery because of how they change through time. With Landsat 8 and high-resolution Google Earth imagery, an RF model was used in [164] on GEE to classify tidal flat types and their distribution in China. The authors reported very high classification rates across tidal flat classes. However, the authors detailed that satellites like Landsat did not fully capture tidal ranges. Coastal wetlands are usually composed of coastal vegetation areas and tidal flats. Coastal tidal flats are natural transitions from terrestrial ecosystems to ocean ecosystems and are vulnerable to anthropogenic activities and natural disturbances such as sea-level rise, land reclamation, and aquaculture. Many existing global land cover data products have a wetland layer, but do not explicitly differentiate coastal vegetation area and coastal tidal flats (no specific layer for coastal tidal flats). The authors in [169] developed a pixel- and frequency-based approach to generate annual maps of tidal flats at 30-m spatial resolution in China’s coastal zone using the Landsat TM/ETM+/OLI images and the GEE cloud computing platform. Tidal flats are unique ecosystems but are threatened due to human disturbances and climate change. Additionally, they are difficult to identify in RS imagery because satellite platforms cannot capture intertidal variability due to their infrequent return times. The authors in [172] addressed this limitation by first processing high-resolution RS and UAS imagery to map minimum and maximum water and vegetation extent. They used Otsu’s thresholding algorithm to automatically detect the best ratio for each index. These two indices were then combined in a composite that showed the total intertidal area in the RS imagery, to which the authors again applied the Otsu thresholding algorithm. The end result was a highly accurate map of tidal flats that did not require any post-processing.

Sebkhas are a type of salty, unvegetated wetland created when desert bodies of water become more salinated over time due to mechanisms of water loss such as evaporation. They are home to specific species of vegetation and fish that can survive in salinated environments, but their drainage networks are often underground, making them hard to identify in RS imagery. An RF model was used in [166] to identify water cavities where sebkhas form in Morocco. Wetland inventory maps are increasingly being used to inform carbon pricing, ecosystem service values, and conservation/restoration decisions. Thus, it is important to make a repeatable processing pipeline that can ingest, process, and visualize data on a day-to-day basis so that monitoring programs and reporting programs (like in a government setting) have up-to-date, accurate information. To this end, there have been many studies identifying wetlands using RS imagery and ML, yet most of them suffer from not being able to distinguish between wetland subtypes. This is a challenging issue because fens, peatlands, bogs, marshes, and swamps can have very different vegetation types and structure. It is important to be able to distinguish between them because they each respond differently to human disturbance and changes in climate. The authors in [167] compared the performance of an XGBoost model to a CNN for wetland type classification.

3.2.7. Infrastructure and Building Detection, Urbanization Monitoring

Infrastructure, building detection, and urbanization monitoring is the 7th-most-well-developed application using GEE and AI (11 studies total).

Table 7. Studies targeting infrastructure and building detection from RS imagery using AI (Note that references marked * denotes novel methods and is detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Goldblatt et al. (2016) [178]	classification	CART, RF, SVM	Google Earth, Landsat 7 ETM+, Landsat 8, WorldPop	India
Huang et al. (2018) [179]	classification	BRT	Google Earth, Landsat 7 ETM+, Landsat 8 OLI	China
Xu et al. (2019) [180]	classification, segmentation	LandTrendr, RF	FROM-GLC, GHSL, Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	China
Zhong et al. (2019) [181]	regression	cubic regression	GPP, GOME-2, Google Earth Pro, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, MOD09A1	China
Lin et al. (2020) [182]	classification	RF	DMSP NTL, GHSL, GlobeLand30, Google Earth, Landsat 8, Sentinel-1, SRTM DEM, VIIRS NTL	China
Liu et al. (2020) [183]	classification, segmentation	CART, Otsu’s thresholding algorithm, RF	Geo-Wiki, GHSL, GlobeLand30, Google Earth, Hansen Global Forest Change, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, OpenStreetMap, SRTM DEM	China
Mugiraneza et al. (2020) [184]	classification, segmentation	LandTrendr, SVM	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Rwanda
Lin et al. (2021) [185] *	classification	CART, gmoMaxEnt, NB, RF, SVM	Landsat 8 OLI	China
Carneiro et al. (2021) [186]	classification	RF	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, Sentinel-2, SRTM DEM	Brazil
Zhang et al. (2021) [187]	classification	RF	Landsat 8 OLI	China
Samat et al. (2022) [188]	classification	SVM	GCL-FCS30-2020, GHSL, Google Earth, Sentinel-2	China

below summarizes those studies and a word cloud generated from the titles and keywords of those papers is provided in Figure 12. The most frequently used terms are “Google Earth Engine”, “urban”, “land”, “building”, “impervious”, etc. The vast majority of the studies in this domain take place in China and are both static mapping and change-detection applications. Infrastructure and urban area identification is often done by comparing these classes to other LULC classes, so we notice that “vegetation” and “forest” also appear in the word cloud. From our interactive web app (see Appendix A) and Table 7, most frequently used RS datasets are Landsat 8 OLI, Landsat 7 ETM+, and Google Earth. The most popular AI models are RF, CART, and SVM, and the most frequently used evaluation metrics are OA, Kappa, PA, and UA. A brief summary of those studies is provided below Table 7. More detailed textual summaries for some selected studies are provided in Appendix C.7.

Materials like parking lots, roads, and buildings (i.e., concrete, asphalt) can be classified as “impervious surfaces” in RS analyses and are often indicative of human development and urban extent. Impervious surfaces change the hydrological cycle and produce heat effects, affecting overall ecosystem health and well-being. To monitor these materials, researchers have tried using night-time lights to estimate their extent, but this process leads to overestimates as light scatters. To investigate how best to identify impervious materials in RS imagery regardless of cloud cover, the authors in [182] combined nighttime light, DEM, and SAR data and an RF model on GEE. Their resulting maps were more accurate than commonly used maps like GlobeLand30. The authors in [180] put forward a new scheme to conduct long-term monitoring of impervious−relevant land disturbances using Landsat archives.

While greenhouses are used to grow food and help ensure food security, their proliferation can have environmental consequences. Previous attempts to classify greenhouses from RS imagery as part of LULC research have focused on small-scale proof-of-concept applications and have not emphasized identifying the structures in complex terrain types. To explore the possibility of identifying greenhouses in RS imagery over a large area in China, an ensemble ML model was designed in [185] to distinguish them from water, forest, farmland, and construction sites. Urban green spaces have a multitude of benefits, such as regulating urban climate, improving air quality, and reducing stormwater. RS has proven useful for studying the landscape structure of urban green spaces. The authors in [179] assessed the impact of urban form on the landscape structure of urban green spaces in 262 cities in China. The results revealed that cities with a high road density tended to have a smaller area of urban green spaces and be more fragmented. In contrast, cities with complex terrains tended to have more fragmented urban green spaces.

Rapid urban expansion around the world has led to worsening human and ecosystem health, affecting forests, air and water pollution levels, and overall levels of biodiversity. However, the currently available maps for mapping urban settlements and their expansion are mostly static, whereas it would be more useful to have up-to-date information to be able to make better urban planning and land-use decisions. The authors in [186] designed a workflow for mapping urban sprawl over time in Brazil using an RF on the GEE platform. Increasing rates of urbanization put pressure on conservation targets and biodiversity levels as land previously occupied by ecosystems is converted into built-up areas. RS imagery makes it much easier for urban planners and researchers to monitor rates of urbanization and urban sprawl over wide areas. However, few labeled datasets are available using ML to identify buildings and built-up areas. To address this problem, a large, vectorized, ground-truth verified dataset was created in [178] in India in order to train different ML models on GEE. A semi-automatic large-scale and long-time-series (LSLTS) urban land mapping framework was demonstrated in [183] by integrating the crowdsourced OpenStreetMap (OSM) data with free Landsat images to generate annual urban land maps in the middle Yangtze River basin (MYRB) from 1987 to 2017.

Research on urbanization and urban sprawl will often focus on how urban spaces are replacing agricultural land and forested spaces. Vegetation maps, on the other hand, are often produced using “urban”, “built-up areas”, or “impervious surfaces” as classes to predict for, distinctly separating vegetation and zones of human inhabitation. Much less work has gone into monitoring vegetation prevalence and distribution within urban spaces themselves. This is an important and timely research topic given the environmental and psychological benefits people get from having access to green spaces within cities, such as stress reduction, better air quality, and lower temperatures. Using different vegetative indices (EVI, Gross Primary Production, etc.) derived from Landsat and MODIS data, the authors in [181] showed that urban sprawl in Shanghai had increased significantly in the last decade and a half.

3.2.8. Wildfires and Burned Area

Wildfires and burned areas have less than 10 studies using GEE and AI (eight studies total).

Table 8. Studies targeting wildfires from RS imagery using AI. (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Parks et al. (2019) [189]	regression	RF	Landsat 4 TM, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Canada, United States
Quintero et al. (2019) [190]	segmentation	FormaTrend, LandTrendr	Landsat 5 TM, Landsat ETM+, Landsat OLI, MCD64A1, SRTM DEM	Spain
Long et al. (2019) [191] *	classification	RF, SVM	CBERS-4 MUX, FireCCI51, Gaofen-1 WFV, GFED4, Google Earth, MCD12C1, MOD44B, MTBS, Landsat-8	Global
Bar et al. (2020) [192]	classification	CART, RF, SVM, Weka clustering	FireCCI51, IRS 1C, Landsat 5, Landsat 8 OLI, MODIS, ResourceSat 2, Sentinel-2, VIIRS	India
Sulova and Jokar Arsanjani (2020) [193]	classification	CART, NB, RF	CGLS-LC100, FIRMS, MOD13Q1, Sentinel-2, SRTM DEM	Australia
Zhang et al. (2020) [194]	classification	RF	Landsat 5	Global
Seydi et al. (2021) [195]	classification	KNN, RF, SVM	Landsat 8, MODIS, Sentinel 2	Australia
Arruda et al. (2021) [196]	classification	DNN	INPE, Landsat 8 OLI, MODIS	Brazil

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the eight papers is provided in Figure 13. The most frequently used words are “Google Earth Engine”, “burn/ed”, “fire(s)”, “forest”, and “change”, which show that the main focus of these studies is on monitoring forested areas pre- and post-fire. The data products “Sentinel-2”, “Landsat”, and MODIS maps like “MCD64A1” (a burned-area product) indicate that wildfire and burned area mapping analyses rely mainly on optical imagery. From our interactive web app (see Appendix A) and Table 3, the mostly used RS datasets are Landsat 8 OLI, and MODIS. The most popular AI models are RF, SVM, and CART, and the most-used evaluation metrics are OA, commission error (CE), Kappa, omission error (OE), and R². A brief summary of those eight studies is provided below Table 8. More detailed textual summaries for each of the eight studies are detailed in Appendix C.8.

Wildfires cause damage to ecosystems and human health, in addition to releasing greenhouse gasses when they burn. Climate change increases the number of wildfires across the globe. The recent massive wildfires, which hit Australia during the 2019–2020 summer season, raised questions to what extent the risk of wildfires can be linked to various climate, environmental, topographical, and social factors and how to predict fire occurrences to take preventive measures. An automated and cloud-based workflow was developed in [193] for generating a training dataset of fire events at a continental level using freely available RS data on GEE. Landscape fires have been a major natural hazard affecting West-Central Spain, and therefore, it is critical to be able to map and characterize landscape fires. Using the LandTrendr (Landsat-based Detection of Trends in Disturbance and Recovery) and FormaTrend (Forest Monitoring for Action—Trend) algorithms on the GEE cloud-computing platform, a method was proposed in [190] for identifying fire-induced disturbances. Wildfires are a common occurrence in the Brazilian Cerrado, often determining and changing the natural plant species in burn cycles. However, the Cerrado has been undergoing increasing anthropogenic conversion into cropland and pastures, which has changed hydrological and biogeochemical cycles within this ecosystem. This in turn has led to changes in fire size, pattern, frequency, and severity, so it is more important than ever that methods to quickly and reproducibly monitor the fire landscape within savannah are created. A completely cloud-based DL workflow combining Google Cloud and GEE was designed in [196] to classify burn scar areas in Brazil.

Traditional wildfire mapping field surveys and digitization efforts are time-consuming and hard to reproduce over time. Burned area indices can be created to monitor post-fire landscapes and their subsequent recovery, but their thresholds are not dynamic and so perform differently in different locations. Sentinel-2 data was used in [195], along with two different burn areas and LULC maps to train different ML classifiers (k-nearest neighbor (KNN), RF, SVM) to map wildfire damage in Australia. As the planet warms, forest fires are increasing in occurrence and severity. This has negative consequences for ecosystems, biodiversity, and human health. To estimate the damage caused by forest fires and their subsequent recovery rates, RS imagery is needed to monitor forests and burn scars over large areas. However, to date, most fire products are created with coarse RS imagery, making regional and local fire monitoring difficult. To determine the impact of using higher-resolution RS data products, how Landsat and Sentinel optical imagery affected an ML model’s performance in burn area classification was compared in [192].

Burned area maps showing where wildfires have occurred are important in being able to analyze global wildfire trends. However, many burned area maps derived from RS imagery are from the MODIS platform. The 250 m spatial resolution of products like FireCCI51 leave out a lot of detail, so the authors in [191] used CBERS, Gaofen, and Landsat imagery to create a 30 m burned-area dataset for 2015. However, the authors noted that their method had difficulty recognizing burned areas from recently plowed fields in agricultural areas, so crop-type masks should be used to remove potential false positives. Additionally, Landsat data was used for both the data collection and validation stage. Thus, the authors were not able to assess the suitability of using Landsat imagery for data collection purposes despite their high accuracy rates. Later on, [194] adapted the exact same processing steps on GEE to produce a burned area map for the year 2005, illustrating how sharing and storing code on GEE makes it easy to re-run analyses or adapt them for new use cases.

Satellite-derived spectral indices such as the relativized burn ratio (RBR) allow fire severity maps to be produced across multiple fires and broad spatial extents. In order to better interpret the fire severity in terms of on-the-ground fire effects compared to non-standardized spectral indices, [189] produced a map of composite burn index (CBI), a frequently used field-based measure of fire severity.

3.2.9. Heavy Industry and Pollution Monitoring

There are seven studies about heavy industry and pollution monitoring using GEE and AI.

Table 9. Studies targeting heavy industry and pollution from RS imagery using AI.

References	Method	Model Comparison	RS Data Type	Study Area
Waller et al. (2018) [197]	regression	RF	DART, Google Earth Pro, Landsat 5 TM, NLCD	United States
Lobo et al. (2018) [198]	classification	CART	RapidEye, Sentinel 2	Brazil
Xiao et al. (2020) [199]	segmentation	LandTrendr	Google Earth, Landsat 5, Landsat 7, Landsat 8	Mongolia
Balaniuk et al. (2020) [200]	classification	CNN	Sentinel-2	Brazil
Fuentes et al. (2020) [201]	classification	RF	Landsat 5, Landsat 8, Natural Resource Canada DEM, Sentinel-1, Sentinel-2	Canada
He et al. (2020) [202]	segmentation	LandTrendr	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	China
Zhou et al. (2021) [203]	classification	CART, RF, SVM	Sentinel-2	China

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the seven papers is provided in Figure 14. The most frequently used words form the phrase “Google Earth Engine”. Most applications in this area are focused on monitoring reclamation or pollution at active or previous mining sites, so “mine” and “mining” feature prominently in this word cloud. The algorithm LandTrendr was used by several papers after identifying mine sites to monitor pollution and water levels or vegetation changes through time. From our interactive web app (see Appendix A) and Table 9, the most-used RS datasets are Sentinel-2, Landsat 8 OLI, Landsat 5 TM, Landsat 5, and Google Earth. The most popular models are RF, CART, and LandTrendr, and the most-used evaluation metrics are Kappa, OA, PA, UA. A brief summary of those seven studies is provided below Table 9. More detailed textual summaries for each of the seven studies are detailed in Appendix C.9.

Mining can lead to lots of environmental degradation during the actual mining process itself, but often continues to do so if mines are not properly reclaimed after the mine is no longer active. Field techniques for monitoring environmental damage operate on a limited spatial and temporal scale, failing to fully capture what is happening. RS can help monitor ecological changes during mining and ensure that mining companies clean up after mining has stopped during the reclamation process. A mapping study was performed in [198] for mining areas in the Brazilian Amazon using Sentinel-2A images and the CART classifier in GEE. To monitor mining disturbances at a coalfield in Mongolia, the LandTrendr algorithm was used in [199] to analyze Landsat data. The authors designed a fast, efficient method on the GEE platform to monitor surface mining operations and show that only 26% of promised reclamation was undertaken at the Shengli Coalfield. Heavy industry projects like mining normally require reclamation after the fact to ensure that local ecosystems can heal and regenerate. Monitoring sites that have undergone mining is made much easier with RS imagery because they are often large, spatially distributed ecological disturbances. This is especially the case for underground mining projects where subsidence occurs but is difficult to track without an aerial view. Landsat imagery and the LandTrendr algorithm were utilized in [202] to monitor water accumulation in subsidence areas of past mining in China. Mining is economically important because of the many jobs and resultant materials it provides but is associated with various environmental and health risks. One such danger comes from the failure of tailings dams, which store water with toxic levels of waste solids. Even though these failures can cause significant damage to the environment, human health, and infrastructure, there is not a global database containing active tailings dams. This in turn can make it easier for illegal mines to operate as legal mining operations with tailings dams are not heavily monitored. In order to keep track of mines and dams in Brazil, two different CNNs were used in [200] to first classify potential mining sites and then to classify perceived/potential environmental risk.

As cities expand and develop, construction and demolition waste is often stored until it can be further processed, reused, or gotten rid of. Sometimes these waste piles are orderly and are trackable, but many are not, making it hard to manage them and their potential negative environmental or social effects. Current methods to take stock of waste piles and dump sites rely on field investigations, which take a lot of time, effort, and money to produce. More work needs to be done to identify them using RS imagery and ML methods, but tuning different ML methods and their respective parameters can lead to different results. To test the efficacy of different ML algorithms for identifying waste and dump sites in optical imagery, the parameters for the CART, RF, and SVM algorithms available on GEE were optimized in [203].

Oil and gas pads are developed for production and then capped, reclaimed, and left to recover when no longer productive. Understanding the rates, controls, and degree of recovery of these reclaimed well sites to a state similar to pre-development conditions is critical for energy development and land management decision processes. The authors in [197] used time series data of the Soil Adjusted Total Vegetation Index (SATVI), calculated from Landsat 5 imagery, to track changes and assess vegetation regrowth on 365 abandoned well pads located across the Colorado Plateau. Previous estimates of particulate matter for the Canadian Air Pollutant Emissions Inventory (APEI) were based on the exposed mine disturbance areas that had been calculated using outdated mine area extents. With GEE JavaScript API, RF classifiers were used in [201] to produce maps of mine waste extents with Landsat-8 and Sentinel-1 and Sentinel-2 archives.

3.2.10. Climate and Meteorology

There are seven climate and meteorology studies using GEE and AI.

Table 10. Studies targeting climate and meteorology studies.

References	Method	Model Comparison	RS Data Type	Study Area
Chrysoulakis et al. (2019) [204]	regression	polynomial regression	MCD43A1, MCD43A2, MOD09CMA	Global
Chastain et al. (2019) [205]	regression	major axis regression	Landsat 7 ETM+, Landsat 8 OLI, Sentinel-2 MSI	France, Portugal, Spain, United States
Demuzere et al. (2019) [206]	classification	RF	DMSP-OLS NTL, Global Forest Canopy Height, Landsat 8, Sentinel-1, Sentinel-2	Australia, Brazil, Canada, China, France, Japan, Mexico, Poland, Singapore, Spain, Sudan, United Kingdom, United States
Ranagalage et al. (2019) [207]	classification	SVM	Landsat 5, Landsat 8	Sri Lanka
Medina-Lopez and Ureña-Fuentes (2019) [208]	regression	DNN	Sentinel-2	Global
Besnard et al. (2019) [209]	regression	LSTM, RF	Landsat 4, Landsat 5, Landsat 7, Landsat 8, MCD43A4	Global
Elnashar et al. (2020) [210]	regression	ANN, GBR, SVR	MCD12Q1, MOD13A2, SRTM DEM	China, India, Myanmar, Thailand, Vietnam

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of those papers is provided in Figure 15. The most frequently used words are “Google Earth Engine”, “changes” and “climate”. There are also specific keywords for ocean-related studies like “sea” and “salinity”, and studies focused more on land and atmosphere characteristics like “surface”, “land”, “temperature”, “LST” and “albedo”. From our interactive web app (see Appendix A) and Table 10, the most-used RS datasets are Landsat 8, Landsat 5, and Sentinel-2. The most popular AI models are RF, and the most frequently used evaluation metrics are mean absolute error (MAE), OA, root mean square error (RMSE), R². A brief summary of those studies is provided below Table 10. More detailed textual summaries for each of the seven studies are detailed in Appendix C.10.

Forests store much of the world’s terrestrial carbon, but globally they are under threat due to the effects of global warming and human disturbance. While forests release carbon immediately when they are cut down or otherwise disturbed, they also release carbon through secondary effects. This type of climate “memory” or lag in carbon flux is much less studied and so not well-known. To study this mechanism further, the authors in [209] used an LSTM and compared the performance to an RF for carbon fluxes in global forests.

Accurate satellite-derived albedo estimations are needed to parameterize and in turn to validate climate simulation models. MODIS satellite observations from 2000 to 2015 were analyzed in [204] using GEE to derive global snow-free land surface albedo estimations and trends at a 500 m resolution. A method was presented in [208] to obtain high-resolution sea surface salinity (SSS) and temperature (SST) by using Sentinel-2 Level 1-C Top of Atmosphere reflectance data. The consistency between Tropical Rainfall Measuring Mission (TRMM) multi-satellite precipitation and monthly gauged precipitation has been confirmed worldwide. A downscaling framework (from 25 km to 1 km) was proposed in [210] for TRMM precipitation products by integrating GEE and Google Colaboratory (Colab).

Furthermore, 30-m Landsat imagery has a long history of coverage between the seven ETM+ and eight OLI sensors. Sentinel-2 Multispectral Instrument (MSI) imagery has a higher resolution of 10-m and faster revisit frequency (10 days instead of 16 days for Landsat). Being able to use all of these sensors together for a given EO analysis would greatly increase the available spatial and temporal resolution, but the sensors have differences that need to be calibrated before they can be integrated. Still, this is one of the most-requested datasets we found in our review. Major-axis regression was performed in [205] on these datasets in pairs (seven ETM+/8 OLI, 7 ETM+/2 MSI, and eight OLI/2 MSI) across the entire coterminous United States and they were able to determine cross-platform correction coefficients for the Blue, Green, Red, NIR, and SWIR bands present in all three satellites.

Urbanization has changed the urban landscape and resulted in increasing land surface temperature (LST). In [207], the authors investigated the impacts of landscape changes on LST intensity (LSTI) in a tropical mountain city in Sri Lanka. There are several ongoing attempts to classify cities around the world based on various characteristics like urban canopy cover, total built-up area, neighborhood sizes, and urban heat island effects (for example, see Urban Atlas, World Urban Database Access and Portal Tools (WUDAPT)). These datasets can help planners and policymakers make more informed decisions as they consider implementing sustainability measures in their respective cities. However, these types of spatial datasets often rely on surveying methods that need to be continually updated. A cloud-based workflow was implemented in [206] and compared to the traditional method of using SAGA GIS for producing local climate zone city maps based on data like WUDAPT.

3.2.11. Disaster Management

Disaster management has six studies using GEE and AI.

Table 11. Studies targeting disaster management from RS imagery using AI.

References	Method	Model Comparison	RS Data Type	Study Area
Yu et al. (2018) [211]	classification	RF	Landsat 5 TM, Landsat 7, Landsat 8, SRTM DEM	Nepal
Cho et al. (2019) [212]	classification	RF	Landsat 7 ETM+, Landsat 8 OLI, MODIS Terra, Sentinel-1, SMOS	United States
Uddin et al. (2019) [213]	segmentation	CART, GEOBIA	Google Earth, Landsat 8, Sentinel-1, SRTM DEM	Bangladesh
Vanama et al. (2020) [214]	segmentation	Otsu’s thresholding algorithm	Global Precipitation Measurement satellite data, JRC Global Surface Water dataset, Landsat 8, Sentinel-1 GRD, Sentinel-2, WorldView 3	India
Ghaffarian et al. (2020) [215]	classification	RF	GeoEye 1, Google Earth Pro, Landsat 7 ETM+, Landsat 8 OLI, WorldView 1, WorldView 3	Philippines
Kakooei and Baleghi (2020) [216]	classification	CART, RF	NAIP, NOAA NGS Emergency Response	United States

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the six papers is provided in Figure 16. The most frequently used words are “Google Earth Engine”, “recovery”, “landslide”, “post-disaster”, “hurricane”, and “damage”. Many studies were focused on mapping buildings after flood or landslide events. One of the main challenges to doing disaster management research on GEE is that there is a delay to uploading RS data from the time it is recorded and the time it is uploaded to the platform. This limits the utility of doing time-sensitive research or deploying time-sensitive applications on GEE (i.e., the keywords “rapid”, “assess”). From our interactive web app (see Appendix A) and Table 11, the most-used RS dataset is Landsat 8. The most popular ML models are RF and CART, while the most frequent evaluation metrics used are OA, PA, and UA. A brief summary of those studies is provided below Table 11. More detailed textual summaries for each of the six studies are detailed in Appendix C.11.

RS imagery has long been used to monitor community recovery after natural disasters. Decision makers can use RS imagery and analyses to redirect resources during the recovery process. Even so, many studies focused on disaster recovery use VHR imagery that increases data storage and compute needs. To explore the suitability of GEE for disaster recovery, the authors in [215] used an RF model trained on Landsat imagery to do change detection on pre- and post-disaster areas in the Philippines. Building detections in post-disaster scenes are a valuable resource for timely assessing damages in disaster management. Using RGB images as input, an automatic building detection method was proposed in [216] to find buildings and their irregularities in pre- and post-disaster (sub-) meter resolution images.

Landslides are a major natural hazard in mountainous regions. Traditionally, landslide mapping heavily relies on field surveys and visual interpretation of satellite imagery. A new method was proposed in [211] for mapping landslides in Nepal using RF on GEE. Many agricultural landscapes have incorporated surface drainage systems to stop fields from flooding during heavy precipitation and runoff. These underground drainage networks have caused flood forecasting to become harder to do since it is more difficult to track water in space and time, as drainage networks are not always well mapped. The authors in [212] created surface drainage maps through running an RF model on the GEE platform, by analyzing vegetation, thermal, moisture, and climate datasets, along with surface drainage records.

Producing flood maps is critical to giving advanced warning to those who may be in affected areas. However, producing these maps in real time is hindered by the fact that many mapping applications focus on too small an area due to lack of computational resources. The authors in [214] presented a case study for the 2018 Kerala flood in India. They demonstrated how GEE can be used to process large optical and SAR RS datasets, in conjunction with field and precipitation data, using image processing techniques to produce high-resolution flood maps over a large area. Flood forecasting in Bangladesh currently involves running hydrological inundation simulations based on DEMs in order to produce early warning notifications. However, these simulations are compute-intensive and require access to high-resolution, up-to-date DEMs. Cloud-based mapping using RS imagery has the potential to provide quicker inundation forecasts over a large spatial area. This is especially true for analyses that utilize SAR imagery, since floods are often caused by heavy rains leaving optical imagery obstructed by clouds. The authors in [213] produced flood maps in Bangladesh, by taking advantage of the easy-to-find data and freely available compute on the GEE platform.

3.2.12. Soil

There are six soil studies using GEE and AI.

Table 12. Studies targeting soil from RS imagery using AI.

References	Method	Model Comparison	RS Data Type	Study Area
Padarian et al. (2015) [217]	classification, regression	CART, Serial Rifle Classifier	SRTM DEM	United States
Ivushkin et al. (2019) [218]	classification	CART, RF, SVM	Landsat 5 SR, Landsat 8 SR, SoilGrids	Global
Poppiel et al. (2019) [219]	regression	RF	ALOS DEM, Landsat 4 TM, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Brazil
Cao et al. (2019) [220]	regression	KNN, QRF, RF	LANDFIRE, Landsat 7	United States
Greifeneder et al. (2021) [221]	regression	GBRT	ASTER DEM, CGLS-LC100, GLDAS, Landsat 8 OLI, Landsat 8 TIRS, MOD13Q1, Sentinel-1, SRTM DEM	Global
Zhang et al. (2021) [222]	regression	ANN, RF, SVM	HydroSHEDS DEM, MODIS, Sentinel-2A	China

below summarizes those papers and Figure 17 below is the word cloud generated from the title, keywords, and abstract of the six soil studies. The most frequently used are characteristics that soil researchers try to monitor: “SOM”, “organic”, and “stocks” for soil organic matter, as well as “litter”, “moisture”, “thermal”, and “salinity”. MODIS and Sentinel-2 data feature heavily in the word cloud because they are the most-used data products (alongside Landsat). From our interactive web app (see Appendix A) and Table 12, we found that the most-used RS datasets are Landsat 8 OLI and SRTM DEM. The most popular ML models used are RF and CART, while the top evaluation metrics used are R² and RMSE. A brief summary of those studies is provided below Table 12. Much more detailed textual summaries for each of the six studies are detailed in Appendix C.12.

Many authors come to GEE curious to test out the new cloud computing platform for their domain-specific application. GEE provides freely available compute and data to interested researchers, which they then use to explore the strengths and limitations of GEE. An early soil mapping study was performed in [217] on GEE in 2015. Collecting field samples for soil mapping can be time- and labor-intensive and can be bound to small areas given their costs. These data collections also need to be repeated, representing a barrier to presenting up-to-date information that covers large spatial areas to decision-makers. To address these issues, the authors in [219] used field observations, DEM data, and Landsat imagery on GEE to map different soil types and soil attributes across a large region in Brazil.

Soil plays a critical role in the carbon and water cycles, along with providing areas for habitat or agricultural use. The spatial distribution of litter and soil carbon (C) stocks is important in greenhouse gas estimation and reporting and inform land management decisions, policy, and climate change mitigation strategies. The effects of spatial aggregation of climatic, biotic, topographic and soil variables on national estimates of litter and soil C stocks were explored in [220]. The authors also characterized the spatial distribution of litter and soil C stocks in the conterminous United States (CONUS). Litter and soil variables were measured on permanent sample plots from the National Forest Inventory (NFI) from 2000 to 2011. Beyond mapping litter and soil carbon (C) stocks, it is also important to map soil organic matter at a large scale, but traditional field collection techniques are cost- and effort-intensive. Many researchers have thus turned to RS imagery and/or ML to map soil organic matter, but there is still some difficulty in selecting the right input data or ML model for prediction. To determine how different datasets and ML models perform on GEE in predicting soil organic matter, an ANN, RF, and SVR model was compared in [222] with MODIS, Sentinel-2A, and DEM data as input.

Accurate soil moisture content information is crucial to being able to correctly model water, energy, and carbon cycles, as well as being key to understanding and predicting natural hazards like drought, floods, and landslides. However, most soil moisture datasets are created with medium or coarse spatial resolution. Using optical, thermal, and SAR imagery in addition to DEM data, a global, high-resolution soil moisture map was produced in [221]. The authors concluded that optical RS imagery and land-cover information play the most important roles in determining soil moisture content, but SAR imagery and soil data also contribute significantly to the model’s overall performance. This finding highlights other studies’ results ([95,161,182]) that the combination of optical and SAR data improves predictive outcomes. Soil salinity can impact agricultural yields and is a global issue, but current datasets like the Harmonized World Soil Database have low spatial resolution and need to be updated. As one of the main soil salinity datasets in use, this makes it difficult to estimate up-to-date soil salinity levels even as they change due to increasing drought severity from global warming. The authors in [218] explored GEE’s potential to make a global soil salinity map based on field data and Landsat thermal infrared imagery.

3.2.13. Cloud Detection and Masking

There are five studies (four are novel methods) that used GEE and AI for cloud detection and masking.

Table 13. Studies targeting cloud detection from RS imagery using AI (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Gómez-Chova et al. (2017) [223] *	regression	kernel ridge regression, linear regression	Landsat 8, RapidEye, SPOT 4	Argentina, China, Jordan, Spain
Mateo-García et al. (2018) [224] *	classification	ACCA, Fmask, k-means	Landsat 8 Biome Cloud Masks, Landsat 8 TOA	Global
Yin et al. (2020) [225] *	classification	DeepGEE-CD, FMask, RS_Net	Landsat 8 OLI	NS 1
Li et al. (2022) [226]	classification	Cloud-Score, Fmask, SVM, QA60	Sentinel-2	Amazon tropical forest, China, Sri Lanka
Zhang et al. (2022) [227] *	regression	DeepGEE-S2CR, DSen2-CR	SEN12MS-CR 2	Global

¹ Not Specified, ² this is a co-registered set of Sentinel-1 and Sentinel-2 images.

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of those papers is provided in Figure 18. The most frequently used words are “Google Earth Engine”, “cloud(s)”, and “masking”. The main task in this literature domain is masking clouds in optical imagery (which is not a problem for SAR data), so the words “optical”, “Landsat-8”, and “Sentinel-2” are also prominent in this word cloud. From our interactive web app (see Appendix A) and Table 13, the most-used RS datasets are Landsat 8 OLI, SRTM DEM, and Google Earth. The most popular models are Fmask and the most-used evaluation metrics are CE, OA, OE, RMSE. A brief summary of those studies is provided below Table 13. More detailed textual summaries for each of the five studies are detailed in Appendix C.13.

Many mapping and identification tasks that use RS imagery and ML rely on optical cloud-free imagery. Detecting and removing clouds in optical RS imagery is a difficult but important task, as many other classification and detection methods rely on masking clouds and on obtaining cloud-free imagery. Many algorithms, including Fmask, which is a commonly used algorithm to create a cloud mask in RS imagery, rely on using thresholds for single RS images, which makes them prone to errors when applied to entire RS time series. The authors in [223] treated cloud detection as a change detection problem across time using a kernel ridge regression model.

Optical RS imagery has many applications across several environmental and earth science domains. However, Optical RS imagery is often occluded by clouds, limiting its utility. While processing techniques like taking monthly composite images to remove clouds works to some extent, it relies on having enough cloud-free imagery to make the composites, which is not always available. Recently, DL models have shown the ability to reconstruct scenes in optical RS imagery that is blocked by clouds. However, researchers looking to use DL models in cloud environments often have to coordinate across different storage, analysis, and ML platforms (e.g., Google Cloud Storage, Google Colab, Google AI), which can be cumbersome and expensive. The authors in [227] thus decided to implement their cloud-removal DL model directly in GEE. Their model, DeepGEE-S2CR, is a cloud-optimized version of the DSen2-CR model presented in [228] and fuses co-registered Sentinel-1 and Sentinel-1-2 images from the SEN12MS-CR dataset.

Cloud detection is a well-studied task and GEE has several cloud detection/masking algorithms available on its platform. However, some of them have shown to be unstable leading to considerable under- or overestimation. To explore how CV algorithms and ML models can be used together on GEE, [226] combined the existing Cloud-Score algorithm with an SVM to detect clouds in imagery ranging from Amazon tropical forests, Hainan Island, and Sri Lanka. Fmask is the most commonly used method but has limited use in mountainous regions where terrain and shadows can be confused for clouds or when sudden changes in the Earth’s surface occur in time-series imagery. A convolutional neural network (CNN) called DeepGEE-CD was built in [225] to detect clouds in RS imagery directly on the GEE platform. Cloud screening may be cast as an unsupervised change detection problem in the temporal domain. A cloud screening method based on detecting abrupt changes along the time dimension was introduced in [224], assuming that image time series follow smooth variations over land (background) and abrupt changes are mainly due to the presence of clouds.

3.2.14. Wildlife and Animal Studies

Wildlife and animal studies is one of the less developed application domains using GEE and AI (four studies total).

Table 14. Studies targeting wildlife and animal studies.

References	Method	Model Comparison	RS Data Type	Study Area
Carrasco-Escobar et al. (2019) [229]	classification	RF	DJI Phantom 4 Pro, 3DR Solo	Peru
Ascensão et al. (2019) [230]	classification	binomial logistic regression	Hansen Global Forest Change, MCD12Q1, MOD13Q1	Brazil
Lyons et al. (2019) [231]	classification	RF	DJI Phantom 3 Professional	Australia
Pérez-Romero et al. (2019) [232]	classification, regression	KNN, RF	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	Spain

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the four studies is provided in Figure 19. The most frequently used words are “insect”, “bird”, “roadkill”, and “malaria” which show what scientists are trying to monitor in this domain. Meanwhile, specific place names like”, “forest”, “water”, “Peru”, and “Amazonian” reflect where the studies are being done. From our interactive web app (see Appendix A) and Table 14, the most popular AI model is an RF and the most-used evaluation metric is OA. A brief summary of those studies is provided below Table 14. More detailed textual summaries for each of the four studies are detailed in Appendix C.14.

UAS (i.e., drones) are able to collect high-quality data over large aggregations of wildlife, as they offer an attractive opportunity for improving methods and increasing cost effectiveness of monitoring wildlife populations. The authors in [229] explored the use of UAS for identifying Ny. darlingi breeding sites with high-resolution imagery (~0.02 m/pixel) and their multispectral profile in Amazonian Peru. Land use changes such as deforestation, irrigation, wetland modification and road construction, may drive infectious disease outbreaks and interfere with their transmission dynamics. Accurate classification of Ny. darlingi -positive and -negative water bodies would increase the impact of targeted mosquito control on aquatic life stages. Researchers in [231] developed a semi-automated framework for monitoring large complex wildlife aggregations using drone-acquired imagery over four large and complex waterbird colonies.

The success of conservation and mitigation management strategies may greatly depend on the knowledge of the temporal and spatial patterns of roadkill risk, and its relationship with key environmental drivers. The authors in [230] used a set of freely available environmental variables, namely habitat information from RS observations and climatic information from weather stations, to assess and predict the roadkill risk.

Pest outbreaks are causing more damage to forests around the world as winters get warmer and summers are drier and start earlier. These conditions allow pests to proliferate, though pests do not always kill trees outright. They often defoliate trees, which weakens them before future pest outbreaks or drought conditions. However, forest defoliation is understudied and much of the research done in this area relies on coarse resolution data. Using Landsat RS imagery, climate variables, and government environmental data, Ref.[232] analyzed Pine Processionary Moth outbreaks in pine forests in southern Spain.

3.2.15. Archaeology

Archaeology is also one of the less researched applications using GEE and AI (three studies total).

Table 15. Studies targeting archeology from RS imagery using AI. (Note that references marked * denotes novel methods and will be detailed in Section 3.3).

References	Method	Model Comparison	RS Data Type	Study Area
Liss et al. (2017) [233] *	classification	Canny edge detection, RF	WorldView 2	Jordan
Orengo and Garcia-Molsosa (2019) [234] *	classification	CART, RF, SVM	DJI Phantom 4 Pro	Greece
Orengo et al. (2020) [235] *	classification	RF	Google Earth, Sentinel-1, Sentinel-2 MSI, WorldView 2, WorldView 3	Pakistan

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the three papers is provided in Figure 20. The most frequently used words are “Google Earth Engine”, “detection”, “satellite”, “drone”, and “survey” while terms like “automated” and “mounds” are also common. This reflects the papers we reviewed and their focus on using the GEE platform to scale up and automate exploratory surveys using RS data, both from satellite platforms and self-collected drone imagery. From our interactive web app (see Appendix A) and Table 15, the most frequently used RS dataset is WorldView 2. The most popular ML model is an RF and the most-used evaluation metric is visual analysis.

Utilizing RS imagery for anthropological studies can be difficult because of a lack of financial resources, technical training, or compute needed to analyze large RS datasets. More specific to searching for mounded sites and scattered materials that would indicate past human habitation in RS imagery, it is difficult to pair legacy field data with RS imagery. When archaeologists look for potsherds, either in the field or at development sites, the standard practice is to form walking surveys to detect evidence of prior human settlement. This usually involves a large group of people walking in parallel lines over a given area, documenting what they find along the way. This process involves a lot of upfront personnel costs. The authors in [233] demonstrated the potential role of GEE in the future of archaeological research through two case studies. The authors in [234] used drone imagery and GEE to detect potsherds in the field in the hopes of speeding up this process. In [235], the authors utilized optical and SAR data on GEE to create a classifier capable of outputting a likelihood that there is a mounded site in a given region of the Cholistan Desert in Pakistan. More detailed textual summaries for each of those three studies are provided in Appendix C.15, as they are all proposed some novel methods.

3.2.16. Coastline Monitoring

Coastline monitoring is one of the less researched applications using GEE and AI (three studies total).

Table 16. Studies targeting coastline monitoring studies.

References	Method	Model Comparison	RS Data Type	Study Area
Hagenaars et al. (2018) [236]	regression	linear regression, marching squares interpolation algorithm, region growing clustering algorithm	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, Sentinel 2	Netherlands
Vos et al. (2019) [237]	regression	MLP	Landsat 4 TM, Landsat 5 TM, Landsat 7, Landsat 8, Sentinel-2, UAS	Australia, France, New Zealand, United States
Cao et al. (2020) [238]	classification	hierarchical clustering	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	China

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the three papers is provided in Figure 21. The word clouds provide an informative (general and specific) focus of each set of the papers. For example, we can see that the most frequently used general words are “shoreline”, “coastline”, and “tidal”, and “beach”. This type of research is interested in first detecting coastlines, but also monitoring geospatial changes over time (i.e., keywords “detection”, “position”, “changes”, “temporal”, “time”, and “multi-annual”). From our interactive web app (see Appendix A) and Table 16, the most-used RS datasets are Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI.

Observing and quantifying the changing position of the shorelines is critical to present-day coastal management and future coastal planning. The authors in [236] presented an automated method to extract shorelines from Landsat and Sentinel satellite imagery. The authors in [237] evaluated the capability of satellite RS to resolve at differing temporal scales the variability and trends in sandy shoreline positions. In [238], the authors proposed a method to map continuous changes in coastlines and tidal flats in the Zhoushan Archipelago during 1985–2017, using Landsat images on the GEE platform. More detailed textual summaries for each of those three studies are provided in Appendix C.16.

3.2.17. Bathymetric Mapping

There are only two bathymetric mapping studies leveraging GEE and AI.

Table 17. Studies targeting bathymetry from RS imagery using AI.

References	Method	Model Comparison	RS Data Type	Study Area
Traganos et al. (2018) [239]	regression	multiple linear regression	Garmin Fishfinder 160C sonar, Lowrance HDS-5 sonar, Sentinel-2	Greece
Sagawa et al. (2019) [240]	regression	RF	CZMIL airborne LiDAR, HDS-5 sonar, HDS-7 sonar, Landsat 8, Riegl VO-880G airborne LiDAR	Japan, Puerto Rico, USA, Vanuatu

below summarizes those studies and a word cloud generated from the titles, keywords, and abstracts of the two papers is provided in Figure 22. The most frequently used words are “bathymetry”, “satellite” and “satellite-derived”, as well as “validation”. Currently, bathymetric mapping applications are derived from radar, sonar, and light detection and ranging (LiDAR) measurements from boats and small aircraft in conjunction with model simulations. The authors using GEE for bathymetric mapping research are trying to use satellite imagery and ML on the cloud platform to generate bathymetric maps over much larger scales than would be possible otherwise.

Mapping bathymetry across large areas is a difficult problem. This is in part because high-resolution aerial radar data, which produces some of the best bathymetry maps, are expensive to collect and only cover small areas. Researchers in [239] paired field observations of coastal depths with RS imagery to train multiple linear regression models that can then predict in areas where no depth information is available. Without accurate bathymetry information, ships risk getting stranded in shallow water areas around the globe. Typically, ships equipped with sonar and planes that have airborne LiDAR are used to get water depth measurements. However, sonar is not suitable for shallow water measurements and airborne LiDAR is expensive to get. Moreover, there are very few bathymetry datasets that have a global reach. The authors in [240] used airborne LiDAR, sonar, and Landsat data to estimate bathymetry in Japan, Puerto Rico, the USA, and Vanuatu using an RF model. More detailed textual summaries for each of those two studies are provided in Appendix C.17.

3.2.18. Ice and Snow

There are only two studies in ice and snow that have leveraged GEE and AI.

Table 18. Studies targeting ice and snow studies.

References	Method	Model Comparison	RS Data Type	Study Area
Tedesche et al. (2019) [241]	classification	CART	ArcticDEM, Landsat 4 TM, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI	United States
Qi et al. (2020) [242]	regression	linear regression	Landsat MSS, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, MOD09GQ, NOAA AVHRR	China

below summarizes the two studies and Figure 23 below is the word cloud generated from the title, keywords, and abstract of the two studies. The authors are interested in measuring “changes” and “trends” in “ablation”, “break-up”, “freeze/-up”, “freezing”, “phenology”, “subsistence”, and “reflectance” levels in “ice”, and “snowfields”. From our interactive web app (see Appendix A) and Table 18, we found that the most-used RS datasets are Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI.

Global warming is putting pressure on Arctic ice and snow cover as the Arctic is heating up much more rapidly than the rest of the planet. In Alaska, changes in perennial snow cover have wide-ranging implications from changing hydrology and vegetation patterns, altering the local topology through more frequent freeze-thaw cycles, and by disrupting the ability of subsistence hunters in the region to find food. The authors in [241] used a CART model to track the changes in the cryosphere in Alaska. The duration and seasonality of lake ice is sensitive to local environmental changes such as wind, air temperature, and snow accumulation. Lake ice phenology (LIP, ice breakup and freeze-up dates and ice duration) is a particularly robust proxy for climate variability. The authors in [242] studied LIP in Qinghai Lake, China. A more detailed textual summary of those two studies is provided in Appendix C.18.

3.3. Advances in Methods

In this section, we provide a summary of all 21 novel method papers (i.e., those marked as * in the tables in Section 3.2 above). Specifically, see

Table 19. Method papers for classification tasks.

References	Evaluation Metrics	Application Area	Model Comparison	RS Data Type
Pekel et al. (2016) [32]	CE, OE	water	expert system	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI
Liss et al. (2017) [233]	accuracy	archaeology	Canny edge detection, RF	WorldView 2
Lee et al. (2018) [84]	OA, PA, UA	LULC	BULC-U	GlobCover, Landsat 5
Mateo-García et al. (2018) [224]	CE, false positive rate, OA, OE, RMSE, ROC, true positive rate	cloud	ACCA, Fmask, k-means	Landsat 8 Biome Cloud Masks, Landsat 8 TOA
Murray et al. (2018) [86]	CA, OA, PA	LULC	RF	Landsat 7 ETM+, Landsat 8 OLI, Landsat 8 SR, SRTM DEM
Orengo and Garcia-Molsosa (2019) [234]	visual analysis	archaeology	CART, RF, SVM	DJI Phantom 4 Pro
Long et al. (2019) [191]	OA, CE, OE, R2	fire	RF, SVM	CBERS-4 MUX, FireCCI51, Gaofen-1 WFV, GFED4, Google Earth, MCD12C1, MOD44B, MTBS, Landsat-8
Alencar et al. (2020) [135]	CE, OA, OE, visual assessment	vegetation	DT, RF	Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI
Orengo et al. (2020) [235]	visual analysis	archaeology	RF	Google Earth, Sentinel-1, Sentinel-2 MSI, WorldView 2, WorldView 3
Liang et al. (2020) [94]	Kappa, OA	LULC	CART, MD, RF	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, SRTM DEM
Wang et al. (2020) [151]	accuracy, CE, F1-score, IoU, Kappa, OE	water	MNDWI, MSCNN, RF	Google Earth, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI
Yin et al. (2020) [225]	CE, MIoU, OA, OE	cloud	DeepGEE-CD, FMask, RS_Net	Landsat 8 OLI
Amani et al. (2020) [69]	CE, Kappa, OA, OE, PA, UA, visual assessment	crop	ANN, SNIC	Canada’s Annual Crop Inventory, MCD12Q1, Sentinel-1, Sentinel-2
Long et al. (2021) [138]	Kappa, OA, PA, UA	vegetation	CART, LandTrendr, MD, NB, RF, SVM	CGLS-LC100, Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, Sentinel-1, Sentinel-2, SRTM DEM
Adrian et al. (2021) [72]	IoU, Kappa, OA, PSNR, SSIM	crop	DnCNN, RF, SegNet, U-Net, 3D U-Net	Sentinel-1, Sentinel-2, WorldView 3
Lin et al. (2021) [185]	Kappa, OA, PA, UA	infrastructure	CART, gmoMaxEnt, NB, RF, SVM	Landsat 8 OLI

below for novel methods for classification tasks, and

Table 20. Method papers for segmentation tasks.

References	Evaluation Metrics	Application Area	Model Comparison	RS Data Type
Isikdogan et al. (2019) [147]	F1-score, precision, recall	water	DeepWaterMapv2, DeepWaterMap, MNDWI, MLP	Landsat 8
Mayer et al. (2021) [156]	F1-score, Kappa, recall, precision	water	U-Net	JRC Global Surface Water datasets, PlanetScope, Sentinel-1

for segmentation tasks, and

Table 21. Method papers for regression tasks.

References	Evaluation Metrics	Application Area	Model Comparison	RS Data Type
Gómez-Chova et al. (2017) [223]	RMSE	cloud	kernel ridge regression, linear regression	Landsat 8, RapidEye, SPOT 4
Pipia et al. (2021) [141]	MAE, RMSE, RRMSE, R2	vegetation	GPR	HyMap, Sentinel-2
Zhang et al. (2022) [227]	MAE, RMSE, PSNR, SSIM	cloud	DeepGEE-S2CR, DSen2-CR	SEN12 MS-CR

for regression tasks. A word cloud for all 21 novel methods papers (i.e., those papers in Table 19, Table 20 and Table 21) is provided in Figure 24. The most frequently used words are “Google Earth Engine”, “classification”, and “machine learning”. For being smaller research domains (in terms of total paper count for this GEE + AI review paper), archaeology and cloud detection and mapping research presented many novel ways to use CV, ML, and DL methods on the GEE platform. In the word cloud above, “cloud”, “masking”, “archaeology”, “archaeological”, and “survey” reflect this influence. “Urban”, “water”, and “fire” are also included in this word cloud as they included papers with novel methods in them. The “cover” and “surface” keywords could be referring to the LULC, vegetation, water, or infrastructure domains. Those novel methods are detailed in subsections in Section 3.2. Inspired recommendations from those novel methods are provided in Section 4.2 and Section 4.3. It is interesting to see that there are only three studies about archaeology (Section 3.2.15), but all three papers have proposed novel methods.

4. Challenges and Research Opportunities

This section provides a summary of the patterns observed (Section 4.1) from reviewing the research discussed above. Section 4.2 and Section 4.3 describe the challenges and research opportunities from application (Section 4.2) and technical (Section 4.3) perspectives.

4.1. Summary and Discussion

4.1.1. Brief Summary of Reviewed Studies

Our comprehensive and interactive review indicates that the integration of GEE and ML (such as RF) is relatively straightforward and can directly run model training on GEE, and that the integration of GEE and DL is not as intuitive and convenient (e.g., DL is not supported directly in GEE, detailed in Section 4.1.2 below; researchers need to train DL models outside GEE, either offline on their local computers or in Google Cloud AI). However, the literature confirms that the integration of GEE and AI is becoming more widespread for geospatial analysis across a range of domains (Section 3.2). The expanding range of applications and increasing integration of AI methods into GEE observed in the reviewed literature affirm their potential to enable effective and accurate RS systems at a variety of scales.

Among the 200 reviewed studies, the most frequently used RS data are (see Figure 4c for details): Landsat 8 OLI (74 studies), Landsat 5 TM (49 studies), and Landsat 7 ETM+ (48 studies). RF (125), SVM (40), and CART (38) are the most popular ML models (see Figure 5a for details). It is not surprising that RF is the dominantly used model, as RF is a widely accepted and efficient ensemble learning model that has demonstrated the ability to cope well with a number of common ML problems (e.g., imbalanced data, missing values, the presence of outliers, and overfitting) [243]. Among the reviewed studies, the majority used ML (181), only a very small portion used DL (22) and CV (16); this is not surprising, due to GEE’s limitations (Section 4.1.2). Note that the number does not add up to 200, because some studies used combinations of ML, DL, and CV, so they were counted multiple times. Among the 22 DL studies, most of them had to run the DL models either offline on their local computers or on the Google Cloud AI platform. Only a very small portion of studies (Section 4.3.1) actually integrated GEE with DL in an indirect way—DL models were trained offline or on Google Cloud AI and then weights were uploaded to GEE and performed online prediction there. The most-employed evaluation metrics are OA (137 studies), PA (101), UA (98), and Kappa (76) (see Figure 5b for details). Of the 200 papers that we reviewed, all utilized GEE for both data processing, where 104 papers also ran computation offline.

While the research investigated in Section 3 has demonstrated the power of using GEE and AI for many different problem domains, most of the studies use GEE’s built-in ML methods (e.g., RF, SVM, and CART). There is still a long way to go before researchers can more easily develop, implement, test, and use novel AI methods (especially DL) on the platform (see Section 4.1.2) due to bottlenecks in integrating GEE with Google AI cloud. Some thematic areas are saturated with application-oriented papers, as is evident by the list and number of citations in each subsection in Section 3.2. Our recommendation is that for these areas (e.g., crop mapping and LULC), journals take less application-based papers unless they are contributing new datasets or processing pipelines for working with multiple datasets and start calling for novel method-based papers. However, other areas (e.g., archaeology and bathymetry) could benefit from more use-cases or proof-of-concept papers that open-source their code and data, speeding up the pace of research in those respective fields.

From our interactive web app tool (see Figure 20 below), we noticed that most work does not include hardware and software specifications (e.g., what CPU/GPU the authors used to run their models, what Python libraries they used to implement the DL models, etc.) and/or processing times [244]. Of the 200 total papers we reviewed, 101 ran strictly in cloud computing environments (i.e., they had no offline component). Of these 101 papers, only 10 papers provided their offline computation specifications (see Figure 25b for details). From Figure 25a, most GEE integrated with AI work ran on the GEE cloud platform. Of these papers, 98 (i.e., those marked as NA, which refers to “not applicable”) ran solely on cloud platform(s) and 92 (those marked as NS, which means “not specified”) ran locally without giving the hardware specification of the machines or runtimes for their analyses. Of those studies that used cloud computation, the majority of them are on GEE while a few combined GEE and the Google AI platform. A visual summary of software used in the reported literate is provided in Figure 26. If a publication only used GEE or its APIs, this is given a value of “NA” for “not applicable” since no additional software was used. We can see from Figure 26 that 96 papers fall into this category. Of the remaining papers that specified software that was used to complete part of an analysis outside of GEE, 27 studies used R, 23 used Python, 19 used ArcGIS, and 10 used the scikit-learn Python package. To make models comparable, reproducible, and to inform the design of RS systems, it is important to report this type of information [245]. This is even true for index-based methods and more traditional ML models so that researchers can fully evaluate the trade-offs between runtime, accuracy, and ease of implementation. The interactive web app tool that accompanies this review is intended, in part, to make future research more reproducible. Most papers have an open-access PDF/HTML version of their manuscripts, though a sizable portion of manuscripts (42 out of 200 of reviewed articles) do not. To increase the rate of progress integrating GEE and AI, we suggest authors seek to provide an open-access version of manuscripts whenever possible.

4.1.2. GEE Limitations

GEE serves as a great free-of-charge cloud platform for EO big data processing and analysis. With the very large amounts of data and combinations of temporal domains utilized in [21], GEE was critical to enabling these investigations. The use of GEE also facilitated the testing of several ML algorithms in a much faster way than would have been possible without it. The oil palm classification demonstrated [107] in GEE is useful to provide a quick understanding of oil palm plantations present in the landscape. This in itself is advantageous for independent monitoring bodies to conduct a survey of the landscape in question and conduct more detailed assessments if necessary. In the near future, it is foreseeable that a growing number of large-scale mapping and monitoring programs, enabled through the integration of AI with GEE, will emerge as critical to tools to help scientists, managers, and policymakers understand and respond to our environment [136].

However, GEE also has multiple noticeable limitations. Many authors reported compute limits, lack of processing methods, the inflexibility of different models, and a lack of data as their main limitations (each detailed below). Some recommendations for future research derived from these limitations are provided in Section 4.2, Section 4.3 and Section 4.4.

●

Compute limits [17,55,78,83,85,87,93,141,155,171,200,217,234,239,240]: Authors often ran into memory errors when analyzing too many field samples/observations. This also happened when the size of authors’ input data was too large more generally and it was difficult to know beforehand if intermediate processing steps would trigger this error. Thus, many authors had to export data as part of their analysis to access functionality not on GEE or because using GEE would make them run out of the amount of free compute provided. For example, every image uploaded to GEE (at the time of this paper’s release) is limited to 10 GB [234]. As the authors used sub-centimeter drone imagery, they had to downsize each image before uploading it, resulting in a loss of resolution. See below for a few quoted limitations:

○

“…The users are limited to approximately 1 million training points…, a limitation in using a high number of trees within GEE when the amount of field samples is high” [17].

○

“One of the disadvantages of using the GEE cloud computing platform is that it limits the number of field samples and input features. This is especially challenging when the analysis is applied to a large domain, which may reduce the efficiency of the implemented method” [171].

○

“…The current GEE pipeline for processing the available data on GEE through the Python or JavaScript APIs requires exporting large volumes of data to cloud or local storage … These processes are time consuming and require extra funds for cloud processing and cloud storage” [87].

●

A lack of processing methods/models/algorithms [17,21,35,46,81,85,93,101,107,110,120,141,152,158,159,160,201,215,217,221,225], reasons listed were:

○

There is a lack of domain-specific models and methods (GEE algorithms are general) because GEE is more developed in some areas (LULC, forest, vegetation, crop) than others [17,46,107,152,158,225,239];

○

No neural networks (NNs) are currently supported on GEE directly, but many authors use DL models for their research [46,55,71,107,136,161,166], and they either have to train their DL models offline or on Google Cloud AI, which is not free of charge. Authors can also use TensorFlow on Google Colab and Google Cloud AI but not directly on GEE. For example, in [225], “… limited by the computation resource of GEE, some specific convolution layers of DNN cannot be implemented in GEE. For example, a dilated convolution layer could not be achieved due to the fact that dilation is not supported in the convolution API provided by GEE. Conversion of other types of convolutions to the convolution used in this study may help to solve this problem and it needs further investigation…”. The authors in [156] mention, “… integration of the Google AI platform with GEE creates a versatile technology to deploy deep learning technologies at scale. Data migration and computational demands are among the main present constraints in deploying these technologies in an operational setting;”

○

SNIC is the only object-based classifier on GEE; authors also want more “advanced methods” or just more options;

○

Hyperparameter tuning is not possible on the platform [21], so many authors use local software (e.g., scikit-learn) for this purpose and then upload the models to GEE afterwards;

○

One of the benefits of using an RF model is that you can run a feature importance analysis afterwards to determine which set of input features contributed most to the model’s learning. However, this extremely common and important operation is not possible on GEE.

●

Inflexibility of models [19,35,46,152,159]: This limitation is similar to lack of models but is different in that it describes issues using models already on GEE. For example, authors in [35] emphasized, “A third limitation to the modeling approach described here is its current incomplete use of cloud-computing services, and reliance on desktop computer power to run the BRT models. Ideally, the modeling would be run within the same environment where the satellite data are preprocessed—Google Earth Engine—or a similar cloud-computing service offering similar levels of access to Sentinel datasets. GEE does currently provide machine-learning algorithms such as random forests, but these do not provide the flexibility that is currently offered within the BRT R functions”. This is both lack of methods and model inflexibility. The authors in [46] found that in general the algorithms on GEE were not very flexible and some preprocessing steps such as dealing with missing data were difficult to implement. Thus, the authors performed all preprocessing steps outside of the GEE platform.

●

Lack of data [32,46,54,67,75,94,120,126,127,160,161,162,183,184,193,215,221]: This related to both a lack of field observations and curated RS datasets.

○

Not every data product is on GEE;

○

Authors specifically called for a Landsat-Sentinel combined dataset. This dataset could serve as the foundation for research in many different application areas by expanding both the spatial and temporal resolution available to researchers;

○

Very-high-resolution imagery is not on GEE, meaning that to validate GEE prediction results authors often need to download this data locally.

●

Importing and exporting data from GEE [83,126,193,198,234]: This process is time-consuming and results in lower resolution classification maps. However, many authors need to import or export data based on storage constraints on GEE.

●

Other limitations:

○

There is a delay from the time RS data are available and the time that they are uploaded to the platform, limiting their utility for time-sensitive applications [213,214];

○

Authors might have a hard time converting programs to GEE from their own environment [81,136,217]. Cited issues were that authors were not familiar with JavaScript, Python, or the GEE programming interface. Authors were concerned that not everyone would have the skillset to implement models in GEE;

○

A concern that data and code will not be kept private for sensitive use-cases [217].

4.2. Challenges and Opportunities from an Application Perspective

Most of the current integrations of GEE and AI utilize data and models already available on GEE (detailed in Section 3.2). Only a few papers proposed novel methods (summarized in Section 3.3). Below, we provide some challenges and opportunities related to application-oriented research.

4.2.1. Proof-of-Concept for Less Researched Applications and Novel Methods for Saturated Application Domains

The authors in [107] point out, “… classification method demonstrated in GEE is useful to provide a quick understanding of oil palm plantations… This in itself is advantageous for independent monitoring bodies to conduct a survey of the landscape in question and conduct more detailed assessments if necessary.” For applications that are not yet well-studied using GEE, it will be useful to run some proof-of-concept experiments on GEE. These types of analyses will shed light on what limitations exist for doing domain-specific research on the platform (e.g., are the main barriers a lack of data, lack of preprocessing models or AI methods, etc.).

Even for very saturated application domains (e.g., wetland mapping, see Section 3.2.6), there are few novel methods. We would like to clarify that it is not that there are no novel methods. Researchers still use interesting preprocessing pipelines, creating new datasets, and often use DL. Again, we take a very narrow view of “novel” in this paper and this definition is confined to how researchers are using AI methods on the GEE platform. Researchers focused on wetland studies seem to be much more focused on using free compute, compiling and scaling up datasets over larger areas than would be possible on local machines, and creating open-source processing and visualization pipelines. However, there is still a lot of room for novel methods for those saturated application domains. For example, it would be useful for a saturated application domain to experiment with novel methods developed for other domains. The web app we developed for this review paper will serve as an important tool to easily find novel methods (check the demo video of the web app for how to find a novel method paper; the link to the video is provided in the Appendix A).

4.2.2. Using ML for Exploration/as an Aid to Human Expertise

At a certain point, it is difficult or impossible for humans to determine meaningful relationships in complex, highly dimensional data. One of the ways AI is most helpful is in data exploration. Still, the goal of EO-AI research should not be to automate away human expertise since AI models cannot understand human values and are often heavily biased. In [233], the authors use an RF set to output probabilities instead of class predictions to identify possible archaeological features in Jordan. The authors in [235] do the same thing in Pakistan over a large desert, saving time and effort that otherwise would have required surveyors to spend time in potentially unsafe conditions. While [233,235] use ML models to prepare for fieldwork identifying archaeological mounds over large areas, [234] used it to identify potsherds in the field. They use drone imagery and GEE to identify potsherds, thus allowing surveyors to focus their attention on finding and cataloging them even over large areas. This exploration method has also been successfully demonstrated for burned area mapping in [191]. The authors use a similar process of first using an RF model in probability mode to find a good “starting point” for classification, and then they tweak the probability to remove false positives before using a pixel-aggregation algorithm to determine the final classification output. The authors are able to show that their method shows good agreement with other commonly used burned area products, but with finer classification boundaries. In order to explore the potential to distinguish between subtypes of surface water body, the authors in [158] use slope, shape, and phenology, and flooding information as input to an RF model to predict for lakes, reservoirs, rivers, wetlands, rice fields, and agricultural ponds. They found that their method does not work very well for wetlands and the OA is not very high (85%) across classes. However, the RF model they use is interpretable and they show which other subclasses are easy or more difficult to predict for. Unfortunately, the entire preprocessing (method) cannot be run directly on GEE, because the shape features cannot be calculated on the GEE platform and are crucial to the overall analysis; the authors first have to do this in a local environment and then upload them.

It is important to note that the authors in these papers are actively changing the results of classification. In some cases, they are doing so many times (over several iterations). Thus, they are introducing bias into their models, but the trade-off is acceptable if the emphasis is on exploration rather than on statistical validity. This methodology is similar to using an expert system where domain experts use ML systems in a “collaborative” way, blending human expertise with the automation capabilities of AI. Still, these models would need to be continuously tested on new data to make sure that their probability threshold values are accurate, and their predictions should not be taken at face value.

4.2.3. More (High Quality) Data

Many authors suggested and called for more data for performance improvement (e.g., [16,127,162]). However, as the studies in [105] showed, more data are not always better. The authors in [105] investigated the difference in ML model performance when using single image mosaics, time series RS imagery, statistical features (median, standard deviation), band ratios, or all of the features listed. They test this by training an RF model on each subset of data to create LULC maps in Brazil. The authors find that inputting a time series of the data is the most accurate, more accurate even than when using all of the computed indices and statistical features. The authors in [239] train four different multiple linear regression models on sonar from field data collection and optical RS imagery to map bathymetric depths in three different locations near Greece. They acquired good results with a very simple, intuitive model. While current trends point to the use of models with increasing complexity, it is important to note that many times a simpler model will perform well given high-quality input data.

The analysis in [246] suggests an increased focus on dataset scaling is needed; the authors further emphasized that scaling to larger and larger datasets is only beneficial when the data are high quality. Meanwhile, as [247] emphasizes, more data and a simpler NN is better than a bigger NN with more data. This echoes the data-centric views of AI [248] proposed by one of the AI pioneers Dr. Andrew Ng. Ng says it is time for “data-centric” solutions to big issues. Ng observes that 80% of the AI developer’s time is spent on data preparation [248]. Thus, domain experts should be involved in creating high-quality datasets, since they know the data sources and relevant input variables much better than AI engineers. Thus, together with the authors in [246,247], we call for responsibly collecting larger datasets with a high focus on dataset quality. We also call for researchers to share their datasets on GEE, which would be useful to a wide variety of domains and researchers. Potential recommendations in this direction: (1) improving the quality of existing datasets, and (2) working on generating more data but with a focus on good quality.

4.2.4. Feature Engineering and Feature Importance

Feature engineering (see Appendix A.1 in [1]) using RS data is difficult because it is time-consuming to do and often relies heavily on human experience, domain knowledge, and technical expertise (in terms of location, what kind of data are being processed, what variable to look for, etc.). Meanwhile, feature engineering is often necessary for a given ML analysis because of the large amount of RS imagery coming in every day. In this way, DL methods can help because they are able to recognize complex patterns in data without the necessity for feature engineering (feature engineering can still help in DL analysis, but is not necessary, i.e., NNs can learn complex patterns from raw data). Still, tradeoffs between traditional ML models and DL methods have not been properly mapped out for the RS space. The authors [43] set a good example by addressing this issue through comparing the performance of an RF model (with feature engineering) to a long short-term memory neural network (LSTM) and U-Net NN models (without feature engineering) to identify pasturelands. The RF model was trained on GEE, while the DL models had to be trained offline as GEE does not currently support DL models (Section 4.1.2). U-Net had the highest generalization across both the validation and testing sets, maintaining high accuracy rates, while the LSTM and RF model underfit the test set. To illustrate the tradeoffs between ML and DL models, the authors include run and inference times. The RF model was able to complete training and prediction in 3 h. As reported in [43], U-Net takes a long time to train, while the LSTM takes a long time at inference time. Specifically, the LSTM took 30 min to train but 23 h to predict on the test set, while the U-Net took 24 h to train but 1.2 h at inference time. Much more work like this should be done to explore the strengths and weaknesses for ML and DL models, as this will be helpful for many research areas that would like to take advantage of GEE and AI.

With proper features from feature engineering, ML algorithms, which require less (good quality) trained data than DL, often perform better than DL. For example, the authors in [136] reported that their results indicated that the classification accuracy of DL was not as good as traditional ML methods (e.g., SVM). We recommend the following three directions for future studies in terms of feature engineering.

(1) Compare multiple ML algorithms or ML vs. DL algorithms: As pointed out in [136], it is worth investigating which methods (ML vs. DL) are better for a specific domain application. Their results indicated that the classification accuracy of DL was not as good as traditional ML methods (e.g., SVM). Several ML models are compared in [115] to map oil palm using Landsat 8 imagery in Malaysia. The authors find that tree-based ML models (e.g., RF, CART) work better than an SVM for the task and are able to classify large areas with high accuracy. Even so, classification errors are traced to the relatively coarse resolution of Landsat data. The authors suggested that higher resolution imagery (e.g., Sentinel) and the ability in the future to use DL methods on GEE will most probably improve higher performance. The authors in [136] developed and implemented a new pixel-based method (Ppf-CM) in GEE using 525 full Landsat scenes (19.96 billion pixels) to monitor S. alterniflora dynamics. They found that Ppf-CM not only enhances the spectral separability between S. alterniflora and others, but also improves the problems caused by the scarcity of entire cloud-free Landsat scenes. These findings echo well with prior GEE-supported pixel-based studies (e.g., [80]) and further confirm that pixel-based methods outperform scene-based methods to monitor S. alterniflora. The classification results in [161] were evaluated using both pixel-based and object-based RF classifications available on the GEE platform. The results revealed the superiority of the object-based approach relative to the pixel-based classification for wetland mapping.

The authors in [46] compare several algorithms on the GEE platform, including CART, IKPamir, LR, a multi-layer perceptron (MLP), NB, RF, and an SVM, for crop-type classification. The authors also use an ensemble NN but have to move off the GEE platform since NNs are not currently supported. The ensemble NN performed the best out of all the models. The authors found that atmospherically corrected Landsat data boosted model performance more than when models were fed Landsat composites data. The authors in [56] compare the performance of an artificial neural network (ANN) to CART, RF, and SVM models on GEE for sugarcane mapping in China using Sentinel-2 imagery. The authors identify that the SVM performs the best, but then go on to show which type of errors each model makes. For example: the ANN tended to overfit the data and give too much preference to the sugarcane class, while tree-based models confuse the forest and water classes. The authors then incorporate Normalized Difference Vegetation Index (NDVI) information into the SVM to show how the model does with this extra information. It is not clear why the authors did not allow each model to see NDVI information, as this extra information may have helped various models learn better. If the authors wanted to show how models learned from phenology information versus phenology combined with NDVI information, they could have trained each model on separate subsets of the data. While GEE allowed [159] to train several ML models, some models failed to run due to computational constraints or inflexibility. The authors show that in all cases, ML models do much better at binary than multi-class classification. The authors in [66] utilize many ML algorithms available on GEE and compare specific time windows for phenological analysis and find that the closer the data comes to planting and harvesting time, the better the ML models performed.

(2) SAR + optical RS images for better model performance: In addition, many studies reported [17,57,68,72,74,165,166,182,212,214] or suggested in future work [46,56,107,215] that SAR combined with optical RS images would improve model performance. Three classification methods (SVM, RF, and decision fusion) were used in [52] for the pixel-wise classification for crop mapping. The SVM classifier resulted in the lowest accuracy. The integration of multispectral and SAR data improved the classification accuracy. To improve the results in this study, the authors in [56] identify that using SAR data would be helpful in removing the impact shadows have on classification errors for sugarcane mapping. The authors in [95] compare the contribution of SAR data and different indices (e.g., NDVI, EVI, Soil Adjusted Vegetation Index (SAVI), Normalized Difference Water Index (NDWI)) derived from optical data on overall classifier performance. They find that including SAR data moderately improves performance, while only NDWI gives the ML model a significant performance enhancement. Using optical, thermal, and SAR imagery in addition to DEM data, [221] produces a global, high-resolution soil moisture map. The authors use a gradient boosted regression tree (GBRT) model to train on in-situ observations paired with RS imagery to then predict soil moisture in other locations. After running a relative variable importance analysis, the authors can conclude that optical RS imagery and land-cover information play the most important roles in determining soil moisture content, but that SAR imagery and soil data also contribute significantly to the model’s overall performance. This finding highlights other studies’ results ([95,161,182]) that the combination of optical and SAR data improves predictive outcomes.

(3) What input for what algorithms (feature importance): This section is separate from feature engineering in that it is less concerned with computing new features from existing data than with determining which input variables contribute to model learning.

In [58], the random samples extracted from the training pool along with RS-derived features and climate variables were then used to train ecoregion-stratified RF classifiers for pixel-level classification. Evaluation of feature importance indicated that Landsat-derived features played the primary role in classification in relatively arid regions while climate variables were important in the more humid eastern states.

To investigate how best to identify impervious materials in RS imagery regardless of cloud cover, [182] combine nighttime light, DEM, and SAR data and an RF model on GEE. Their resulting maps are more accurate than commonly used maps like GlobeLand30. More importantly, though, the authors quantitatively show that using multiple sources of data are better than single sources for this task; optical data are the most important, but SAR data improve accuracy rates across all metrics. In future studies, more work like this needs to be done so that researchers can save time and effort by knowing which data will be useful for a task beforehand. The authors in [178] compare different combinations of input data and their impact on model performance. For their application, Landsat 8 data serve as better input than Landsat 7 alone or Landsat 7 data with computed indices like NDVI. Having access to datasets like the one produced by [178] will make it much easier for future researchers to create more accurate building detection models, either by allowing researchers to add to this dataset and training ML models or by using it as one of several other datasets incorporated into the same analysis.

It is important not only to be able to map the current state of wetlands vegetation, but how that vegetation is changing over time. However, different sets of input data and ML methods used for change detection of wetland vegetation need to be evaluated more fully as choices made during preprocessing and hyperparameter tuning can affect the end result of an analysis. The authors in [138] use an adaptive stacking algorithm to train an ML classifier on optical, SAR, and DEM data to identify wetland vegetation. Adaptive stacking is using one ML classifier to identify the optimal combination of ensemble classifiers and hyperparameters to be used for a given task. In this case, the authors use an RF model to determine the best combination of the CART, Minimum Distance (MD), NaiveBayes (NB), RF, and SVM classifiers on GEE. The authors find that the adaptive stacking method is much more accurate than the RF and SVM models alone. The resulting classification map is then combined with a trend analysis performed by the LandTrendr algorithm, which allows them to identify wetland vegetation distribution as it is now and also how it has changed over time. The authors in [138] also test their workflow on different subsets of input data and show that adding more data helped the adaptive stacking algorithm learn better (the best combination of input data was all of the data). The authors note that forest and reed classes were not identified well with their adaptive stacking algorithm, and that the LandTrendr algorithm will most likely need to be re-tuned in different environments.

The authors in [15] integrated single-date features with temporal characteristics from six time-series trajectories (i.e., two Landsat shortwave infrared bands and four vegetation indices), to produce an intact-disturbed forest map to track degraded forests. The whole processing pipeline is done on GEE using an RF. The authors also ran a relative variable importance analysis for each ecoregion. The authors are able to show that past maps are a bit outdated due to their inability to separate forest classes by intact and degraded, although their results vary from ecoregion to ecoregion. The purpose of the study in [21] was to determine how the inclusion or exclusion of data for training RF models with RS and temporally variable climate variables influences model outcomes. Cloud computing on GEE was utilized in [35] to create an open-source, reproducible map of wetland occurrence probability using LiDAR and RS data for the entire Alberta. Using a BRT, the authors are able to match a current governmental effort in Alberta while also producing a relative variable importance showing which RS variable might be the most useful for future wetland mapping efforts in the area.

The authors in [55] used a CNN–LSTM hybrid model to predict soybean yield in the contiguous United States using RS imagery alongside weather data and show that the hybrid approach works better than either CNN or LSTM alone, although the results were better in some states than others. Additionally, the authors create combinations of input data to determine which variables are most important in training their NN.

A low-cost method was demonstrated in [107] for monitoring industrial oil palm plantations in Indonesia using Landsat 8 imagery that allowed them to distinguish between oil palm, forest, clouds, and water classes using the CART, RF, and MD algorithms. Their results demonstrated that CART and RF had higher OA and Kappa coefficients than the MD algorithm. In addition, the authors [107] compared model accuracy based on different combinations of spectral bands (particularly red-green-blue (RGB) and infrared bands include shortwave infrared (SWIR), thermal infrared (TIR), and near infrared (NIR)), including all bands, to determine which would help specifically with oil palm plantation monitoring.

The authors in [136] used a specific invasive species in China as a case study for developing an ML pipeline that takes into account both cloud cover and phenological information. They compared the ability of a stacked autoencoder and an SVM to classify vegetation types. While the SVM was trained on GEE, the DL model had to be trained offline as the platform does not currently support DL models. The authors find that the DL model performs better than the SVM and that both models perform better with phenological information. The same species of plant can look different at different stages of its life while also being submerged under water in some RS scenes. The authors in [140] argue that phenology information in RS time series can better capture tidal flat wetland vegetation and so can compare phenology information to statistical (min, max, median) and temporal features (quartile ranges). They then feed this data into an RF while analyzing its effect on model performance during different periods of time (all data, green and senescence seasons) for wetland vegetation classification. The authors showed that the phenological information was the most important input feature to the RF, while combining all three sets of features led to the highest accuracy. In addition, the model performed best when predicting over both the green and senescence periods, most likely providing the model with a better estimate of the total variance needed to identify wetland vegetation. More research like this should be done to isolate the importance of individual input features and time periods in ML model performance. To explore the potential to distinguish between surface water body subtypes, [158] use slope, shape, and phenology, and flooding information as input to an RF model to predict for lakes, reservoirs, rivers, wetlands, rice fields, and agricultural ponds. Their method does not work very well for wetlands and the OA is not very high (85%) across classes. However, the RF model they use is interpretable and they show which other subclasses are easy or more difficult to predict for. The authors in [192] found that Landsat 8 data led to higher fire burn estimates while still improving fire burn detection accuracy, though both Landsat and Sentinel-2 catch more small fire patches than MODIS.

To determine the impact of using higher-resolution RS data products, the study in [192] compared how Landsat and Sentinel optical imagery affected an ML model’s performance in burn area classification. The authors used Weka clustering output and different spectral and index information as input into the CART, RF, and SVM models available on GEE. They find that both Landsat and Sentinel imagery produce much better maps that capture small burn areas that current maps and fire monitoring products like MODIS are not able to capture, though Sentinel imagery leads to an underestimation in burn area. The authors also find that the tree-based algorithms perform comparably to each other but much better than the SVM model. This study highlights the importance of analyzing different data sources and ML models to show their respective contribution to predictive performance.

4.2.5. Creative Integration of Existing Algorithms Available on GEE

Through in-depth exploration, it is promising for domain experts to propose creative integration of existing CV/ML algorithms available on GEE. The GEE has provided an extensive cloud platform to train and classify using ML algorithms. The authors in [192] have studied and evaluated the potential utilities of medium resolution satellite imageries of Landsat-8 OLI and Sentinel-2 to estimate precise forest burnt area over Uttarakhand, Himalaya. Specifically, they used the pre- and post-fire differential reflectance to capture fire patches using “differenced” burn sensitive spectral indices (dNBR, dNDVI, dNDWI and dSWIR, see the abbreviations list right before the References). The unsupervised Weka cluster layer used as input to ML algorithms along with differential indices, which played an important role to recognize the pattern and expansion of fire patches. Among three ML algorithms, CART and RF achieved better performance in terms of accuracy (Kappa) than SVM. To explore how CV algorithms and ML models can be used together on GEE, the authors in [226] combine the existing Cloud-Score algorithm with an SVM to detect clouds in imagery ranging from Amazon tropical forests, Hainan Island, and Sri Lanka. The Cloud-Score algorithm first masks input RS imagery, then is used for input to train the SVM. This process led to much higher accuracy rates than any of the other CV algorithms for cloud detection and does so with considerably lower error rates. The authors in [150] analyze to what degree different preprocessing steps affect the output water maps using both SAR and DEM data and two variations of Otsu’s thresholding algorithm. They showed that SAR data included radiometric terrain correction (RTC) as a preprocessing step yield more accurate results and that Bmax Otsu thresholding is more stable to different inputs than Edge Otsu. However, their analysis was limited in time and space, so more work needs to be done to test their results in different locations and varying terrain types at different times.

Model stacking, ensemble learning, and label estimation: Many authors use or test multiple CV, ML, and DL models in their research. Still, it is difficult to tune hyperparameters or choose threshold values that affect the end result of a given analysis. To alleviate these problems, several authors we identified in our review use different models to automate the hyperparameter tuning process. For example, in [135], the authors used two different RF models to produce maps of vegetation change, as detailed in Section 3.2.4. The authors in [138] used an RF model to train a separate ensemble classifier made up of CART, MD, NB, SVM, and another RF model. The first RF was able to choose the best combination of models and each model’s respective hyperparameters. This ensemble model performs better at wetland detection than any of the models individually. In [185], the authors use almost the exact same method (this time for building detection), though the final ensemble is chosen via a manual weighting process.

4.2.6. Beyond ML: Modeling in GEE

A majority of papers we reviewed used data and algorithms on GEE to complete their analyses. These “proof-of-concept” papers are often explorations by authors into how to use the platform or for showing that research typically done offline can be done in the cloud. However, many of the most straightforward classification and regression applications have now been sufficiently demonstrated. In the future, one area of research that should be given much more focus is on implementing more complex production applications built on top of GEE that make use of modeling. For example, after mapping out urban areas using ML classification techniques, the authors in [94] then went further and implemented an ecosystem service value model on the platform. This is a creative way to use the available parallel processing capabilities of GEE and is an under-researched application area.

4.3. Challenges and Opportunities from a Technical Perspective

Making the integration between GEE and AI more seamless would allow researchers and practitioners in various domains to better take advantage of GEE and AI. From our systematic review, we provide some identified future challenges, opportunities, and recommendations below for researchers, practitioners, and engineers (including GEE engineers at Google) to consider. Note that some of the recommendations are general directions (Section 4.3.1, Section 4.3.2, Section 4.3.3, Section 4.3.4 and Section 4.3.5) and others are more specific (Section 4.4).

4.3.1. Model Implementation and Online Learning in GEE

GEE has a large data catalog and houses many preprocessing methods, as well as various CV and ML algorithms. However, an often-cited limitation (Section 4.1.2) to research on GEE is that there are not enough methods or models implemented on the platform. Thus, a promising applied research direction for using GEE and AI is to implement and test CV or ML models that would be useful to other researchers. For instance, the authors in [141] developed a GPR model on GEE that works for both vector and tensor input. This model was also cloud-optimized specifically for the size limits of GEE, making it a lightweight but accurate option. Similarly, the authors in [84] implemented an unsupervised Bayesian model in GEE for LULC classifications.

Many GEE and AI analyses rely heavily on optical imagery. Obtaining enough cloud-free RS imagery can be difficult, though there are methods like Fmask that help remove clouds when they are present in RS scenes. Still, options for cloud-removal algorithms on GEE are limited. The authors in [223,224] both propose new methods for cloud removal over large areas that can be run directly on GEE, without the need to download data. In [224], the authors show that their proposed method outperforms popular algorithms like Fmask and Automated Cloud Cover Assessment (ACCA) by 4–5% accuracy. The authors in [223] tested their algorithm on both Landsat and Satellite pour l’ Observation de la Terre (SPOT) optical imagery. Along with the authors in [107,239], we call for more domain-specific novel AI methods to be implemented on the platform, which would be useful to a wide-variety of researchers. To move towards the seamless integration of GEE and AI, we suggest the following three directions.

(1) Simple but robust ML/CV methods: From current GEE limitations (see Section 4.1.2), one promising way to make integration of GEE and AI smoother, deeper, and more robust, is to develop simple, novel, and robust CV/ML methods, for example, Canny edge detector (developed by John F. Canny in 1986 but still being used in nowadays’ edge detection applications, cited a total of 39,549 times as of 17 April 2022 and by 1953 papers in 2021; his Google Scholar citations as of today total 66,796, though the author’s Canny edge detector paper’s citations are about 60% of the total). We explicitly list this example here to show that it is worth devoting the time to develop simple but robust CV/ML algorithms. The Canny edge detector is robust and is used in many image processing applications. However, this specific algorithm may not be appropriate for RS images or for RS images in a specific domain. We call for AI and RS researchers and engineers to develop robust CV/ML methods for novel and, ideally, computation-optimized, RS-image processing algorithms towards the smooth and robust integration of GEE and AI.

(2) Reimplementing and/or optimizing (both classic and state-of-the-art) CV/ML methods on GEE: The authors in [107] pointed out a need for more and better algorithms on the GEE platform. The authors in [141] implemented GPR, which is increasingly used because it is a transparent ML model that also outputs model uncertainties. The method in [141] has been optimized for green Leaf Area Index (LAI) in RS imagery, in a way that is optimized for GEE. First, they created the model so that it can run on vector or tensor time series imagery. Then, the authors used active learning (AL) for feature reduction so that the model only learns on important data while creating a model that can run within GEE’s memory confines. This GPR model is then used to gap-fill RS imagery focused on LAI, meaning the model is able to “see” through clouded optical imagery. More work like this should be done, either in creating new models to upload to the cloud that other researchers can use or optimizing these models so that they are memory efficient and thus can leverage GEE on the cloud, instead of needing to preprocess and model training on local computers or on Google Cloud AI. The authors mentioned that better GEE code documentation and error messages could help future researchers interested in developing custom ML models for the platform. (detailed in Section 3.2.4).

(3) DL with GEE: DL models are not currently available on the GEE platform (Section 4.1.2). However, some authors [69,151,225,227,228] have found an interesting workaround that allows them to use NN models directly in the cloud. All of these authors first train an NN model outside GEE, and then upload the weight matrices as data files that can be read by the JavaScript or Python development environments. Then, it is necessary to implement each layer in the network (convolutional layers, activation layers, etc.), so that imagery can be run through the NN at inference time to produce predictions. This method has worked across domains like water extraction, cloud detection, and crop mapping. Still, there are several caveats to this approach. First, researchers need to have access to the compute needed to train the NN model in the first place. Often researchers are drawn to GEE because of the freely available compute, so this method is mainly geared towards those looking specifically to use NNs. Researchers also need to know how to implement and test different layers in an NN, a task that many EO researchers may not have the experience for. Lastly, none of the authors listed above implemented the full training process on GEE (e.g., forward and backpropagation).

Novel model architectures: Both [72,147] used the GEE platform to download and process data with which they could then use to train novel NN models. The authors in [147] trained a CNN called DeepWaterMapv2 that can handle flexible input sizes of optical RS imagery and evaluate images with a constant runtime. Additionally, their CNN can filter out clouds to fill in obstructed scenes and predict where water is with high accuracy. The authors in [72], on the other hand, used both optical and SAR data from GEE to train a 3D U-Net model for crop-type classification. The 3D CNN architecture shows an improvement over the more traditionally used 2D convolution operations. Neither author used GEE itself for the DL part of their analysis, because NN models are not currently supported on GEE. However, their research shows that GEE makes it easy to locate data for a variety of applications.

Transfer learning (TL): TL is one powerful technique that makes models trained on large sets of data and compute available for applications without these resources. TL was initially proposed in [249] and recently received significant attention due to recent advances in DL [250,251,252,253,254,255]. Inspired by humans’ capabilities to transfer knowledge across domains (e.g., the knowledge gained while learning violin can be helpful to learn piano faster), the main idea behind TL is that it is more efficient to take a DL model trained on an (unrelated) massive image dataset (e.g., ImageNet [256]) in one domain, and transfer its knowledge to a smaller dataset in another domain instead of training a DL classifier from scratch [257]. A major assumption in many ML and DL algorithms is that the models will generalize to new, unseen data given that it is from the same feature space and distribution [258], and that there are universal, low-level features shared between datasets for different applications. However, this assumption does not hold for many real-world problems. For example, it is not uncommon that a classification task in one domain lacks sufficient data, but a very large set of training data is available in another domain, where the data may be in a different feature space or follow a different data distribution. In such situations, knowledge transfer, if done successfully, would greatly boost the learning performance by avoiding expensive and labor-intensive data-labeling efforts [250]. The authors in [71] showed that TL works the best when they use a U-Net to map sugarcane in Thailand, meaning that the pre-trained weights resulted in the highest accuracy, F1-score, precision, and recall. More work should be done towards evaluating the effectiveness of TL within the EO studies as it could potentially save large amounts of compute from not having to constantly train DL models from scratch. The authors note that their model does not take into account phenological information, which would have required changing the NN architecture, but that this is an area for future research using their method.

4.3.2. Web Interface Tools to Support ML Exploration

The authors in [107] noted that there is a need to make intuitive, easy-to-use tools for specific tasks that incorporate input from the public and other stakeholders like non-governmental organizations (NGOs) and government agencies. We recommend the following two directions for future research while researchers and practitioners develop such tools.

(1) Humans-in-the-loop: As authors in [1,116] emphasized, one big research direction we recommend is human-in-the-loop ML Human-in-the-loop computing aims to achieve what neither a human being nor a machine can achieve on their own. The authors in [259] emphasized that a human-centered understanding of ML can lead not only to more usable ML tools, but to new ways of learning computationally.

The authors in [234] used drone imagery and GEE to detect potsherds in the field in the hopes of speeding up this process. They train a CART, RF, and SVM on this drone imagery, but only the RF model produces adequate results. The authors test their workflow in two separate locations in Greece. This research is interesting because the overall goal of the paper is not to optimize accuracy per se, or to even replace human experts in the field. As the authors note, “It is important to note here that this method does not aim to substitute archaeological fieldwalking but complement much of the non-specialist work conducted by groups of people for long periods of time in conductive environments so there is more time and resources available to dedicate to specialized work”.

To explore how GEE could be used to create an open-source processing pipeline for deforestation mapping in Liberia and Gabon, [116] used two different RF models to create data masks and then predictions for various land types there. However, the output classification maps were then shown to local experts to correct, boosting the accuracy of the final accuracy rates. The authors showed that their method is more accurate than other efforts to classify deforestation rates in these two countries, though there were still some model misclassifications between classes due to not enough ground-truth data. This presents a future area of research, where ML/DL/CV models are used to generate first-order maps that are then verified by experts in that field (i.e., expert systems). Building land classification maps in this way saves experts’ time but also keeps humans in-the-loop where human values and knowledge can still be represented and included.

(2) Smart GEE + AI data annotator: ML, especially DL methods, are only as good as the amount of labeled training data that they have access to. To accelerate the integration of GEE with AI to generate informative insights, development of smart GEE data annotators for GEE and AI is one of the most important directions to go. Humans and machines they each have their own strengths, in the smart GEE-AI data annotator systems, the classifiers should be able to select what samples are most confused based on the current learning status and thus to ask the human annotator (e.g., domain expert) for annotation; this is what AL is good at. The main idea behind AL is to take advantage of a large set of not-annotated images by selecting which images would help to improve the performance of ML/DL and thus need annotation through an uncertainty selection strategy (see [1] for a detailed introduction about the selection strategy).

While DL receives lots of attention, these models still require a lot of input data and large amounts of compute to train them. However, as compute becomes publicly available in cloud-based platforms like GEE, obtaining large amounts of labeled training data remains a key bottleneck to using DL models. One novel way to make the data labeling process less time- and resource-intensive was illustrated in [156], where the authors used current water maps and a segmentation algorithm to automatically collect data labels from Sentinel-1 imagery. These data are then used to train variations of U-Net in an offline environment. Due to computational constraints, the authors were not able to compare their model to more traditional ML models like an RF. Even with their automated data labeling pipeline, the authors note that their study lacked sufficient data to adapt their method to more than one country and manual validation was still necessary to validate the model post-prediction.

4.3.3. Open-Source GEE-AI Library Development

One promising way to accelerate the integration of GEE and AI is the availability of open-source libraries in multiple languages (e.g., Python and R). All the studies we have investigated using GEE with DL have trained their DL models offline (detailed in Section 4.3.1), not directly on the GEE cloud computing environment.

A strong need of Python-based GEE application/package/framework: As [225] pointed out “some specific convolution layers of DNN cannot be implemented in GEE. For example, dilated convolution layer could not be achieved due to the fact that dilation is not supported in the convolution API provided by GEE. Conversion other types of convolutions to the convolution used in this study may help to solve this problem and it needs further investigation”, to make the integration of GEE and AI seamless, we need open-source GEE and AI libraries that help make sure existing AI (especially ML and DL) algorithms can be used in the GEE environment. Good examples of this are the Geemap [260] and Rgee [261] libraries, which make it easy to access GEE JavaScript functions for researchers who use Python and R. The authors in [35] noted that some uncertainties in their underlying training dataset, a lack of subsurface soil information, and having to move between GEE and offline analysis may have contributed to errors in this analysis. It will be easier to avoid errors like these if there are more open-source Python/R libraries that make it easier to connect GEE and local computers. GEE does provide JavaScript and Python APIs, but more work needs to be done to incorporate the wealth of well-tested AI algorithms already available in Python into these APIs [94].

4.3.4. Model Deployment Using GEE as Backend

Several publications have used GEE as a backend to their applications (e.g., see [262,263]), taking advantage of the parallel processing capabilities, freely available compute, and large number of datasets. However, one of the main benefits of using GEE is the wide variety of CV and ML algorithms available. In [32], the authors built a custom expert system to map global surface water changes using the platform. Using GEE as a backend allowed them to both run their analysis and then host their resulting maps on an interactive web browser. Another example is Remap [86], an application that allows users to crowd-source LULC observations while using GEE to browse data and make predictions using an RF model.

4.3.5. Vectorizing Data Boundaries

Both [72,147] implemented novel DL architectures using semantic segmentation. In the future, it would be very useful to instead design models capable of instance segmentation. That way, results can be vectorized and used to create global datasets for future mapping research. Having digitized boundaries for individual ecological features is the first step to monitoring them and measuring how they have changed over time. While the authors in [156] demonstrated a novel way of creating data labels via segmentation algorithms, this is still semantic segmentation and the data labels required additional verification. The authors in [233], however, showed this is possible. First, the authors used an RF model for detecting archaeological mounds. They then used an edge detection algorithm after the supervised classification to automatically digitize/vectorize boundary features. Obtaining an accuracy score before digitizing boundaries can give a higher level of confidence in using the resulting dataset in future studies.

4.4. Overarching Challenges and Opportunities

We have provided some separate recommendations for future research for both application-oriented (Section 4.2) and novel/technical (Section 4.3) research above. However, some higher-level combination of the recommendations for application-oriented and more technical perspectives will strengthen the integration of GEE and AI and thus further advance many domain areas of research. There are several opportunities for researchers and practitioners who have interdisciplinary backgrounds and expertise or for research groups with the required complementary expertise to team up and work on problems at the intersection of RS and AI. For example, domain experts working on using ML for exploration and as an aid to human expertise (detailed in Section 4.2.2) will be significantly more productive if there are intuitive, interactive, and visual open-source web app tools to support them in their work. Another example is deep and careful investigation of RS sensors, imagery, and AI towards novel and effective models and algorithms tailored for RS imagery. Particularly, simple but robust ML/CV methods (Section 4.3.1) will be more effective if the researchers and practitioners who have interdisciplinary background and expertise can work towards developing RS image-specific tools, since most CV algorithms are initially designed for camera images and videos, not for remotely sensed satellite images.

One more promising overarching direction is implementing open-source web tools (Section 4.3.2) so that users who do not have a programming background can explore and use existing CV/ML algorithms available on GEE. This could include reimplementing both classic and state-of-the-art CV and/or ML algorithms and deploying them on GEE. A lack of models and model flexibility are two of the most-cited limitations researchers give when using GEE (Section 4.1.2), so by building out the number and type of algorithms on GEE, scientists and practitioners will be better able to do their research in a more seamless way on the platform. Additionally, the GEE Python and JavaScript API documentation pages are not for domain expert users; they are made for web app developers, ML engineers, or for those researchers and practitioners who have an interdisciplinary background. We will stop here, as we do not want to confine research and practitioners’ imagination to propose and develop creative and effective overarching opportunities that will significantly advance various domains, which can leverage the power of GEE and AI.

5. Conclusions

To leverage RS big data for large-scale important challenges such as global climate change, intelligent methods and computation-intensive and -supportive cloud platforms (including cloud storage of huge RS datasets) are critical. GEE is a pilot platform that has great potential to support both challenges (i.e., AI methods and cloud computing platform). Yet to date, many application domains (Section 3) still remain at the proof-of-concept stage regarding leveraging GEE and AI. This trend may relate to a steep learning curve for researchers. Overall, based on our systematic and interactive (Appendix A) review, we contend that GEE integrated with AI has great potential to provide a collaborative and scalable platform for researchers, practitioners, and policymakers to solve critically important problems in various areas. However, many challenges, and thus opportunities, still remain for a deeper and more seamless integration of GEE and AI. This is especially true of the integration between DL and the GEE platform, which is detailed in Section 4.2 and Section 4.3. Up to now, to take advantage of DL with GEE, the time-consuming training process still has to take place outside GEE. Researchers and practitioners either have to train DL models offline on local computers or on a separate cloud computing platform (e.g., Google cloud AI), which is often not freely available to the public. In summary, the deeper and smoother integration of GEE and AI has considerable potential to address major scientific and societal challenges such as climate change and natural hazards risk management.