**1. Introduction**

Wetlands have been identified as valuable resources that provide a variety of ecological and socioeconomic benefits [1], but they are also threatened due to human activities, such as agricultural intensification and climate change [2]. These threats and others make monitoring the spatiotemporal variation of wetlands' hydrological processes crucial to their effective management. Here, by hydrological processes, we refer to wetlands' highly variable environments characterized by hydric soils temporarily or permanently flooded by water. When dry, wetlands resemble surrounding uplands, whereas when inundated, they can have either moist soils or surface water that ranges from centimeters to meters deep. There are also high levels of diversity in wetland cover classes, wherein some inundated wetlands are filled with emergent or submerged vegetation, and others are absent of all vegetation.

Though the dynamic nature of wetlands makes them ecologically valuable to numerous flora and fauna, this also makes them difficult to monitor [3,4]. Monitoring depressional

**Citation:** Sahour, H.; Kemink, K.M.; O'Connell, J. Integrating SAR and Optical Remote Sensing for Conservation-Targeted Wetlands Mapping. *Remote Sens.* **2022**, *14*, 159. https://doi.org/10.3390/rs14010159

Academic Editors: Alban Kuriqi and Luis Garrote

Received: 12 December 2021 Accepted: 28 December 2021 Published: 30 December 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

wetlands can also be challenging because these highly dynamic systems are primarily dependent on climate and local weather systems for ponding, and can often be relatively small (<40 ha) [5,6]. The interplay among water, vegetation, and soil results in wetlands that share spectral reflectance characteristics of both aquatic and terrestrial environments. Accurate and unbiased estimates of wetland surface water across the range of natural conditions have therefore eluded scientists.

The Prairie Pothole Region (PPR) is one example of a high-risk, dynamic wetland system composed of millions of temporary, seasonal, and semi-permanent depressional wetlands, called potholes. These potholes are known for their cycles of drought and deluge, which drive important ecosystem functions, such as the abundance of aquatic invertebrates [5]. The PPR covers an extensive area of approximately 750,000 km2, including parts of five US states and three Canadian provinces (Figure 1), and provides habitat for over 50% of North America's migratory waterfowl [7,8]. Hydroperiods in the potholes vary from days to years, but seasonal wetlands that maintain water for less than four months are common [9,10]. Reduced surface water area and changes in hydrology are common in PPR wetlands, for example, as caused by tile draining to allow for higher agricultural production [11], or upland sediment erosion into wetlands, which, though a natural process, is often accelerated by agricultural activity, which fills potholes, and reduces their volume [12]. The total wetland loss in the PPR caused by climate change and human activity was estimated to be 30,000 ha between 1997 and 2009 [10]. A resulting shift towards smaller wetlands and shortened hydroperiods [13–15] has underscored a need to understand how these altered hydrological conditions affect ecosystem services and habitat provisioning at broad spatial scales, which starts with an accurate and repeatable estimate of spatial variation in wetland surface water.

Remote sensing analysis can provide broad-scale spatial and temporal information about wetland surface water [16,17]. Previous studies utilized various remote sensing technologies to monitor wetlands across the PPR [8,18]. For example, [8] used highresolution NAIP data and LIDAR Digital Elevation Models (DEMs) to map PPR wetland inundation, and tested the results with the Wildlife Service National Wetlands Inventory (NWI). However, though NAIP and DEMs can provide fine spatial resolution data (<1 m), these methods cannot capture temporal variation within a season, as NAIP and LiDAR data are not collected intraannually. Optical sensors, such as Sentinel-2 and Landsat, can detect surface water, and have often been used with success for deep, permanent, large water bodies [19,20]. For example, the Joint Research Centre (JRC) provided Landsat-derived surface water products useful for capturing large wetlands. However, the JRC and other products that rely on moderate resolution spectral data often underperform in detecting water in small potholes with dense vegetation canopies and mixed pixels. Others have used Sentinel-1 synthetic aperture radar (SAR) data (spatial resolution: 10 m) to map water extent in the PPR with reasonable success [21,22], as SAR data is robust to cloud cover, and 10 m data provide reasonable spatial resolution. However, no study has solved all of the challenges for mapping the spatial and temporal variation of surface water in the PPR, and made their algorithm available for long-term monitoring by the research and conservation community. There is a need for open-science algorithms that capture the variation of surface water, can map water even below emergent vegetation, and still represent surface water in smaller potholes.

This study relies on geospatial informatics, which is an expanding field, and includes remote sensing of landscape-scale big data, the development of machine learning tools, and integration with High-Performance Computational (HPC) cloud computing resources. Geospatial informatics offers a unique opportunity for the fast processing of broad-scale remote sensing data in a short time, providing a more comprehensive set of applications, and addressing the limitation of traditional methods [23,24]. The Google Earth Engine (GEE) cloud geospatial computing platform provides a web-based interface to fast parallel processing on Google HPCs with planetary-scale analysis capabilities. The GEE provides a multi-petabyte catalog of global satellite and geospatial datasets [25], such as Landsat, MODIS, and Sentinels. It also gives users the ability to analyze, manipulate, and map the results, and create web-based applications to repeat the analysis [26]. As part of our work, we utilized the capabilities of GEE to create an open-source algorithm for mapping wetlands that can readily be shared with conservation managers and the science community for continued use and development.

**Figure 1.** The location of the Prairie Pothole Region (PPR) (**A**); the location of the study site in the US and the state of North Dakota (**B**); distribution of ground truth points in the study site (**C**).

To help solve the historical problems of surface water mapping in the PPR, this paper presents a multi-sensor fusion approach that integrates selected fine-resolution (10-m) bands of Sentinel-2 with 10-m Sentinel-1 SAR data, allowing an estimate of both large and small inundated areas. The integration of SAR with optical data also offers complementary information, and can significantly improve the interpretation and classification of results [27,28], for example, by allowing surface water estimates beneath closed-canopy herbaceous vegetation. Altogether, this study aims to provide scalable surface water estimates that can assist with habitat models for wetland-dependent organisms, such as waterbirds or aquatic invertebrates. We will provide our algorithm in a format that can be freely shared and readily implemented by those with minimal coding and modeling experience, such as conservation managers. We achieved this through the following objectives: (1) we developed an open-source framework to map the spatial variation in wetland surface inundation and vegetation based on Sentinel-1 SAR data and Sentinel-2 high-resolution bands within the GEE platform; (2) we deployed this algorithm over a portion of PPR in the high priority conservation area of the PPR; (3) we analyzed the accuracy of this algorithm for generating the information needed for setting conservation targets.

#### **2. Study Area**

Our study area was a portion of PPR in North Dakota, USA (Figure 1). The area is dominated by natural grasslands, agricultural areas, and a relatively high density of potholes, which, in this area, often present as small and elliptical water bodies. These numerous small wetlands provide natural habitats for wetland-dependent animals and plant species. We selected this area due to the high density of small potholes, high conservation priority, and availability of ground truth data. We mainly focused our algorithm on a subset of the PPR identified as a high priority conservation site for waterfowl by the United States Fish and Wildlife Service.

#### **3. Data**

The data includes a set of aerial imagery to serve as ground truth data, the highresolution bands (bands 2, 3, 4, and 8) of Sentinel-2, and C-band SAR data Sentinel-1 sensor. We describe the details of the dataset below.

#### *3.1. Ground Truth*

Researchers from Duck Unlimited Inc., a non-profit conservation organization, provided the ground truth data. These data include georeferenced aerial photographs of the PPR wetlands in North Dakota collected through a partnership with the United States Fish and Wildlife Service (USFWS). The USFWS used a fixed-wing aircraft to collect imagery in a 1.5 m spatial resolution. If necessary, the images were orthorectified by technicians or research scientists, and used to estimate wet areas during spring and summer for the research projects. We used the summer data of two years (2016 and 2017). These datasets were provided in shapefile formats, and showed wetland boundaries, delineating dry and inundated wetland areas. Some of these wetlands also contained emergent vegetation cover, as identified by field observers (range: 0–80% vegetation cover).

We examined the spectral reflectance of wetland and non-wetland classes, which differed substantially, as indicated by a plot generated for a portion of the study area (Figure 2). The spectral characteristics of wetlands and open water especially differ due to mixed pixels, differences in water depth, the potential presence of vegetation, and variation in water turbidity. Compared to forest and agriculture, deep open water exhibited lower spectral reflectance, as water rapidly absorbs electromagnetic radiation, especially longer wavelengths, and attenuation increases with water depth. The spectral reflectance of wetlands is intermediate to upland vegetation and open water, making wetlands a distinct and highly variable land cover type. Wetlands and moist soils show a dampened near-infrared (NIR) and shortwave infrared (SWIR) reflectance compared with upland vegetation, but are too shallow to attenuate all electromagnetic radiation, as often occurs in deep open water. The spectral characteristics of wetlands will also change rapidly with inundation and vegetation status. To account for this in our ground truth point selection, we selected random points within the digitized wetland surface water area polygon shapefiles to provide the ground truth pixels in GEE. We also included nonwetland training data that represented agriculture, forest, and urban areas. We collected those points using visual observation of high-resolution Google Earth images. The total

deviation of spectral reflectance of pixels for each land cover type.

number of points (including wetland and non-wetland classes) for the years 2016 and 2017 were 895 and 2231, respectively.

**Figure 2.** Spectral reflectance during summer months from Sentinel-2 optical bands of large bodies of deep open water compared to inundated wetlands and other land cover types in a portion of the study area. Wetland water shows different spectral characteristics compared to deep open water, likely due to the presence of submerged and emergent vegetation. The error bar shows the standard

Additionally, we provided two out-of-sample subsets for small-vegetated (1440 points) and non-vegetated wetlands (1680 points). Ducks Unlimited provided the vegetation data within the surface water polygons. We used those additional points in a separate accuracy assessment process to evaluate the performance of our method for the smallest wetlands, which are the most challenging to classify as they contain the highest proportion of mixed pixels. The average time difference between ground truth data (wetlands and non-wetlands) and satellite data acquisition was one month.

#### *3.2. Sentinel-1*

Sentinel-1 obtains C-band synthetic aperture radar (SAR) images at various polarizations and resolutions. C-band Level-1 Ground Range Detected (GRD) data were obtained through GEE. These data were collected in the Interferometric Wide (IW) swath mode with a spatial resolution of 10 m, a swath width of 250 km, and a repeat cycle of 12 days. These data are available in GEE as preprocessed datasets that express each pixel's backscatter coefficient (σ◦) in decibels (dB). The preprocessing steps include applying orbit files, thermal noise removal, radiometric calibration, and orthorectification (terrain correction). This study used two polarization modes: single co-polarization with vertical transmits and receive (VV), and dual-band co-polarization with vertical transmit and horizontal receive (VH). A total of 20 ascending orbit Sentinel-1 SAR scenes spanning two months were collected over the study area. We used median values of the S1 temporal time series in the multisensory band composite. A median composite can provide a cleaner image with reduced speckle noise [29]. These data were acquired from July to September 2016. The descending orbit data were excluded from the study because they lacked sufficient coverage orbit over the study area (Table 1). Unlike optical sensors, SAR data can be acquired day and night and during cloudy conditions, completely independent of solar radiation, which is particularly important in high latitudes, and increases the availability

of multi-temporal observations for assessing wetland hydroperiods. Moreover, SAR data is sensitive to both open water and below-canopy inundation, making it advantageous to identify inundation in vegetated wetlands [30]. The C-band SAR data of Sentinel-1 is also known to be useful for the discrimination of water and non-water classes in non-forested wetlands with short herbaceous vegetation (e.g., bog and fen) [31]. This is in contrast to the longer wavelengths, such as L-band SAR data, that are preferred to detect inundation areas in forests due to higher penetration depth [32].

**Table 1.** Multisensor satellite data and spectral reflectance indices were used for supervised classification to identify the water bodies in the study area.

