*4.5. Spatial* −*Temporal Time-Series Algorithm*

Advances in computing technology have fostered the development of new and powerful data fusion, gap filling, machine learning, and deep learning techniques that have demonstrated promising results in a wide range of applications [229]. Models to fuse data from multiple sensors and fill gaps can improve UHI monitoring using an ensemble of dense time series of thermal data with a high spatial resolution [199,230]. Zhou et al. [200] presented a new algorithm that focuses on data gap filling using clear observations from

orbit overlap regions to obtain Landsat LST data. Multiple linear regression models were established for each pixel time series to estimate the stable predictions and uncertainties. Liu et al. [83] comprehensively quantified the spatial–temporal patterns of surface urban heat island by investigating the relationship between LST and the land cover types, and the associated landscape components. Such approaches have been used to generate temporally dense and high-resolution LST over long time periods by integrating the observations of Landsat, MODIS, AVHRR, VIIRS, and ECOSTRESS [191,231]. These datasets facilitate subtle analyses of monthly, seasonal, and yearly trends in UHI intensity at regional levels. Machine learning (ML) has become popular in UHI and UHIRIP, but its use has remained restricted to predicting, rather than understanding, the natural world. ML techniques may not be the solution to all the problems remotely sensed data might have. However, these techniques provide a powerful set of tools that deserve serious attention to deal with some relevant UHI and UHIRIP remotely sensed data problems [232]. Lucas [233] points out that ML differs from the broader field of statistics in two respects: (1) the estimation of parameters that relate to the real world is less emphasized, and (2) the driver of the predictions is expected to be the data rather than expert opinion and careful selection of plausible mechanistic models. The Google Earth engine (GEE) is a cloud-based platform for planetary-scale geospatial analysis that brings Google's massive computational capabilities to bear on a variety of high-impact societal and environmental issues [234]. GEE has many functions that could be used to analyze UHI and UHIRIP at local, regional, and global levels (Figure 5). Some research has generated consistent large-scale UHI and UHIRIP analysis based on optimal data and ML algorithm selection using GEE [235,236]. The advanced GEE cloud-based platform and the large number of geosciences and remote sensing datasets archived in GEE were used to analyze land the cover dynamics (236), and the results showed the advantages of using GEE to analyze the spatiotemporal dynamics of the LULCC, vegetation cover, LST, and climate for a long time series, and highlighted the importance of environmental protection. The power here lies in the way a scientist defines their questions and uses machine learning alongside other methods. Techniques for data analysis and interpretation that fully incorporate the temporal dimension remain an area of intense research and represent an important challenge for operational UHI and UHIRIP monitoring.

**Figure 5.** Simplified system diagram using the Google Cloud platform and Google Earth engine for monitoring UHI and UHIRIP.

### **5. Summary of UHI and UHIRIP Based on Remotely Sensed Data**

This review provides an overview of research on UHI and UHIRIP based on remote sensing techniques, sensors, and algorithms, as listed in Tables 3–6, respectively. Much work has been completed on UHIRIP in the last four decades, and we have endeavored to keep updated with new methods and results. A significant research limitation still exists: the quantification of UHI and its regional impacts using high-resolution timeseries remotely sensed thermal data in the urban and non-urban interface. Some of the algorithms listed in Table 5 may be the most practical approaches to assess UHI in core urban areas of cities and surrounding areas, but characterization of UHI across broad areas is necessary in order to inform monitoring, reporting, science, and policy. Being able to relate LST to UHI is especially important when such datasets are being used to inform policy decisions or to communicate outside of the scientific community. The increasing availability of remotely sensed data across a range of spatial resolutions and temporal frequencies, and technological improvements in image processing capacity and storage, have led to advances in the methods used to monitor UHI more frequently and accurately.

Assessing the uncertainty and accuracy of UHI data is important. A sensitivity analysis not only provides a framework for assessing the potential for bias and the extent of uncertainty in UHI estimates, but also reveals significant factors that determine the extent of UHIRIP in the urban and non-urban interface. Oleson et al. [237] developed an approach to evaluate the robustness of models used to simulate urban heat islands in different environments. The findings indicated that heat storage and sensible heat flux are most sensitive to uncertainties in the input parameters within the atmospheric and surface conditions considered. Sensitivity studies indicate that it is important to not only to accurately characterizing the structure of the urban area, but also to ensuring that the input data reflect the thermal admittance properties of each of the city surfaces.

Currently, a wide variety of methods are employed to characterize UHI for major cities worldwide (Table 5), although most of the applications cited were limited to small areas because of data availability and constraints of storage and computing resources. With the development of gap filling and data fusion models [238], advances in high-performance computing (HPC), and cheaper storage, applications based on high-resolution time series at larger or even regional scales will become the mainstream in the near future [199,231]. While much of the methodological variation described here will persist, future methods will evolve and adapt to greater data volumes and processing capabilities [239]. Legacy change mapping methods that rely on analyst interactions with individual scenes should decline over time given the improved ability to process and characterize time series of rich high-resolution thermal data. However, such spatial −temporal methods that are based on gap filling and data fusion should match the institutional requirements for accuracy. Near-term research objectives will require robust validation datasets in establishing which data-intensive methods are the most appropriate for quantifying UHI over large areas. Techniques for LST data analysis and interpretation that fully incorporate the temporal dimension still require intense research and represent an important challenge for operational UHI research in order to meet managemen<sup>t</sup> needs.

Technological advances that include machining leaning and artificial intelligence in UHI and UHIRIP using remotely sensed data have led to an explosion of UHI and UHIRIP profiling data from large numbers of multiple data sources [233,240]. This rapid increase in the remotely sensed data dimension and acquisition rate is challenging conventional analysis strategies. Modern machine learning methods, such as deep learning, promise to leverage very large datasets for finding a hidden structure within them, and for making accurate predictions [232]. Deep learning methods are a powerful complement to classical machine learning tools and other analysis strategies, and have been used in a number of applications in UHII and remotely sensed image analyses [241]. The explainable artificial intelligence in UHII and UHIRIP modeling has become more and more important [242]. Interpretable machine learning methods either target a direct understanding of the model

architecture (i.e., model-based interpretability) or interpret the model by analyzing the model behavior (post hoc interpretability) [242].

Currently, most of the time-series algorithms used to map UHIRIP include data from the temporal domain of AVHRR and MODIS, and the spatial domain of the data is almost entirely neglected. Although these datasets with a lower spatial resolution and higher temporal frequency can detect a change of UHI in real time, they often lack pertinent spatial detail. Even though many UHI analysis algorithms have been developed [60], most of the UHI monitoring data derived from the Landsat archive are provided in a time frame that is not near enough to real time to be relevant for specific managemen<sup>t</sup> needs. With the advances in HPC and cheaper storage, applications based on Landsat time series at continental or even global scales will be the mainstream in the next few years.

To date, information from Landsat time-series thermal data has taken the form of statistical metrics, change metrics, pattern distribution, or trend components used in UHI impact applications [243]. Improvement of existing approaches, as well as the inclusion of novel techniques, often imported and adapted from other disciplines, are important to fully capitalize on the thermal data in order to produce monthly, seasonal, and annual LST results that meet a wide range of UHI and UHIRIP research needs. Landsat-9, which will be launched in September 2021, will continue collecting images of the Earth's surface in visible, near-infrared, and shortwave-infrared bands, as well as the thermal infrared radiation, or heat, of the Earth's surface from two thermal bands. The future European Space Agency's LSTM (Land Surface Temperature Monitoring) or Sentinel 8 mission will carry a high spatial −temporal resolution thermal infrared sensor to provide records of land-surface temperature. Land-surface temperature measurements are key variables to understand and respond to climate variability and natural hazards, such as urban heat island issues. The main objective of LSTM is to deliver global high spatial −temporal dayand night-time land surface temperature measurements. LSTM will operate from a low-Earth, polar orbit, to map both land-surface temperature and rates of evapotranspiration. It will be able to identify the temperatures of individual fields and image the Earth every three days at a 50 m resolution. Another future thermal sensor is Thermal infraRed Imaging Satellite for High-resolution Natural resource Assessment (TRISHNA), which is a future high-resolution space-time mission in the thermal infrared (TIR) led jointly by the French (CNES) and Indian (ISRO) space agencies. One of scientific objectives guiding the definition of the mission is the monitoring of the urban environment. TRISHNA will be positioned on a polar orbit and provide a revisit of three passages over 8 days with global coverage. The time of passage around 13:00 p.m. LST allows thermal data to be collected in the middle of the day, but also in the middle of the night. The instrument will offer four thermal channels (8.6 μm, 9.1 μm, 10.4 μm, and 11.6 μm) and six optical channels (485 nm, 555 nm, 650 nm, 860 nm, 1380 nm, and 1650 nm) with a spatial resolution between 50 m and 60 m for all channels. All of these observations acquired from thermal remote sensing will provide more valuable information for natural resource management, hazard monitoring, and scientific research and applications.
