Reconstruction of ECOSTRESS Land Surface Temperature Using Deep Learning and Support Vector regression (Case study: Bar-Arieh and Latian watersheds)

Land surface temperature plays a crucial role in regional planning as it serves as a key parameter for various environmental applications, agriculture, meteorology, and other projects. Given the constraints of traditional meteorological stations, remote sensing technology emerges as a valuable tool for obtaining meteorological data. Surface temperature captured in satellite imagery represents the average temperature of a specific area on Earth, calculated by analyzing the energy detected by the sensor (radiance) within the thermal band. This data holds significance across scientific endeavors and facilitates comprehensive analysis and planning processes. Land surface temperature plays a pivotal role in determining the quality of water, soil, vegetation, and watershed health. Therefore, studying the spatial and temporal changes in the land surface temperature under various atmospheric and climatic conditions is essential. Remote sensing has emerged as a powerful tool for monitoring these changes. However, the limited temporal and spatial resolution in satellite data presents a significant challenge when analyzing and estimating temperature variables. In this study, the main aim was to tackle this challenge by leveraging deep learning and machine learning techniques. Specifically, the MODIS, VIIRS, and ECOSTRESS satellite images were used simultaneously to produce daily maps of land surface temperature with a spatial resolution of 70 meters in the Bar-Arieh watershed in Neishabur, Razavi Khorasan province and the Letian Dam watershed in Tehran province during periods with unavailable data. The models employed in this study are highly effective and extensively utilized in environmental modeling applications. The Support Vector Regression (SVR) model stands out as a robust machine learning algorithm, while the Long Short-Term Memory (LSTM) model is recognized as a dominant deep learning model known for its efficacy in various environmental modeling tasks. To accomplish this, we selected three pairs of dates for each basin, resulting in a total of six dates. The images from each pair were imported into the models. Subsequently, we assessed the performance of each model on each date using RMSE, R2, and NRMSE statistics. The findings revealed that the LSTM model outperformed the SVR model in both regions, as evidenced by lower RMSE values and NRMSE coefficients and higher R2 scores. For instance, the LSTM model produced the most accurate results for the Bar-Arieh watershed on June 17, 2020, with an RMSE of 1.81 degrees, R2 of 0.66, and NRMSE of 11.94%. Similarly, in the Latian watershed on June 28, 2019, the LSTM model achieved an RMSE of 1.61 degrees, R2 of 0.83, and NRMSE of 8.6%. The LSTM model demonstrated greater accuracy compared to the SVR model. Being a robust deep learning model, the LSTM model effectively addressed the complexities of the modeling process by extracting features from raw data and considering the time series nature of the data. Given the intricate nature of the Earth's surface temperature dynamics influenced by numerous factors, the modeling results exhibit satisfactory accuracy and precision. The forecasting of the Earth's surface temperature has been carried out for the upcoming 3 to 4 days, considering the potential complex events within this timeframe. This timeframe underscores the validity of the results and the efficacy of the utilized models. Evaluation metrics such as R-squared (R2), Root Mean Square Error (RMSE), and Normalized RMSE (NRMSE) highlight the LSTM model's capability in reconstructing ECOSTRESS land surface temperature values. The observed RMSE values falling within the range of 1.5 to 3 degrees Celsius for the LSTM model signify a commendable level of accuracy in predicting the Earth's surface temperature. Moreover, the normalized RMSE values, spanning from 8 to 17%, serve as a testament to the model's proficiency in capturing the nuanced temperature fluctuations inherent in environmental systems. These values indicate that the model's predictions typically align closely with the observed temperature data, showcasing its ability to effectively account for the complex interactions and dynamics influencing the Earth's surface temperature.