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

Introduction: Land surface temperature (LST) is a critical variable in regional and local planning, playing a key role in climate change studies, hydrological modeling, vegetation monitoring, urban heat island effects, urban development, global warming, agricultural conditions, and influencing plant growth rates and timing. As an essential input parameter for most environmental models, the accuracy of LST estimation directly impacts the reliability of model outputs. Therefore, a precise and optimized study of the spatiotemporal variations of LST under different weather and climatic conditions is necessary. However, due to the limited coverage of meteorological stations, remote sensing serves as a fundamental tool for acquiring large-scale meteorological data. LST derived from satellite imagery represents the average pixel temperature over a specific land surface area, calculated based on the thermal band radiance detected by the sensor. Despite its advantages, a major challenge in LST estimation is the lack of simultaneous high temporal and spatial resolution in satellite data, which complicates accurate temperature analysis.
Materials and Methods: This study utilized deep learning and machine learning techniques to generate daily LST maps from MODIS, VIIRS, and ECOSTRESS satellite imagery, focusing on data-gap periods in two watersheds: Bar-Arieh (Neyshabur, Razavi Khorasan Province) and Latyan Dam (Tehran Province). The ECOSTRESS sensor, with a high spatial resolution of 70 meters, was selected for its ability to provide detailed LST measurements. However, due to temporal gaps in ECOSTRESS data, reconstruction was necessary. Two prominent models were employed: Support Vector Regression (SVR), a widely used machine learning model for regression tasks, and Long Short-Term Memory (LSTM), a powerful deep learning model specialized in processing sequential and time-series data. For each watershed, three pairs of dates (six dates in total) were selected, ensuring that the images were temporally aligned for model training and validation. The models were fed with corresponding satellite data from MODIS, VIIRS, and ECOSTRESS to predict missing LST values. Performance evaluation was conducted using three statistical metrics: Root Mean Square Error (RMSE), coefficient of determination (R²), and Normalized Root Mean Square Error (NRMSE).
Results and Discussion: The comparative analysis of RMSE, R², and NRMSE values demonstrated the superior performance of the LSTM model over SVR in reconstructing LST values. In the Bar-Arieh watershed, the best results were obtained for June 17, 2020, with RMSE = 1.81°C, R² = 0.66, and NRMSE = 11.94%. For the Latyan Dam watershed, the most accurate predictions were recorded on June 28, 2019, with RMSE = 1.61°C, R² = 0.83, and NRMSE = 8.65%. The LSTM model's success can be attributed to its inherent ability to extract complex features from raw data while accounting for temporal dependencies, making it highly effective for time-series forecasting. Unlike traditional machine learning models, LSTM captures long-term patterns and nonlinear relationships within the data, enabling it to handle the inherent complexities of LST dynamics. The model's robustness was further validated by its ability to reconstruct LST for periods with missing data (3–4 days ahead), accounting for the temporal intervals between successive ECOSTRESS overpasses. Despite the challenges posed by the variability of factors influencing LST—such as land cover changes, atmospheric conditions, and diurnal temperature fluctuations—the model's predictions remained within an acceptable error range. The RMSE values (ranging between 1.5°C and 3°C) indicate a reasonably accurate estimation, given the natural variability of LST (0°C to nearly 35°C in the study areas). Additionally, the NRMSE values (mostly between 8% and 17%) confirm the model's reliability, especially considering the complexity of environmental processes.
Conclusion:  The study highlights the effectiveness of deep learning, particularly the LSTM model, in reconstructing high-resolution LST data from multi-sensor satellite imagery. The evaluation metrics (R², RMSE, and NRMSE) consistently demonstrated LSTM's superiority over SVR, reinforcing its suitability for time-series-based environmental modeling. Given the dynamic nature of LST and the multitude of influencing factors, the model's ability to maintain low error margins (RMSE ~1.5–3°C) and high explanatory power (R² up to 0.83) underscores its potential for operational use in remote sensing applications. Furthermore, the normalized error values (NRMSE between 8% and 17%) suggest that the model performs reliably even under complex environmental conditions. These results are particularly significant for applications requiring high spatiotemporal resolution LST data, such as urban heat island monitoring, precision agriculture, and climate change studies. Future research could explore the integration of additional satellite datasets or hybrid modeling approaches to further enhance prediction accuracy. Overall, this study provides a robust framework for LST estimation in data-scarce scenarios, contributing to improved environmental monitoring and decision-making processes.