Data-driven approach for land surface temperature retrieval with machine learning and sentinel-2 data

This research endeavors to advance land surface temperature (LST) prediction accuracy through the development of a sophisticated machine learning model. Leveraging the potential of Sentinel 2 data and atmospheric parameters, we augment Landsat-based LST with MODIS-based LST, enriching the temporal dimensions of our dataset. A distinctive feature of our study is the pioneering use of Sentinel 2 data as inputs for LST prediction, a facet scarcely explored in the existing literature. Our investigation delves into the correlation dynamics between LST and atmospheric parameters. Notably, the study employs a diverse set of machine learning models, including Extra Trees, Random Forests, LightGBM, XGBoost, and Support Vector Regressor. These models collectively exhibit superior performance, with Extra Trees emerging as a standout performer, with a minimal mean absolute error (MAE) of 0.423, a root mean square error (RMSE) of 1.340 °C, and an impressive coefficient of determination ( R 2 ) of 0.984. The exploration of Sentinel 2 data as an input source for LST prediction not only refines predictive accuracy but also opens novel research avenues in the realm of LST dynamics. This study contributes to the existing body of knowledge by introducing innovative methodologies and providing a comprehensive understanding of the intricate correlations influencing LST. • Enables accurate LST estimation using Sentinel-2 data and atmospheric parameters. • Demonstrates the effectiveness of ensemble methods for LST prediction. • Augments Landsat LST data with MODIS LST for improved spatio-temporal resolution. • Highlights the innovative use of Sentinel-2 data for improving LST prediction. • Provides insights into the strong LST correlation with various environmental factors.