Downscaling MODIS land surface temperature to 90 m using random forest regression to assess transferability

The observable spatiotemporal trade-offs among different satellite sensors necessitate the downscaling of optical and thermal data to identify Earth’s surface features and analyze temperature variations at finer scales. This study employs a random forest (RF) regression model to downscale Moderate Resolution Imaging Spectroradiometer (MODIS) land surface temperature (LST) data from a 1-km spatial resolution to 90 m. The model is trained on a primary region and applied to a spatially distinct secondary region to evaluate its transferability. Six acquisition dates, each representing a different month, were selected for downscaling. For contextual performance assessment, eXtreme Gradient Boosting (XGBoost) and Categorical Boosting (CatBoost) models were also trained and evaluated using the same setup. In the primary region, the RF model showed minimal root mean square error (RMSE) deviation across all acquisition dates, underscoring the importance of incorporating multiple variables. Comparable performance was observed with XGBoost and CatBoost in the primary region. In the secondary region, RF demonstrated strong generalizability, with the lowest RMSEs recorded in November (0.85) and February (0.96), and R 2 values of 0.65 or higher across all months. In contrast, XGBoost and CatBoost exhibited inconsistencies in the secondary region, with CatBoost failing notably in June, further highlighting RF’s robustness in transferability. The study’s outcomes suggest that RF regression is a promising and effective method for downscaling coarse-resolution LST data in similar geo-environmental conditions.