A Machine Learning-Based Method for Modis LST Downscaling and Reconstruction in Diverse Regions and Seasons

Land Surface Temperature (LST) is a critical environmental parameter for describing biophysical processes at the Earth's surface, applicable at regional or global scales. High-resolution LST downscaling is of paramount importance for research in areas such as urban heat islands, crop health, natural disasters, and climate change. However, sensors with a high revisit time are limited in their ability to provide detailed spatial information. Therefore, downscaling of LST products with coarse spatial resolution is considered a necessary and inevitable process to address this challenge and offer valuable data solutions. In this study, we have developed a LST downscaling and reconstruction method based on machine learning. By introducing predictor variables to characterize the spatial distribution of LST, we have successfully downscaled and reconstructed LST data from the Moderate Resolution Imaging Spectroradiometer (MODIS) with a resolution of 990m to 90m. Our analysis was carried out on various land surface cover types, including urban areas, croplands, mountainous regions, and water bodies. The comprehensive comparison and analysis of model performance have shown that in different seasons, the downscaling and reconstructed performance in mountainous areas is the best, with R-squared (R
²
) values reaching 0.71 and 0.80, and Root Mean Square Error (RMSE) values of 1.72 and 1.64, respectively. The proposed model was subsequently applied for downscaling and reconstructing surface temperatures in cloud-covered areas, and the results were confirmed through both visual assessments and statistical measurements of reconstruction accuracies.