A validation procedure to assess the precision and accuracy of satellite-derived land surface temperature

Land Surface Temperature (LST) is a fundamental variable for the understanding surface–atmosphere interactions, urban climate dynamics and the Earth’s energy balance. As the use of satellite-derived LST products continues to expand, ensuring their reliability becomes increasingly important, particularly for recently developed downscaled products that provide enhanced spatial detail. This thesis presents a comprehensive validation framework designed to evaluate both the accuracy and precision of satellite-derived LST datasets and their highresolution downscaled counterparts. The study is organized into two complementary phases. The first phase focuses on the validation of original-resolution LST products. A ten-year archive (2013–2022) of Moderate Resolution Imaging Spectroradiometer (MODIS) Terra and Landsat 8-9 observations was assessed against "GOLD" standard quality ground measurements collected at seven climatically diverse reference stations in the United States. This analysis was complemented by a cross-satellite comparison conducted across Italy during 2023. The results show a clear seasonal dependence of retrieval performance, with higher accuracy and precision during winter than in summer. In addition, both sensors exhibit sensitivity to land cover and climatic conditions. The second phase investigates the performance of a Machine Learning (ML) downscaling algorithm developed to increase the spatial resolution of satellite LST products. The algorithm enhances Landsatderived LST from 100 m to 10 m using Sentinel-2 imagery, and MODISderived LST from 1 km to 100 m using Landsat data. Three successive product versions generated between 2023 and 2025 for the cities of Milan, Florence and Naples were evaluated through a multi step validation procedure. The assessment included comparisons with original-resolution satellite products, independent air temperature observations from regional meteorological networks and a final statistical evaluation based on Ordinary Least Squares regression combined with Huber–weighted outlier detection. This approach replaced the previously adopted RANSAC methodology, providing a more transparent and reproducible framework for residual analysis and model diagnostics. Overall, the downscaled products successfully preserved the spatial and temporal characteristics of the original LST maps, with coefficients of determination generally exceeding 0.90. The products also reproduced the expected differences between surface temperature and air temperature, including their pronounced seasonal and diurnal variability. Nevertheless, the analysis highlighted limitations associated with specific sensors and product versions, some of which exhibited a reduced temperature range. The results demonstrate that the proposed validation framework can be effectively applied across different sensors, spatial resolutions and levels of spatial enhancement. The methodology provides a robust basis for assessing the quality of downscaled LST products and supports their application in local climate studies. At the same time, the findings identify several directions for future improvements, including the refinement of the Terra-Aqua compositing procedure and the explicit integration of land cover information into the pixel-by-pixel quantitative analysis of downscaled LST products. This research has been conducted within the framework of the Italian National PhD Program in Earth Observation funded by the European Union — NextGenerationEU under the Italian National Recovery and Resilience Plan (NRRP).