A Machine Learning Model Integrating Remote Sensing, Ground Station, and Geospatial Data to Predict Fine-Resolution...

Heat-related morbidity and mortality are increasing due to climate change, emphasizing the need to identify vulnerable areas and people exposed to extreme temperatures. To improve heat stress impact assessment, we developed a replicable machine learning model that integrates remote sensing, ground station, and geospatial data to estimate daily air temperature at a spatial resolution of 100 m × 100 m across the region of Tuscany, Italy. Using a two-stage approach, we first imputed missing land surface temperature data from MODIS using gradient-boosted trees and spatio-temporal predictors. Then, we modeled daily maximum and minimum air temperatures by incorporating monitoring station observations, satellite-derived data (MODIS, Landsat 8), topography, land cover, meteorological variables (ERA5-land), and vegetation indices (NDVI). The model achieved high predictive accuracy, with R² values of 0.95 for Tmax and 0.92 for Tmin, and root mean square errors (RMSE) of 1.95 °C and 1.96 °C, respectively. It effectively captured both temporal (R²: 0.95; 0.94) and spatial (R²: 0.92; 0.72) temperature variations, allowing for the creation of high-resolution maps. These results highlight the potential of integrating Earth Observation and machine learning to generate high-resolution temperature maps, offering valuable insights for urban planning, climate adaptation, and epidemiological studies on heat-related health effects.