Downscaling Landsat LST to 10 m in Google Earth Engine: assessing SAR contribution and reducing spatial artifacts using Sentinel-1 and Sentinel-2 in Warsaw

Land surface temperature (LST) at high spatial resolution is needed for urban heat island analysis, but Landsat thermal sensors measure at approximately 100 m before resampling to 30 m, limiting effective thermal detail. This study presents a framework for downscaling Landsat LST from 30 m to 10 m using machine learning in Google Earth Engine (GEE), applied to Warsaw, Poland, across five summers (2021-2025). Eight predictors derived from Sentinel-2 (Normalized Difference Vegetation Index [NDVI], Normalized Difference Built-up Index [NDBI], Bare Soil Index [BSI], Modified Normalized Difference Water Index [MNDWI], and albedo) and Sentinel-1 Synthetic Aperture Radar (SAR) data (co-polarized VV backscatter, cross-polarized VH backscatter, and the VV/VH ratio) were used. Four GEE-native algorithms were compared: Random Forest (RF), Gradient Boosting Trees (GBT), Support Vector Machine (SVM), and Classification and Regression Trees (CART). RF achieved the best performance (mean root mean square error [RMSE] = 1.135°C, mean absolute error [MAE] = 0.827°C, coefficient of determination [R 2 ] = 0.896). NDBI was the most important predictor (23.1%), while SAR features contributed 27.6% of total importance. However, an ablation experiment showed that the full model (with SAR) outperformed the optical-only model in three out of five years (2021-2023), while the optical-only model performed better in 2024-2025. The mean RMSE difference was 0.110°C in favor of the full model, indicating that the SAR benefit is condition-dependent. Three strategies for computing shortwave infrared (SWIR)-based indices were compared: computing at native 20 m then resampling to 10 m reduced spatial artifacts at land cover boundaries while achieving the best average accuracy. Indirect validation using temperature transects confirmed that the downscaled LST captured sharper thermal gradients at urban-park and land-water boundaries than the 30 m input. The entire workflow is provided as open-source GEE code, requiring no local computing resources.