Machine learning land surface temperature downscaling method based on Landsat 9 and Sentinel-2 satellite feature interaction

As a core essential climate variable (ECV) in the Global Climate Observing System, land surface temperature (LST) plays a pivotal role in climate change monitoring, urban thermal environment assessment, agricultural management, and ecosystem surveillance. To obtain high-precision LST data, a novel machine learning-based LST downscaling framework integrating feature interaction optimization and Shapley additive explanations (SHAP) scoring was proposed. SHAP scoring was employed for feature importance analysis to identify optimal predictors, while 10 distinct models were comparatively evaluated to establish a high-resolution downscaling framework adaptable to homogeneous surface characteristics. The results show that the SHAP-based feature selection significantly enhanced prediction accuracy by prioritizing nonlinear determinants. The red-blue band interaction feature demonstrated consistent dominance across all algorithms (XGBoost, LightGBM, GradientBoost), exhibiting both the broadest SHAP value range (−2.0 to 2.0) and the highest relative contribution weight. By explicitly addressing spatial heterogeneity, the spatial random forest (SRF) model achieved superior downscaling performance, particularly in vegetated regions. It generated reliable 10 m-resolution LST estimates (R2 = 0.74, RMSE = 6.28°C), demonstrating robust generalization capabilities in complex terrain conditions. The SHAP-based land surface temperature downscaling method can effectively capture the nonlinear interactions among spectral, topographic, and other features, demonstrating high accuracy and strong physical interpretability in high-resolution temperature retrieval over areas dominated by homogeneous vegetation.