Enhanced Land-Surface Temperature Recovery Through Multisensor Data Fusion and Spatial Resolution Improvement

This study presents a novel multimodel fusion approach for enhancing land surface temperature (LST) recovery and spatial resolution improvement. We compare and integrate linear Ridge regression and nonlinear random forest (RF) ensemble models to retrieve high-resolution LST from thermal infrared (TIR) remote sensing data. Initially, LST is computed using a single-channel (SC) algorithm at 30 m (SCLST-30 m) and 90 m (SCLST-90 m). The SCLST-90 m is then enhanced through a multimodel fusion framework that combines both ensemble models with high-resolution predictor features, including surface reflectance, multiple indices (soil adjusted vegetation index, normalized difference vegetation index, normalized difference water index, normalized difference built-up index, and urban index), and digital elevation model. The fusion approach integrates the complementary strengths of both models, with RF capturing nonlinear relationships and Ridge regression providing stability in linear patterns. Cross-validation between SCLST-30 m and the fused LST demonstrates superior performance, with the integrated approach achieving higher accuracy than individual models. Validation against in-situ LST confirms the effectiveness of the fusion framework, yielding R² values of 0.89 and 0.84 for RF and Ridge models respectively. The RMSE values for SCLST-30 m, RF-DLST, and Ridge-DLST are 0.37, 0.38, and 0.45 K, respectively. These results demonstrate the effectiveness of our multimodel fusion approach in achieving accurate recovery and resolution enhancement of LST, offering valuable insights for environmental monitoring and urban planning applications.