Physics-guided thermal infrared and multispectral fusion for Land surface temperature monitoring

Accurate land surface temperature (LST) monitoring increasingly depends on the integration of complementary Earth observation modalities. Thermal infrared observations provide direct evidence of land-atmosphere energy exchange and night-time thermal patterns, whereas optical and multispectral imagery supplies land-cover, vegetation, and structural context. However, differences in imaging mechanism, acquisition time, cloud contamination, and spatial resolution can weaken the physical consistency of fused products. This letter presents a physics-guided multimodal deep learning model for fusing ECOSTRESS thermal infrared observations with Sentinel-2 optical/multispectral imagery for high-resolution LST reconstruction. The model combines modality-specific encoders, a radiative-transfer confidence term embedded in cross-attention, multi-scale spatial aggregation, and surface-energy-balance regularization. Experiments on ten urban and agricultural regions show that the proposed model improves LST retrieval accuracy and physical consistency over five adapted baselines. Across repeated runs, it achieves an average MAE of 1.47 +/- 0.04 deg C and RMSE of 2.51 +/- 0.06 deg C, while reducing energy-balance consistency error by more than 35% relative to non-physical fusion models. The scope of the study is therefore focused on physically consistent LST reconstruction under night-time, partially cloudy, and cross-sensor acquisition conditions, rather than general environmental risk mapping.