A Physical Mechanism-Constrained Deep Learning Hybrid Model for Retrieving Surface Temperature Under Nonprecipitation Cloud

The wide range acquisition of all-weather land surface temperatures (LSTs) from passive microwave (PMW) remotely sensed data contributes to understanding the land-atmosphere interactions, surface energy balance, and the global water cycle. Although significant progress has been made in PMW-based LST retrieval using statistical models, physical models, and machine learning methods, there remains a need to propose a model with high accuracy alongside strong physical interpretability and good generalization ability. This article aims to develop a physics-constrained deep learning (DL) hybrid model to obtain accurate LSTs under nonprecipitation clouds and then compare it with the pure physical and DL models. The hybrid model is developed by incorporating the physical loss function into the convolutional neural network, inheriting the advantages of the physical and DL models. Results show that the constructed model achieved good performance with a root mean square error (RMSE) of 1.60 K and a mean absolute error (MAE) of 1.26 K in the simulated data. Sensitivity analysis revealed that the hybrid model is less sensitive to input parameters than the pure physical and pure DL models and exhibits robustness across varying land surface and atmospheric conditions. Furthermore, during the evaluation using U.S. Surface Radiation Budget (SURFRAD) site data, the hybrid model yielded RMSEs of 4.37 and 3.48 K for day and night, respectively, with Advanced Microwave Scanning Radiometer 2 (AMSR2) and ERA5 data from 2012 to 2024, while outperforming the other two models at each SURFRAD site. The spatiotemporal applicability of the hybrid model further highlighted its superior generalization ability in mapping LSTs. We believe that the evident strength of the developed model over the traditional pure physical and DL models is attributed to its good accuracy, robustness to uncertainties in input parameters, and physical interpretability, which will benefit the other parameter estimates.