A Novel Fully Coupled Physical–Statistical–Deep Learning Method for Retrieving Near-Surface Air Temperature from Multisource Data

Abstract

Retrieval of near-surface air temperature (NSAT) from remote sensing data is often ill-posed because of insufficient observational information. Many factors influence the NSAT, which can lead to the instability of the accuracy of traditional algorithms. To overcome this problem, in this study, a fully coupled framework was developed to robustly retrieve NSAT from thermal remote sensing data, integrating physical, statistical, and deep learning methods (PS-DL). Based on physical derivation, the optimal combinations of remote sensing bands were chosen for building the inversion equations to retrieve NSAT, and deep learning was used to optimize the calculations. Multisource data (physical model simulations, remote sensing data, and assimilation products) were used to establish the training and test databases. The NSAT retrieval accuracy was enhanced using the land surface temperature (LST) and land surface emissivity (LSE) as prior knowledge. The highest mean absolute error (MAE) and root-mean-square error (RMSE) of the retrieved NSAT data were 0.78 K and 0.89 K, respectively. In a cross-validation against the China Meteorological Forcing Dataset (CMFD), the MAE and RMSE were 1.00 K and 1.29 K, respectively. The actual inversion MAE and RMSE for the optimal band combination were 1.21 K and 1.33 K, respectively. The proposed method effectively overcomes the limitations of traditional methods as the inversion accuracy is enhanced by adding the information of atmospheric water vapor and more bands, and the applicability (portability) of the algorithm is enhanced using LST and LSE as prior knowledge. This model can become a general inversion paradigm for geophysical parameter retrieval, which is of milestone significance because of its accuracy and the ability to allow deep learning for physical interpretation.