A normal form for synchronous land surface temperature and emissivity retrieval using deep learning coupled physical and statistical methods

The innovative normal form, named deep learning-couple-physical and statistical methods (DL-C-PS), has been developed for synchronously retrieving land surface temperature and emissivity (LST&E) from Japan's geostationary meteorological satellite Himawari-8 carrying the Advanced Himawari Imager (AHI). First, geophysical logical reasoning (GLR) and expert knowledge were used to formulate radiative transfer equations (RTEs). Then, a hybrid approach integrating physical and statistical methods was employed to derive the solution, with deep learning (DL) optimizing the solution process. Three band combination schemes were designed to assess the effectiveness of the DL-C-PS normal form. Simulation data from the MODerate spectral resolution atmospheric TRANsmittance mode (MODTRAN) yielded promising results during validation. Root mean square error (RMSE) values were below 1 K for LST and below 0.008 for LSE when using band combinations of four thermal infrared (TIR) bands or at least three TIR bands combined with water vapor information. Cross-validation and in situ validation showed consistent findings with simulation validation. Compared to MODIS LST&E products (MYD21), in most cases, the RMSE values for LST&E were approximately 2 K and less than 0.015 during daytime, and below 1.3 K and 0.017 during night, respectively. Validated against in situ observations, nighttime RMSE values for LST were approximately 1.5 K with correlation coefficient (R) values better than 0.93. The higher RMSE observed in daytime compared to nighttime can be attributed to the influence of the sun's illumination angle on satellite scanning imaging. Overall, this research presented a normal form for multi-parameter estimation by leveraging the optimal calculation of DL and incorporating physical interpretations.