Physics-Informed Deep Learning for FY-3D MWRI Brightness Temperature Reconstruction with Static-Dynamic Physical Constraints

The integrity and accuracy of brightness temperature (BT) are critical prerequisites for reliable high-precision retrieval products. The Microwave Radiation Imager (MWRI) onboard the FY-3D satellite is capable of providing all-weather BT observations. However, constrained by orbital gaps, extreme weather and instrument detection er-rors, brightness temperature observations are prone to widespread missing values and anomalous noise, greatly limiting their practical application. To address the above da-ta defects, this study constructs a U-Net-based convolutional neural network and pro-poses a targeted reconstruction method for MWRI brightness temperature data. Con-sidering that the low-frequency channels of MWRI are highly sensitive to complex underlying surface conditions, the proposed network incorporates static prior con-straints, including terrain elevation and vegetation type, alongside dynamic physical constraints, such as diurnal temperature variation and latitudinal/longitudinal gradi-ents. Utilizing the 10.7 GHz vertically polarized channel BT data of FY-3D MWRI from 2023, ablation studies and accuracy evaluations are conducted. The results demon-strate that the incorporation of physical constraints significantly improves the model's restoration performance and effectively resolves the discontinuity issue in the recon-structed BT. Validations demonstrate that guided by synchronous land surface tem-perature (LST), the proposed method not only accurately reconstructs the spatial dis-tribution of the original BTs but also faithfully captures their instantaneous dynamic states. The mean biases in the land and ocean test regions are -0.580 K and -0.064 K, respectively. Furthermore, the annual average standard deviation of the bias is 1.163 K for the land and reaches 0.598 K for the ocean. The reconstruction network proposed in this paper effectively enhances the completeness and reliability of FY-3D MWRI BT data, thereby providing a robust data foundation for the subsequent retrieval of oce-anic, atmospheric, and land surface parameters.