A Physics-Constrained Deep Learning Framework With GAN-Based Data Augmentation for FY-3G MWRI-RM Precipitation Retrieval

Abstract

Precipitation is a key component of the global water cycle, and spaceborne passive microwave remote sensing is one of the main means for global precipitation monitoring. This article proposes a semiempirical and semiphysical precipitation retrieval framework for FengYun-3G (FY-3G) Microwave Radiance Imager-Rainfall Measurement (MWRI-RM) that integrates stepwise physical mechanisms with the advantages of deep learning. First, observed data and precipitation samples generated by generative adversarial networks (GANs) are used to correct and augment numerical weather prediction (NWP) forecast fields, thereby constructing a “hydrometeor–brightness temperature” dataset that is physically distribution-consistent and class-balanced. Second, during retrieval, brightness temperature channel and polarization differences as well as precipitation type are introduced as physical constraints. A back-propagation (BP) neural network and an AttentiveRainNet model are developed to independently retrieve three hydrometeor parameters (cloud water, rain water, and graupel), explicitly resolving the “cold cloud–warm cloud” structure of precipitation systems. The surface rain rate is finally obtained through fusion calculation. The retrieval results are validated using FY-3G MWRI-RM Level 2 (L2) products and Integrated Multi-satellite Retrievals for GPM (IMERG) global precipitation products. Results show that the retrieved rain rates over ocean and land achieve strong consistency with the L2 product (correlation coefficients of 0.911 over ocean and 0.784 over land). Compared with IMERG, the correlation coefficient and root-mean-square error (RMSE) are 0.681 and 1.909 mm/h over the ocean and 0.713 and 0.763 mm/h over land. Notably, compared with the official L2 products, the proposed framework reduces the RMSE by 16.1% over the ocean and 27.8% over land. Furthermore, for Liaoning rainstorm and two typhoon cases in 2024, the model achieves excellent retrieval performance with correlation coefficients of 0.816, 0.866, and 0.771, respectively.