Improving FY-4B Satellite Precipitation Retrieval over Coastal Complex Terrain of Eastern China: Deep Learning Approaches with Multi-Source Underlying Surface Data

This study proposes a deep learning-based precipitation retrieval framework to improve precipitation retrieval using Fengyun-4B (FY-4B) multi-channel infrared brightness temperatures and cloud-top temperature (CTT). Four deep learning models are evaluated, ranging from lightweight architectures (DS-UNet, SmaAt-UNet) to more computationally intensive ones (U-Net, Attention U-Net). Unlike previous FY-4B/deep learning precipitation retrieval studies that mainly emphasize satellite-derived cloud-top features, this study explicitly evaluates the contribution of underlying-surface information within a unified retrieval framework. To quantitatively assess the impacts of underlying-surface information on retrieval performance, a digital elevation model (DEM), topographic factors (slope and surface roughness), and land use/land cover (LULC) are integrated into the framework. The impacts of surface information on precipitation retrieval were further examined through validation against the Global Precipitation Measurement (GPM) mission’s Integrated Multi-satellitE Retrievals for GPM (IMERG) product and rain-gauge observations. Results show that precipitation detection can be improved by incorporating underlying-surface information. Model architecture strongly influenced the results. Lightweight models (DS-UNet and SmaAt-UNet) significantly enhanced retrieval performance by providing clearer spatial constraints because of their greater sensitivities to discrete LULC boundaries. Deeper models (U-Net and Attention U-Net) more effectively encoded continuous topographic gradients because they benefited more from terrain-related variables. Precipitation retrieval skill may not necessarily be improved by a simple combination of DEM and LULC, probably due to thermodynamic forcing and spatial scale mismatches. By comparison, incorporating multiple factors can effectively alleviate these conflicts and improve precipitation detection. This study highlights that precipitation monitoring over complex regions can be enhanced through the proper integration of underlying-surface information. Considering both network architecture and operational objectives is therefore important for developing optimal integration strategies, which can support regional hydrological modeling and flood early-warning systems by providing more reliable input datasets.