Global Ice Water Path Retrieval Using Fengyun series Satellite Data: A Machine Learning Approach

Abstract. This study presents a novel machine learning framework (RobustResMLP) for retrieving the global ice water path (IWP) and cloud ice water path (CIWP) from 2009–2024 via passive microwave observations from China's Fengyun-3 series satellites' microwave humidity sounders (MWHS-I/II). The framework employs a lightweight multilayer perceptron architecture enhanced with gated residual units and hierarchical differential dropout to address the challenges associated with high-noise satellite data. By establishing rigorous spatiotemporal collocation with CloudSat 2C-ICE products, we generate three operational products: (1) synoptic type that orbital-resolution IWP/CIWP (15 km; 2009–2024), (2) climatic type that gridded monthly composites (1°×1°; 2011–2024), and (3) cloud layer mask (CLM) products. Notably, the 89 GHz channel emerges is the most influential predictor despite theoretical limitations. This approach achieves a critical compromise between pointwise accuracy and spatiotemporal completeness, enabling unprecedented decadal-scale cloud feedback analyses. All the datasets are open available in the netCDF4 format for community sharing.