HAF-Net: A Hierarchical Adaptive Fusion Network for Satellite-Based Radar Reflectivity Reconstruction

Abstract

Existing satellite-based radar reflectivity reconstruction methods tend to underestimate strong convective echoes and have difficulty preserving fine spatial structures. This study proposes HAF-Net, a Hierarchical Adaptive Fusion Network for generating satellite-derived proxy fields of composite radar reflectivity (CREF) from FY-4B AGRI observations and auxiliary geospatial variables. HAF-Net is designed to address three challenges in satellite-to-radar reconstruction: the scale mismatch between compact convective cores and extended cloud systems, false alarms associated with irrelevant background responses transmitted through skip connections, and the loss of echo-structure detail when predictions rely only on the final decoder output rather than complementary multi-level representations. Experiments conducted using spatiotemporally matched FY-4B and CINRAD observations show that HAF-Net achieved the best overall performance among the evaluated models across continuous, structural, and threshold-based metrics. Ablation results suggest that Multi-Receptive-Field Convective Feature Extraction block (MCFE) is associated with improved strong-echo representation, while the changes observed after removing Convective-Saliency-Guided Skip Refinement (CSGR) and Hierarchical Echo-Structure Reconstruction Fusion module (HERF) are consistent with their intended roles in false-alarm control and multi-scale structural reconstruction. The reported metrics characterize performance on a radar-covered, day-level held-out test set screened to retain echo-containing scenes within the sampled geographic domain; they should not be interpreted as geographically independent validation or as climatologically representative all-weather performance.