PERM‐Net: A Learning Framework for Rainfall Intensity Estimation From Doppler Radar Sensor Observations

ABSTRACT

Rainfall estimation plays a pivotal role in meteorological early warning systems, urban flood mitigation and hydrological scheduling. To tackle the challenges of heavy rainfall sample scarcity, long‐tailed distribution imbalance and noise sensitivity in Doppler radar sensor (DRS) spectral‐moment block‐level features, a physics‐guided evidence routing mixture network (PERM‐Net) framework for adaptive rainfall estimation is proposed. By integrating evidence‐driven dynamic routing (EDR), heterogeneous expert modelling (HEM) and fusion with uncertainty estimation (FUE) into a unified framework, PERM‐Net enhances feature representation capability and improves the robustness and reliability of rainfall estimation under complex precipitation scenarios. Experimental results demonstrate that the proposed framework consistently outperforms some learning methods in terms of overall estimation accuracy and adaptability to complex raining condition. These superior performance further validates the effectiveness and generalization capability of PERM‐Net for DRS rainfall estimation.