Dynamic Contribution-Matrix ResNet-Based Retrieval Algorithm for Ocean Surface High Wind Speed From Spaceborne Microwave Radiometer

Deep neural network (DNN), equipped with powerful distinctive generalization ability, is one of the most popular retrieval tools for processing ocean remote sensing data from spaceborne microwave radiometers. Considering the signal received by microwave radiometers at different frequencies changes with atmospheric effects, such as rainfall and water vapor in both horizontal (H) and vertical (V) polarization (pol.), and differential signals between H and V pol, these physical factors cannot be ignored in high wind speed retrieval. This study focuses on the integration of physically driven statistical functions with a residual neural network (ResNet) to derive a specialized type of DNN framework for high wind speed retrieval. Specifically, a dynamic contribution-matrix ResNet (DCM-ResNet) model is proposed, which dynamically adjusts the contribution (C) matrix coefficients within the model. These coefficients, reflecting the physical information of radiation transfer, are used to automatically regulate the model’s estimation of high wind speeds. Experiments are conducted using brightness temperatures (TBs) from the Advanced Microwave Scanning Radiometer-2 and wind speeds from the Stepped-Frequency Microwave Radiometer. The experimental results show that the maximum wind speed can reach up to 60 m/s with a root mean square error (RMSE) of 2.92 m/s. In addition, the wind speed RMSE remains stable within 6 m/s as the rainfall rate increases. The contributions of different frequency bands to the C-matrix results are closely related to radiative transfer.