Wind-Concerned Sea Ice Detection and Concentration Retrieval From GNSS-R Data Using a Modified Convolutional Neural Network

Abstract

Spaceborne global navigation satellite system reflectometry method has been increasingly utilized for sea ice parameters retrieval. The coupling effects between wind speed and retrieval parameters was not considered in both traditional threshold-based methods and neural network models. To address this, a wind-concerned convolutional neural network (WCNN) model for sea ice detection and concentration retrieval is proposed. The model is based on convolutional layers for the feature extraction from delay-Doppler maps, along with fully connected layers for fusing the flattened feature map and wind speed parameters. After data training and testing, the WCNN model achieved sea ice concentration (SIC) retrieval accuracies with RMSE values of 8.61% in the Antarctic and 11.70% in the Arctic, with correlation coefficients of 0.97 in both regions. The sea ice detection accuracy reached 98.19% and 97.08%, respectively. In summary, the comparison of WCNN SIC retrieval performance across varying wind speeds demonstrates that incorporating wind speed data into the WCNN model significantly reduces the misclassification of seawater as sea ice in low wind conditions (0–10 m/s) and lowers the misclassification of sea ice as seawater in high wind conditions (10–20 m/s). Furthermore, the spatiotemporal distribution characteristics of the retrieval results, the advantages and weaknesses of the model are discussed.