TSDCU-Net: A Deep Learning Model for Sea Ice Classification Based on SAR Images

The accuracy of sea ice extraction is of great significance for ensuring navigation safety and assessing climate change. Synthetic Aperture Radar (SAR) imagery is an essential tool for remote sensing monitoring of sea ice. This research focuses on classifying sea ice and open water, extracting sea ice information from complex ice zones, and overcoming the effect of speckle noises in SAR images. In this paper, a deep learning model called TSDCU-Net is proposed for classifying sea ice and open water in SAR images. Based on the MDOAU-Net backbone, the TSDCU - Net model mainly consists of three distinct parts. First, a multi-texture scale feature fusion module is used at the input to extract features from SAR images. Moreover, dilated convolution blocks are used to increase the receptive field to reduce the effect of speckle noises in SAR images. Furthermore, the four skip connections of the MDOAU-Net backbone are replaced by four channels and spatial attention for improving representational power. Three datasets of GaoFen-3 in Bohai Bay, GaoFen-3 in the Arctic, and Sentinel-1 in the Arctic are validated and evaluated. The experimental results demonstrate that the proposed TSDCU-Net outperforms other models in classifying sea ice and open water, identifying complex marginal ice zone, and overcoming the influence of speckle noises. It is confirmed that the proposed network has a strong robustness ability.