<p>The satellite synthetic aperture radar (SAR) sensor is one of the most critical tools for monitoring Arctic sea ice. Classifying sea ice types based on SAR images has been a research hotspot. Most existing deep-learning-based sea ice classification models rely on the polarimetric information of SAR images while ignoring the gray-level co-occurrence matrix (GLCM) feature. This study develops a three-branch U-Net model for classifying sea ice in SAR images. By integrating polarimetric information, GLCM features, and auxiliary data, the model can classify open water (OW), young ice (YIC), first-year ice (FYI), and old ice (OIC). The model is trained and tested on the well-known AI4Arctic sea ice challenge dataset. Experiments on 57 testing SAR images demonstrate that the proposed model achieves an overall classification accuracy of 91.45% and an Intersection over Union (IoU) of 0.846 4 for the four-type classification. Ablation experiments were conducted to evaluate the sensitivity of various GLCM features to sea ice classification. The effectiveness of the three-branch input for fusing polarimetric information, GLCM feature, and auxiliary data is validated. Results indicate that incorporating HV_mean significantly enhances classification performance, with an accuracy increase of approximately 0.7% and an improvement in IoU of 0.9%. The three-branch input structure is more effective than the single-branch structure in fusing three types of inputs, resulting in an accuracy increase of 4.7% and an improvement in IoU of 7%. Therefore, the proposed three-branch U-Net model demonstrates stable and reliable capabilities for classifying OW, YIC, FYI, and OIC in SAR images, providing a new approach for Arctic sea ice monitoring.</p>