To address the challenge of simultaneously extracting local and global features in multi-task segmentation of Arctic sea ice, a hybrid architecture integrating residual networks and state space models was proposed in this paper. In this architecture, residual networks were employed as the encoder to extract multi-scale local features. Enhanced Visual State Space (EVSS) modules were adopted in the decoder to achieve global context modeling through the combination of 2D Selective Scan and dual attention mechanisms. In the skip connections, Cross-Scale Contextual Fusion (CSCF) modules were designed to integrate multi-level features from the encoder. Attention mechanisms were utilized to enhance the representation of key regions. In the training phase, a Decoder Stage-Wise Supervision (DSWS) was introduced to optimize the calculation of loss functions. Finally, sea ice concentration, development stages, and floe size predictions were generated simultaneously through parallel segmentation heads. Experiments on the AI4Arctic dataset demonstrate that the proposed model achieves a comprehensive score of 84.014. Its performance exceeds that of mainstream methods such as Swin Transformer and SegFormer in all three prediction tasks. Fine-grained segmentation of sea ice distribution is achieved in this paper. The results provide technical support for subsequent identification and delimitation of navigable water areas. The obtained segmentation results hold significant application value in Arctic route planning and ship navigation safety assessment.