A Dual-Branch Architecture for Adaptive Loss Multitask Mapping Based on AI4Arctic Sea Ice Challenge Dataset

Automated sea ice mapping is increasingly critical for global climate change research and Arctic shipping route planning. In this article, a dual-branch deep learning architecture with channel attention is proposed for multisource sea ice mapping, incorporating an adaptive multitask loss function weighting mechanism. The Ready-To-Train (RTT) AI4Arctic Sea Ice Challenge dataset is used to evaluate the performance of the proposed model. Experimental results demonstrate that compared with the current state-of-the-art model, the proposed model achieves a 1.03% improvement in the combined score. Specifically, the stage of development (SOD) $F1$ score increases by 1.33%, the floe size (FLOE) $F1$ score improves by 4.39%, whereas $R^{2}$ for sea ice concentration (SIC) decreases by 1.35%. Finally, ablation experiments are conducted to validate the effectiveness of the proposed model and the adaptive multitask loss function.