Region-wise query-guided adaptive multimodal fusion network for fine-grained arctic sea ice recognition

Although multimodal remote sensing data fusion has achieved significant progress in Arctic sea ice recognition tasks, challenges remain in achieving effective cross-modal feature alignment and adaptive fusion. Specifically, 1) Differences in imaging mechanisms across modalities can easily lead to the mixing of foreground features from fusion with non-target background semantic information, and 2) existing fusion strategies largely rely on static weighting schemes, which makes it difficult to dynamically adjust modality contributions according to the task context. To address these issues, we propose a region-wise query-guided adaptive multimodal fusion network, termed RQAFNet. The network employs modality-specific experts (MSEs) to independently model feature distributions within each modality. It integrates region-wise query-guided generation (RWQG2) and spatial reconstruction (SR) mechanisms to achieve adaptive alignment of cross-modal semantic and structural information at the regional scale. Simultaneously, it couples the feature extraction pipeline with a multi-task decoder to enable collaborative modeling of multimodal features, ranging from global to local levels. Experimental results on the AI4Arctic dataset demonstrate that RQAFNet significantly outperforms existing state-of-the-art methods in cross-scene sea ice recognition, achieving optimal performance across all evaluation metrics. Furthermore, seasonal analysis shows that the proposed feature-level fusion approach effectively suppresses seasonal error fluctuations and mitigates systematic bias.