Cross-Modal Dynamic Feature Fusion for Visible-Infrared Debris Flow Segmentation

Gully type debris flows are sudden, highly destructive geological hazards requiring accurate, real-time monitoring for effective early warning. However, single-modal visual monitoring is sensitive to complex environments, while existing multi-modal fusion approaches rely on static strategies, limiting adaptability and modal complementarity. Blurred boundary segmentation, class imbalance, and real-time deployment challenges also remain unaddressed. To overcome these issues, this study proposes a cross-modal dynamic feature fusion framework integrating visible and infrared imagery, consisting of a shared backbone for multi-scale feature extraction, a dynamic feature aggregation module for adaptive heterogeneous fusion, a lightweight context-aware semantic segmentation network, and a composite loss function to enhance boundary delineation and mitigate class imbalance. Validated on a self-constructed dual-modal debris flow dataset and public benchmarks, the method achieves an mIoU of 75.6%, outperforming state-of-the-art methods by 3.1%. It meets real-time monitoring requirements and exhibits strong generalization, providing a practical solution for debris flow monitoring with great potential for disaster early warning deployment.