Adaptive Prototype‐Based Few‐Shot Segmentation for Debris Flow Early Warning

ABSTRACT

Debris flows are highly destructive geohazards requiring rapid, accurate perception for effective early warning and emergency response. However, deep learning methods depend heavily on large datasets, whereas debris flow imagery is inherently scarce. Although few‐shot semantic segmentation (FSS) addresses data limitations, existing methods suffer from semantic ambiguity and weak adaptability to complex target morphologies. To address these challenges, this study proposes a cross‐modal few‐shot segmentation framework and constructs a visible‐light benchmark dataset named DebrisFlow‐VisSeg (DFVS). The framework incorporates a Cross‐Modal Feature Enhancement Unit (CFEU) to improve visual‐semantic discrimination, an Adaptive Prototype Construction Mechanism (APCM) to generate robust prototypes for heterogeneous structures, and a Query Feature Disentanglement and Alignment Module (QF‐DAM) to suppress background interference. Extensive experiments on DFVS, PASCAL‐5 ⁱ , and COCO‐20 ⁱ demonstrate that the method outperforms state‐of‐the‐art approaches. Specifically, it achieves 76.31% mIoU under the 1‐shot setting and 80.97% mIoU under the 5‐shot setting on DFVS. Field deployment further verifies its practical applicability for real‐world debris flow monitoring under data‐scarce conditions.