AVANET: A Multiscale Weak-Texture-Aware Network for Snow Avalanche Recognition Using Aerial Imagery

Accurate recognition of snow avalanches from aerial imagery is crucial for cryosphere hazard monitoring and assessment. However, avalanche traces in snow-covered mountain scenes often exhibit weak texture, low contrast, blurred boundaries and large-scale variation, which degrade the reliability of existing visual recognition models. This paper proposes AVANET, a multiscale weak-texture-aware network for aerial snow-avalanche recognition. Specifically, AVANET combines a Multi-branch Weak-Texture Response Fusion (MWTRF) mechanism and a lightweight Multiscale Contextual Enhancement (MCE) module. The MWTRF mechanism is introduced during training to strengthen subtle discriminative cues using high-frequency residual, Sobel gradient, and multi-orientation Gabor responses. The MCE module aggregates cross-scale contextual information while preserving fine-grained texture details. Smooth Gradient-weighted Class Activation Mapping++ (SmoothGrad-CAM++) is further adopted to visualize model attention and support interpretability analysis. Experimental results show that AVANET achieves 86.13% overall accuracy (OA) and 90.67% F1-score, outperforming the YOLOv11s baseline by 3.06% in OA and 2.21% in F1. Interpretability analysis indicates that the model focuses on plausible avalanche-related regions, including release zones, track areas, and debris deposits. These results demonstrate the effectiveness of AVANET for avalanche recognition in aerial remote sensing imagery.