SiamNet: A Double-Temporal SAR Avalanche Detection Method Integrating Multiscale Features and an Attention Mechanism

Avalanches pose serious threats to critical infrastructure in alpine-cold regions. Synthetic aperture radar (SAR) provides all-day, all-weather observations for avalanche mapping under persistent cloud cover. However, existing SAR change detection methods report accuracies from 0.49 to 0.87 and remain less effective for small avalanches. Their strong dependence on region-specific parameters also limits rapid and transferable avalanche mapping. To address these limitations, this study proposes a Siamese multiscale feature fusion network (SiamNet) for avalanche change detection from Sentinel-1A SAR images. SiamNet employs a shared-weight Siamese encoder to extract multiscale features from the pre- and post-event images and enhance avalanche-related feature responses. Experiments were conducted in the perennial snow-covered area of the typical alpine-cold area in the south-east Tibetan Plateau. A total of 33 avalanches were identified, with an overall accuracy of 98.76%. Field validation showed high agreement between the mapped areas and the actual avalanche tracks and deposits. Compared with representative change detection methods, including FC-Siam-Diff, SNUNet-CD, BIT, ChangeFormer, HANet, and EfficientCD, SiamNet effectively suppresses false detections while reducing missed detections and maintains superior performance across different scenarios. The results demonstrate that SiamNet can effectively map small avalanches with irregular boundaries and complex internal structures, providing technical support for rapid large-area avalanche surveys, spatial mapping, and risk assessment.