FAViT: An Embedding Fusion of Vision Transformer and Attention U-Net for Landslides Segmentation Using Sentinel-1

Abstract

This research introduces fusion attention vision transformer (FAViT), a novel hybrid deep learning framework for the automated rapid mapping of landslides from multitemporal Sentinel-1 synthetic aperture radar (SAR) amplitude data. A significant challenge in this domain is the need for methods that are both computationally efficient on DEM-dependent methods and robust to SAR-specific challenges, such as speckle noise and geometric distortion, without relying on external topographic data, both large regional landslide patterns and small, detailed landslides. To address this, FAViT uniquely integrates a vision transformer through dual-stream feature extraction block to an attention U-Net, merging global contextual modeling via self-attention with precise local feature extraction to overcome the receptive field limitations of conventional convolutional neural networks including U-Net. This framework enables direct landslide segmentation from SAR amplitude data pre- and post-event using dual-polarization combinations. Rigorously evaluated on landslides from the 2016 Mw 7.8 Kaikoura, New Zealand and 2022 Mw 6.8 Luding, China earthquakes, FAViT significantly outperforms attention U-Net (p < 0.01, paired bootstrap). It achieved a 25% relative improvement in F1-score for the Luding test area (0.65±0.01 versus 0.52±0.02 baseline) and a 8% improvement for Kaikoura (0.56±0.01 versus 0.52±0.02 baseline) using VV + VH data, demonstrating its enhanced accuracy. The results confirm that the VV + VH combination provides complementary scattering information for comprehensive landslide characterization, with VV capturing surface geometry and VH detecting vegetation disturbance. Successful spatial hold-out validation within each region underscores the model robustness and generalization capability across two diverse terrains, learn more complex relationships establishing FAViT potential for rapid, large-scale post-disaster assessment and global landslide monitoring.