Swin Transformer-Based Multiscale Attention Model for Landslide Extraction From Large-Scale Area

Landslides, a frequent and devastating natural disaster, pose significant risks to human populations and the environment, often leading to substantial loss of life, property damage, and ecological disruption. Creating a comprehensive landslide inventory is essential for disaster response planning and understanding landslide mechanisms. Current methods for landslide extraction, often designed for specific events and primarily focused on vegetative backgrounds, face challenges in practical applications. Landslide extraction from large-scale areas encounters two primary challenges: data imbalance between landslides and background objects and the confusing features of small-scale landslides with complex background objects. This article introduces a two-phase framework for extracting multiscale landslides across large areas with intricate backgrounds. Initially, a dual-temporal image-based method is employed to identify candidate landslides, effectively reducing background interference and addressing data imbalance. Subsequently, a Swin Transformer-based multiscale attention model (Swin-MA) is proposed to capture and learn multiscale landslide features comprehensively. We conducted our study in two regions: the Hengduan Mountains in China, a hotspot for frequent landslides, and Hokkaido, Japan, where significant landslides occurred following an earthquake on September 6, 2018. Our approach outperforms seven recently proposed methods, demonstrating at least a 5.36% improvement in intersection over union (IoU) and affirming its effectiveness and significance in large-scale landslide extraction.