Landslide Change Detection from Satellite Images with Deep Learning Classification

Landslide is one of the most common natural hazards in Taiwan. Because of the complicated terrain, geological, geotechnical and weather conditions in Taiwan, landslides are frequently triggered by earthquakes, typhoons or heavy rainfalls almost year-round, posing significant threats to human lives and property and sometimes causing catastrophic damages. Rapid and accurate detection and classification of landslides are crucial for disaster mitigation, management and prevention. In this regards, satellite remote sensing is an effective approach for collecting data. However, accurate mapping and monitoring landslides usually requires analyzing considerable amounts of images, which is time-consuming and labor-intensive. In addition, in some mountainous regions, landslides may occur repeatedly, and old landslides affected areas may be reclaimed by vegetation, making it difficult to fully understand the spatio-temporal characteristics and changes of landslides. To address these issues, this study adopts a deep learning framework, TransUNet, and develops a two-stage training process and data stacking strategy to detect and classify landslide changes from multi-temporal satellite images of a mountainous watershed region is southern Taiwan. TransUNet combines the strengths of Convolutional Neural Networks (CNNs) and Transformers. Three benchmark datasets (Landslide4Sense, HR-GLDD, and Bijie Dataset) were evaluated in conjunction with labelled image titles extracted from collected SPOT satellite images of the study area for transfer learning. Training of the deep learning model was separated into two stages: the first stage focused on initial landslide change detection, while the second stage refined the classifications by applying a weighting scheme. Results of this study show that TransUNet performs well with high-resolution satellite images for landslide change detection, with the best Precision, Recall and F1-Score of 0.92, 0.76 and 0.82, respectively. In addition, despite lacking a temporal feature extraction framework, developed model can effectively distinguishes the changes of landslide affected areas such as old landslides, new landslides, and vegetation reclaimed areas.