Monitoring Landslide Progression in Leyte, Philippines using Sentinel-2 Imagery and AI-Based Semantic Segmentation Bernadette Anne B. Recto (a*), Raymond Freth A. Lagria (a), Jude Vito C. Agapito (a), Likha G. Minimo (b,c)
a) Department of Industrial Engineering and Operations Research, University of the Philippines Diliman, Quezon City, Philippines
*bbrecto[at]up.edu.ph
b) Science and Society Program, University of the Philippines Diliman, Quezon City, Philippines
c) University of the Philippines Resilience Institute, Quezon City, Philippines
Abstract
Mountainous regions are often devastated by landslides especially following triggering events such as intense rainfall, earthquakes, and volcanic activity. While numerous studies have applied artificial intelligence (AI) for detecting landslides immediately after such events, few have focused on monitoring their spatial and temporal progression over time. This study highlights the potential of AI-based semantic segmentation to monitor landslide progression using multitemporal Sentinel-2 imagery, following the impact of Tropical Storm Agaton in the province of Leyte, Philippines, in April 2022. Sentinel-2 Level 2A images captured immediately after the event, as well as one month, three months, six months, and one year later were acquired and clipped to the municipality of Abuyog in Leyte. From each image, the Red, Green, Blue, and Near-Infrared (NIR) bands, along with the computed Normalized Difference Vegetation Index (NDVI) were extracted and stacked with the elevation values and slope derived from an Interferometric Synthetic Aperture Radar (IFSAR) Digital Terrain Model (DTM) to create multiband inputs for analysis. A U-Net model, trained on labeled landslide polygons which were validated by experts, was then utilized to detect the extent and progression of landslide-affected areas across sequential satellite images captured over time. The model demonstrated consistent segmentation performance across all dates, with F1-scores ranging from 0.684 to 0.821. The results show subtle spatial progression of landslides in certain areas in images taken immediately after the typhoon and those captured one month later, likely due to factors such as prolonged rainfall and terrain instability. In contrast, early signs of vegetation recovery become apparent in some regions between six months to one year after the event. This study provides a starting point for further research on post-disaster recovery monitoring and the identification of areas at risk of secondary landslides, offering practical value to local government units in planning and decision-making.
