A Preliminary Assessment of Landslide Detection in Leyte, Philippines Using an AI-Based U-Net Model on Sentinel-2 Imagery

Abstract. The application of artificial intelligence (AI) based models offers a potential improvement over conventional approaches to landslide detection, which typically demand substantial resources and often require significant human expert involvement. Jadala (2019) describes the U-Net model as a convolutional neural network that is extensively used for accurate pixel-level semantic segmentation despite having only a limited dataset for training. This research explores the capability of the U-Net model for detecting the landslides triggered by Tropical Storm Agaton in Abuyog, Leyte in April 2022, using Sentinel-2 imagery and validated landslide inventories from Leyte, Davao de Oro, and Maguindanao del Norte. The model was trained using image patches that included Sentinel-2 Red, Green, and Blue bands, the Normalized Difference Vegetation Index (NDVI), as well as topographic features such as slope and elevation values acquired from an Interferometric Synthetic Aperture Radar (IFSAR) Digital Terrain Model (DTM). The observed results highlight the model’s effectiveness in identifying landslide pixels, achieving a strong F1-score of 72.97. This performance was further supported by a precision of 79.54 and a recall of 67.46. Across all accuracy metrics, the U-Net model likewise achieved higher performance as opposed to other machine learning approaches such as the Support Vector Machine (SVM) and Random Forest (RF) classifiers, which were evaluated using the same dataset. Future studies should focus on incorporating additional training data across regions with varying geological characteristics to further enhance the model’s accuracy.