Evaluation of U-Net transfer learning model for semantic segmentation of landslides in the Colombian tropical mountain region

Abstract

Landslides in tropical regions, like the Colombian Andean region, pose unique challenges due to factors such as intense rainfall, steep slopes, and complex terrains. Mapping historical and current landslide activity through inventory maps is essential in tropical mountainous regions. While satellite data is commonly used for mapping, it can be time-consuming and manual-intensive, limiting inventory availability. Deep Learning (DL) models, especially Convolutional Neural Networks (CNNs), have shown promise in remote sensing applications with High Resolution (HR) imagery, including landslide detection. Despite advancements, their use in this field is still relatively limited. This study assesses the effectiveness of U-Net model, for automated landslide detection using spectral data from optical satellite imagery (RGB bands), two DEM-derived geo-indices (slope and curvature), and two Synthetic Aperture Radar (SAR) layers (VV amplitude pre- and post-landslide event in May 2015) across three image models (3, 5, and 7 bands). Initially, data is combined into multiband images, and the model is trained in the “La Argelia” river basin in Colombia’s Pacific region. Subsequently, the model is tested in the “La Liboriana” river basin in the western Andean region. The landslide detection results within the inference area are validated by comparing them with the landslide inventory and segmentation results. The U-Net model demonstrates good performance (F1-score around 0.70) for landslide detection, as confirmed in various geographical settings. By utilizing DL models and combining high-resolution satellite imagery, topographical, and SAR data, a comprehensive space-time mapping of landslides can be achieved. This approach has the potential to greatly improve the accuracy and effectiveness of landslide mapping, offering a more holistic view of the temporal dynamics related to these natural hazards.