Landslide Segmentation from Multispectral Remote Sensing Data Using U-Net with Multichannel Topographic Information

Abstract

Landslides lie between high frequency normal hazards that cause moderate to high damage to infrastructure, ecosystems and human settlements, thus requiring precise and automated landslide detection and delineation. Conventional landslide mapping approaches are often dependent on manual interpretation, field surveys, and limited geospatial information, which can be time-consuming and problematic to scale over huge areas. This research presents a deep learning-based semantic segmentation framework using the U-Net architecture for pixel-level landslide delineation from multispectral and topographical remote sensing data. The projected method employs the Landslide4Sense dataset with 4,844 image patches of spatial size 128 × 128 pixels and 14 input channels including 12 Sentinel-2 multispectral bands, slope and Digital Elevation Model (DEM) information. The U-Net model is an encoder-decoder based feature extractor with skip connections to preserve spatial information and generate pixel level landslide masks. The model is trained and tested by the performance indicators of the Intersection over Union (IoU), F1-score, precision, recall and accuracy. The experimental results demonstrate that the proposed U-Net model provides 0.4300 IoU, 0.6014 F1-score, 0.5051 precision, 0.7431 recall and 0.9814 accuracy. The results show the efficiency of the U-Net in pixel detection for landslides, especially in terms of recall, while the lower precision indicates the existence of false-positive predictions. The developed framework provides a deep learning-based baseline for automated landslide segmentation using integrated multispectral and topographic information and establishes a foundation for further investigation of advanced Transformer- and attention-based architectures.