Automated Landslide Scar Detection in the Himalayan Region Using Satellite Imagery and YOLO-Based Deep Learning Models
TL;DR
The results indicate that YOLOv8 achieved the highest accuracy (F-score: 0.857), followed by YOLOv5, YOLOv7, and YOLOv6, which highlight the potential of YOLO-based models for rapid and accurate landslide detection, contributing to efficient hazard assessment and disaster response efforts.
Abstract
Landslides in the Himalayan region are the most serious hazards due to their destructive consequences in the context of life and the economy. More landslides are anticipated based on the present environmental and climatic scenario. Hence, a landslide information detection model is the foremost for hazard risk assessment. The application of deep learning algorithms in hazard analysis has gained significant attention worldwide. Moreover, remote sensing technology has emerged as a critical tool in diverse applications. Therefore, this study leverages the recent scientific and technological developments for extracting landslide-related information. We applied four YOLO (You Only Look Once) models, specifically YOLOv5, YOLOv6, YOLOv7, and YOLOv8, alongside eleven additional YOLO variants to detect landslide scars from satellite imagery. The study focuses on the Greater Himalayan Range, Nepal, utilising a dataset of 275 satellite images, divided into training, validation, and testing sets. The standard quantitative indicators, i.e., precision, recall, F-score, and mean average precision (mAP), are used for evaluation. The results indicate that YOLOv8 achieved the highest accuracy (F-score: 0.857), followed by YOLOv5, YOLOv7, and YOLOv6. These findings highlight the potential of YOLO-based models for rapid and accurate landslide detection, contributing to efficient hazard assessment and disaster response efforts.