A Low-Cost and Scalable Landslide Monitoring and Early Warning System for Mountainous Regions Using Deep Learning-based Computer Vision

Landslides pose severe risks to lives, property, and infrastructure, particularly in mountainous regions where monitoring is hindered by harsh environmental conditions and complex terrains. This study proposes a scalable, low-cost landslide monitoring and early warning system that integrates deep learning and computer vision technologies. Using a single optical camera, time-lapse images of slopes were captured and processed through advanced alignment, masking, and histogram equalization to maintain accuracy under adverse conditions. The system combines static monitoring via image differencing with dynamic monitoring using a RAFT-based optical flow model to detect surface deformations and real-time displacements. Field tests conducted on Washington Makahdiot Cliff demonstrated the system’s capability to identify landslide risks across three deformation phases. Static analysis detected significant grayscale changes (> 100) linked to actual slope deformations, while dynamic monitoring achieved a displacement detection accuracy of 94.6% for movements up to 25 cm. A novel early warning algorithm, based on thresholds for abnormal pixel changes, successfully classified risk levels, with landslide risk coefficients exceeding 0.4 during landslide events and remaining below 0.1 otherwise. The system proved resilient to environmental challenges, such as fog and strong winds, and demonstrated a positive correlation between landslide risk coefficients and rainfall, underscoring the importance of integrating real-time environmental data. This method provides an effective, robust, and cost-efficient solution for landslide monitoring and mitigation in complex terrains.