Deep learning based landslide prone area detection: A comparative evaluation

Landslides are among the most hazardous natural disasters, particularly in meteorologically vulnerable regions, due to heavy rainfall, earthquakes, or human activities like deforestation, thus making it difficult to manage and mitigate. Therefore, accurate and automatic detection of landslide-prone areas is essential for effective meteorological hazard monitoring, offering a scalable and reliable solution. This paper evaluates three advanced deep learning based semantic segmentation models, DeepLabV3, SegNet, and U-Net, to analyze their effectiveness on landslide datasets. Using satellite images, these models perform pixel-wise classification to identify landslide-prone regions across both deserted and grassland regions. The results show that DeepLabV3 outperformed the other models, achieving a 5.97% higher precision, 1.10% higher recall, and a 25% improvement in IoU score compared to U-Net. Similarly, it showed a 2.25% increase in precision, a 3.69% increase in recall, and a 14.75% improvement in IoU score over SegNet. These outcomes highlight the superior capability of DeepLabV3 in accurately detecting landslide-prone areas.