A deep learning based landslide prediction model using satellite imagery

Abstract

Landslides have long posed significant threats to communities and infrastructure worldwide, prompting the development of effective early warning systems for timely risk mitigation. This project addressed this challenge by leveraging deep learning techniques and satellite imagery to develop a predictive model for landslide prediction. The primary objective was to design and implement a machine learning model capable of accurately identifying areas at risk of landslides, thereby facilitating proactive measures to mitigate potential damage and protect lives and property. In this project, we evaluated the performance of five deep learning models—PSPNet, ResU-Net, U-Net, Segformer, and FastFCN—for the task of landslide detection using a dataset of 3700 images. Each model was trained for 15 epochs on an NVIDIA T4 GPU, with an average training time of 20 minutes per session. Our results indicated that U-Net and FastFCN were the most effective models, achieving the highest accuracy (0.9859 and 0.9886, respectively) and demonstrating a balanced performance in precision and recall. U-Net excelled in precision (0.8056), while FastFCN achieved the highest F1-score (0.7076), highlighting their strengths in correctly identifying landslide regions. The architectural nuances of each model, such as FastFCN's use of dilated convolutions and U-Net's skip connections, played a significant role in their performance variations.