An Efficient U-Net Model for Improved Landslide Detection from Satellite Images

Landslides are a dangerous hazard that might have devastating results. Thus, detecting landslides from satellite images can be significant for various governing authorities. In the past, different deep-learning models have produced remarkable results in terms of landslide detection. Here, an enhanced U-Net model is suggested for detecting the landslides from the newly introduced open-source Bijie landslide data set. The satellite images of the data set are obtained from TripleSat with a spatial resolution of 0.8 m. Further, the proposed study uses the ResNet-50, ResNet-101, VGG-19, and DenseNet-121 as backbone models. The model is evaluated qualitatively, and five matrices, i.e. precision, recall, f1-score, MCC (Matthews-correlation-coefficient), and overall accuracy (OA) are computed for quantitative evaluation. The obtained results of each model are compared with the earlier studies to prove the potential and novelty of the research work. The performance of U-Net + ResNet-50 is found to be the best in terms of precision (0.98), f1-score (0.98), and OA (1.0).