Landslide Hazard Identification Based on Deep Learning and Sentinel-2 Remote Sensing Imagery

Abstract

Abstract Landslide is one of the common geological disasters, which seriously threatens human life and property safety. It is particularly important to quickly identify landslide information. This paper takes the Wenchuan earthquake landslide area as the research area, and uses 7 deep learning methods(4-Layer-CNN, AlexNet, ResNet152V2, DenseNet201, InceptionV3, Xception and InceptionResNetV2) to discuss landslide detection methods based on Sentinel-2 remote sensing images. Using the marked landslide and non-landslide sample points, the Sentinel-2 remote sensing image was sliced into 80×80 pixel tiles, and then the deep learning method was used for model training, verification and testing. The results show that : (1) Among the 7 deep learning models, the F1-Score of the DenseNet201 model is the largest, reaching 0.8872, and the RMSE is the smallest 0.2503. It can be seen that the DenseNet model has a good recognition effect on landslide samples, with an accuracy of 0.9172; (2) Second It is InceptionResNetV2, the F1-Score is 0.8721, the RMSE is 0.2721, and the landslide sample recognition accuracy is 0.9012; (3) the worst effect is AlexNet, the minimum F1-Score is only 0.7263, the maximum RMSE is 0.4022, and the accuracy is 0.8295. It can be seen that the deep learning method is applied to Sentinel-2 remote sensing images for landslide image detection, and the accuracy can reach 91.72%, which can quickly and accurately identify landslide information, and improve the method reference and decision basis for disaster prevention and mitigation.