Comparison of landslide detection based on different deep learning algorithms

Abstract

Complex terrain and landscape, and mountains in China are prone to landslide disasters, and causing great potential danger to social stability, people's life and property safety, which makes landslide detection be a hot spot of research. With the advent of remote sensing images, landslide data are growing explosively, which provides conditions for landslide detection by applying deep learning algorithms. Firstly, to broaden the landslide identification method, we created a landslide dataset based on open-source Google Earth images. Secondly, we applied a series of object detection algorithms, such as YOLOV5, Faster RCNN, EfficientDet, SSD, and the improved YOLOV5 by embedding CBAM and Ghost module in the backbone, for detecting landslides from landslide dataset. Finally, we analyzed and compared the experimental results. The results show that the SSD algorithm detects landslides with an accuracy of 97.86%, and the training time for each epoch is only 57s. When there is only one landslide target in the image, SSD is advantageous; The improved YOLOV5 in this paper can identify multiple landslides well while reducing the number of model parameters when detecting images containing multiple landslides.