Semantic segmentation of landslide images in Nyingchi region based on PSPNet network
Zhun Li, Yonggang Guo
Abstract
Tibet is located in southwestern China, with many mountains and complex geological conditions, which are prone to natural geological disasters such as landslides and debris flows. How to extract landslide information efficiently and accurately in real time is of great significance to the rapid response after disasters. The application of deep learning semantic segmentation algorithms to information extraction of geological disasters is a new exploration in recent years. An improved PSPNet network model was proposed for semantic segmentation of landslide in Nyingchi, Tibet Province. Considering the segmentation performance and computing speed, the main feature extraction network in front of PSPNet adopts MobileNetV2 structure, which reduced the number of network parameters, reduced the computational complexity of the network, and effectively improved the convergence speed of the network. Based on the landslide images collected in Nyingchi prefecture, Tibet province, the improved model was compared with the classical SegNet, Unet and the original PSPNet algorithm. The improved algorithm effectively reduces the misclassification and divides objects more accurately.