A deep learning recognition model for landslide terrain based on multi-source data fusion

The traditional high-level remote landslide recognition efficiency which relies on the artificial discrimination of geological experts is low. In this paper, an automatic landslide terrain recognition model based on deep learning is developed to improve the efficiency of the screening of potential landslide hazard in a large area. The model takes remote sensing images, DEM data, geological zones, river system and other geological observation data of the target area as input. For the huge difference of different types of observation data, a feature branch network is designed and constructed to accurately extract the corresponding landslide features: Among them, deep network architecture is used to extract complex features from optical image data, and shallow network architecture is used to extract features from structured data such as altitude, geological composition, river and fault zone distribution. Subsequently, a feature fusion module was designed to fuse the extraction results of the two networks to obtain a comprehensive landslide hazard feature. The model performs semantic segmentation of the landslide area based on the extracted landslide features, and achieves accurate pixel-level landslide terrain classification and positioning. The experimental results show that the recognition accuracy(ACC) of the model reaches 0.85, which can provide technical support for automatic landslide identification.