Landslide Detection And Volume Estimation Based On Low-Altitude Photogrammetry And Deep Learning

Location and volume information of landslide are important parameters commonly used in landslide geological disaster research.Rapidly and accurately detecting landslide and estimating its volume are of great significance for rescue landslide disaster.Existing methods were difficult to accurately identify landslide boundaries.In this paper,a method of landslide detection based on deep learning for low-altitude camera measurement was mentioned.The method integrated multi-source data features and used the improved Deep Lab V3+ model as the segmentation framework to achieve landslide detection accurately.Due to the interference of complex terrain factors,it was difficult for traditional interpolation algorithms to accurately reconstructed the surface elevation before landslide.In this paper,a deep neural network interpolation model was constructed to generate the elevation before landslide under complex terrain conditions,and the landslide volume was estimated based on the surface elevation changes before and after the landslide with the differential algorithm.The main research work of this paper was as follows:(1)On the basis of the team’s previous research,the correspondence features were extracted from the overlapping images obtained by low-altitude UAV for image matching.The Sf M algorithm was used to perform space-three solution to obtain sparse 3D point cloud.The SGM dense matching algorithm was used to reconstruct dense 3D point cloud and generate digital ortho image and digital elevation model.At the same time,gamma-enhanced visible vegetation index and bilateral filtering algorithm were combined to extract ground points from dense point cloud and eliminate non-ground points(such as vegetation).(2)Considering the characteristics of landslides in multi-scale images,the landslide boundary was difficult to detected accurately by existing methods.An improved model of Deep Lab V3+ with spectral-topographic fusion of landslide(STFDN)was proposed.Landslide topographic feature factors(elevation,slope and aspect)and RGB images are introduced to form multi-source data as training samples of STFDN model.Automatic landslide recognition and boundary detection were brought forward through semantic segmentation method.(3)For the problem that traditional interpolation methods were difficult to accurately simulated topographic elevation before landslide.A deep neural network(DNN)interpolation model was constructed to represent the nonlinear mapping relationship between two-dimension coordinates and elevation in point data.The model taked landslide boundary point data as training samples.Based on the deep learning method,the topside elevation before landslide was generated under complex terrain conditions.Then,the landslide volume was estimated by using the topside elevation variation before and after the landslide with the differential algorithm.Experimental results showed that the introduction of landslide topographic feature factors could improve the result of landslide boundary detection,and verified that the improved model was better than Deep Lab V3+ model in landslide detection accuracy.Compared with the traditional interpolation method,DNN interpolation method could reconstruct accurately the local terrain before landslide and estimated the landslide volume.In this paper,the methods that landslide detection and volume estimation based on deep learning could detected landslide and estimate volume accurately and quickly,which was of great significance in the emergency monitoring and decision-making of landslide disaster.