Research On Machine Vision Identification And Monitoring Of Landslide In Southeast Tibet

Landslides are widely distributed in China,occur frequently,and cause heavy losses.Recently,several strategic projects have been undertaken in southeast Tibet,including the Sichuan–Tibet railway,hydropower development in high-altitude areas,and the China–Pakistan Economic Corridor.Landslides pose several challenges to disaster prevention and mitigation.This study,therefore,focuses on addressing four problems—“late discovery,” “slow response,” “difficulty monitoring,” and “inaccurate predictions,”—associated with the Tibetan southeastern regional landslide disaster using a machine-vision-based technology.Using regional landslide disaster identification and monitoring,this study explored landslide disaster prevention and control in southeast Tibet.The main contents of this paper are as follows:(1)A high-precision recognizer is developed to detect landslide images in southeast Tibet,which can detect landslides automatically.Original landslide images are insufficient and blurred;image classification is inaccurate in southeast Tibet.To address these problems,the landslide image database of a section from Lulang to the Zamu town of the Sichuan–Tibet highway in southeast Tibet is first established,and then divided into the original and enhanced datasets.A Caffe Net for Landslide Recognition(Caffe Net-LR)model for landslide identification is proposed based on deep learning.Moreover,the model precision is compared with two-scale datasets and different transfer learning schemes.Moreover,the optimized model is selected and combined with the sliding-window scanning strategy to design a landslide image recognizer in southeast Tibet.The recognizer can realize intelligent landslide image recognition and image positioning in unattended regions.(2)An optimized layout scheme of 3D refined modeling ground-control points for high and steep mountain slopes is proposed.Given the difficulty in surveying ground-control points in the high mountain valley area of southeast Tibet,different layout schemes of ground-control points are designed,and conventional and emergency monitoring conditions are divided according to the layout areas.In addition,inspired on point reference,surface reference,and multisource data fusion,a model accuracy evaluation method is proposed,which can solve the problem of the difficulty associated with the evaluation of the model under the condition of limited control points and checkpoint data acquisition.Considering the potential landslide area of Lamu Village,Dazi District,Lhasa as the research object,the optimized ground-control point layout scheme is obtained by comparing the modeling accuracies of different schemes.Based on this method,a high-precision 3D model of the potential landslide area of Tianmo–Gully in Bomi is obtained using the control point layout scheme under conventional monitoring conditions.(3)The calculation formula for single and binocular vision measurement is derived.Further,the single and binocular vision surface displacement monitoring systems are designed,to solve the issues of limited installation points of visual measuring equipment in field-landslide monitoring scene and complex images in a natural scene.Additionally,the influence of the planar target design scheme on the calibration accuracy is explored.According to the different principles and application scenarios of single and binocular vision measurement,the corresponding cooperative targets are offered to improve the measurement accuracy,and an improved Harris corner extraction algorithm is proposed to achieve subpixel detection of the target corner.Furthermore,the single and binocular vision surface displacement monitoring systems are presented and applied in laboratory experiments.A digital image processing program is developed for the application scene,which can realize automatic 3D displacement measurement,using the relationship between the change of cooperative target image and relative displacement.(4)An online monitoring system for southeast Tibet landslide is designed based on the machine vision measurement technology of slope surface displacement.Combined with above research,a field machine vision–GNSS(Global Navigation Satellite System,GNSS)-integrated observation is designed and implemented.It can deal with the difficulties of special monitoring networks,power supply,and hydrometeorological environments in the landslide area of southeast Tibet.Moreover,based on the overall design and subsystem design,a real-time multisensor online landslide monitoring system is established,which integrates machine vision measurement,GNSS,low-power Internet of Things data storage,and transmission technology.The real-time online monitoring is utilized in the Tianmo/Pailong–Gully geological disaster fields.(5)The landslide displacement prediction model is constructed based on gray system processing and RBF(Radial Basis Function,RBF).First,aiming at the limitation of slope displacement prediction model based on a single algorithm,mixed models of slope displacement prediction based on the RBF neural network is developed.Then,the large range and efficient search of the RBF network diffusion function are realized using the particle swarm optimization algorithm.Further,the slope displacement prediction model based on gray system processing and RBF as well as the mixed model considering induced factors are constructed to improve the prediction accuracy of the data-driven model for an incremental change point of displacement.