| Landslide disaster has the characteristics of many points,wide areas and fast disaster rate.Its scattered debris can easily cause secondary disasters such as mudslides,which seriously threaten the safety of life and property of local residents.Landslide emergency rescue and vulnerability assessment need to obtain accurate landslide boundary information quickly.However,the traditional landslide extraction and analysis has two deficiencies in landslide data acquisition and landslide interpretation.At the data level,it is difficult for the traditional satellite to penetrate cloudy and rainy areas to acquire landslide images;At the method level,the existing landslide extraction methods based on remote sensing image visual discrimination and field inspection are difficult to realize rapid response to landslide disaster due to low efficiency and high cost.How to extract landslide quickly and accurately and serve landslide risk analysis has become a hot research topic in the world.1)In view of the "difficult point" in obtaining landslide images from satellite images and the low density of effective landslide data,this paper uses high-timeliness,low-altitude and ultra-high resolution unmanned aerial vehicle(UAV)to obtain landslide images,establishes a landslide disaster sample database,and designs a data augment method:2)Propose a multi-scale depth attention landslide extraction method to solve the new problems of large landslide scalar difference and confusion between landslide and bare ground spectral characteristics caused by UAV images due to their ultra-high resolution:3)The landslide disaster extraction system is established,and the training model and extraction methods are encapsulated to realize the practicability of deep learning landslide range information extraction:4)Comprehensively utilize the above-mentioned extracted landslide range information,geological terrain,earthquake and rainfall data,analyze the spatial distribution characteristics of landslide,and propose a model combining logistic regression and weighted information to solve the problem of insufficient data or excessive dependence on knowledge.In order to verify the effectiveness of landslide range extraction and landslide susceptibility assessment methods,this paper takes Lushan(2013)and Jiuzhaigou(2017)earthquake areas in Sichuan province as research areas,and takes post-disaster landslide images obtained by UAV aerial photography as data sources.The experimental results show that: 1)The landslide extraction method proposed in this paper is superior to FCN-16 s,U-Net and ResNet50 network models in IOU,F1 index and time performance;2)The AUC value of the landslide susceptibility assessment model proposed in this paper is 0.95,reaching the optimum,and 85% of the landslides fall into the extremely high susceptibility region,indicating that the evaluation results are relatively reasonable. |
