Landslide Extraction And Analysis Based On UAV Images After The Earthquake

The terrain conditions in China are relatively complicated,due to the action of earthquakes,secondary geological disasters such as collapses,landslides,and debris flows are induced frequently.Landslides are the most common and the most destructive secondary geological disasters.A major earthquake often triggers thousands of landslides,landslides will not only bring economic losses,but also cause casualties,so the study of landslides has great significance.In recent years,drones have developed rapidly due to their unique flexibility,portability,durability,and high-resolution data,so they have been widely used in geological disasters and environmental investigations.The method is more operable in the investigation of small areas or watershed landslides and single landslides.Therefore,this paper uses drone images with a resolution of 10 cm after the earthquake of the 2017 Jiuzhaigou Ms7.0 earthquake in Sichuan to extract post-earthquake landslide disaster information.The main research contents are as follows:(1)This paper uses UAV images obtained after the MS7.0 earthquake in Jiuzhaigou County as the data source,and takes part of the landslide areas in Jiudaoguai,Shangsizhai,and Rhino Sea in Jiuzhaigou area as the study area.The selected study area The total area is 87.62 km2,and the actual area of the landslide area included in the study area is 12.67 km2.(2)Three methods of object-oriented,supervised classification and deep learning are used to complete the extraction of landslides.Object-oriented mainly analyzes the attribute characteristics of landslides from the aspects of texture,spectrum,tones,etc.First,multi-scale segmentation of the acquired UAV images,and then Construct a feature rule set,and realize the extraction of landslides according to the constructed feature rule set;the supervised classification first builds a sample library of UAV image seismic landslides,and then selects the minimum distance method as the classifier to achieve the extraction of seismic landslides;The deep learning method uses the area of the landslide disaster body as the training sample area.The size of the sample area is 100×100 pixels.There is no need to use the entire landslide as a single unit,which solves the problem of insufficient landslide data samples.The sample area can be roughly divided into three categories: background label 1,vegetation label 2,and other ground features including sky,road,lake,etc.Landslide label 3.The deep learning convolutional neural network VGG-16 model is used to train,calculate and classify the samples.After the model is successfully calculated,the multi-scale segmentation algorithm is used to perform multi-scale segmentation of the image data to be extracted,and then the constructed VGG-16 model is used to extract the post-earthquake landslide disaster.(3)Among the three extraction methods used in this paper,the object-oriented method is used to extract landslides.The extraction results show "over extraction" and "missing extraction".The average error is 6.99%,and the average kappa coefficient is 76.71%.The distance method is used as a classifier for supervised classification to extract landslides,and the phenomenon of "over-extraction" is common,with an average error of 10.91% and an average kappa coefficient of 67.19%;the method based on deep learning extracts landslides with an average error of 1.35% and an average kappa coefficient It is 94.98%,and the extraction accuracy is high.On the whole,the two methods of object-oriented and supervised classification to extract landslides have improved efficiency,but the extraction accuracy is not high,and the phenomenon of "over-extraction" and "missing extraction" is prone to occur,while the deep learning method extracts landslide Not only solves the efficiency problem,but also improves the extraction accuracy.