Loess Landslide Recognition Based On Remote Sensing Image And DEM

Landslide automatic recognition has always been a heated topic in geological disaster filed.Currently,related researches mostly focus on establishing segmentation strategy and landslide recognition model,but ignore the importance of researches on landslide characteristic indexes.Once the landslide disaster occurs,it will not only cause damage to the original surface coverage,but also change its original terrain features,which results in marked differences between landside and adjacent non-landslide that proved to be more stable than differences in surface coverage.The differences caused by land cover damage can be reflected in remote sensing images according to the continuity of color and specific remote sensing indexes,meanwhile,the terrain differences can be obtained by calculating the terrain feature indexes.At present,the terrain features of landslides,which have been introduced to recognizing landslide by more and more scholars,are more and more prominent in the automatic recognition of landslides.However,most of them still remain in macro cognition of the terrain features of the landslide,and have not carried out further quantitative study yet.Tianshui,located in the transition zone of Qinling Mountain and the Loess Plateau,presents poor geological environment.Besides,loess landslide disasters on the North Mountain and South Mountain in Tianshui have arisen with a large number and in a large scale,which will make a huge impact on the people' life and social production.From the perspective of geomorphometry,this paper studies the existing loess landslides on the North Mountain and South Mountain in Tianshui in terms of automatic recognition.Firstly,visuality is introduced to help with visual interpretation of remote sensing images for the selection of landslide and non-landslide samples.Then,combining terrain features with remote sensing features,the paper adopts SEaTH algorithm to select feature indexes that can effectively distinguish between landslide and non-landslide,and then establishes an effective recognition index system.Further,based on the effective identification indexes,the feature set is set to performance the object-oriented image segmentation,and the multi-layer perceptron is taken as automatic recognition model for landslide to recognize the segmented objects.Finally,the accuracyof recognition results is evaluated.The research can not only effectively assist in disaster prevention and mitigation work,but also provide as reference for the subsequent application of quantitative geomorphology in geological disaster recognition.The results show that:(1)It can effectively improve the accuracy of visual interpretation for landslide by calculating the terrain visibility index and combining it with the remote sensing image interpretation method.(2)For the existing loess landslides,six terrain indexes which are roughness,height variation,dissection degree,heave rolling degree,slope and visuality provide with interpretation value,while some other indexes which are height,height difference,surface curvature,profile curvature,NDVI index,SAVI index,NDCI index and three indexes derived from the tasseled cap including of brightness,greenness,humidity do not present recognition value.Compared with the remote sensing characteristics of landslide,topographic characteristics of landslide is the key to recognize the landslide.(3)Based on the feature set for the landslide recognition,a multilayer perceptron neural network landslide recognition model based on object-oriented supervised classification method is proposed.The landslide recognition accuracy and non-landslide recognition accuracy of the model reached 71.03% and 92.02% respectively,which indicates the model's better performance in the recognition of the landslide scope.