Research On Automatic Multi-Scale Early Identification Of Landslide Based On Multi-source Remote Sensing Data

The frequent occurrence of geological disasters poses a great threat to the safety of people ’s lives and property,and restricts the national economic development and ecological civilization construction.According to statistics,thousands of geological disasters occur every year in China,causing immeasurable losses.Among them,landslides are the most frequent geological disasters and cause the greatest losses.Therefore,the research on early identification of landslide is of great significance for landslide early warning,disaster prevention and mitigation.The geological environment of Zhouqu County in Gansu Province is complex and the landslide disaster is developed.Therefore,this paper takes Zhouqu County as the research area to study the early identification of landslides.In this paper,the GF-1 images in 2015 and 2021,the Landsat8 images in 2021 are used as data sources.Radiation correction,orthorectification,image fusion,cropping,registration and other preprocessing are carried out in ENVI and Arc GIS software.Then,the object-oriented classification method in e Cognition software and the change detection method in ENVI software are used to automatically extract landslides in Chengguan Town,Dachuan Town,Dongshan Town,Nanyu Town and Jiangpan Town.Finally,the landslide susceptibility evaluation model based on information content method is constructed to classify the susceptibility of Zhouqu County.The main research results are as follows :(1)Landslide information extraction based on object-oriented.The multi-scale segmentation method is selected,and the ESP tool is used to determine that the optimal segmentation scale is 131,the shape factor weight is 0.2,and the compactness weight is 0.5.Secondly,the mean,aspect ratio,NDVI,elevation,slope,roughness and GLCM image features and fuzzy classification methods are selected to divide the ground objects into shadows,vegetation,water bodies,buildings,roads,landslides and other ground objects.Combined with 457 landslide sample points in the study area,358 were correctly interpreted.Finally,the accuracy of the classification results is evaluated based on the sample confusion matrix.The overall accuracy is 78.65 % and the Kappa coefficient is 0.68.(2)Extracting landslide information based on change detection.The vegetation index change detection method and Otsu automatic threshold segmentation method are selected for change detection.Among them,27199268 changed pixels were detected and512597836 unchanged pixels were detected.The detection results were verified by visual interpretation.The actual changed pixels were 25993074 and the unchanged pixels were513804030.Finally,the confusion matrix was used to evaluate the change detection results.The overall detection accuracy was 97.9 %,and the Kappa coefficient was 0.77.(3)Landslide susceptibility evaluation.The 30 m * 30 m grid evaluation unit was selected in Arc GIS software,and the susceptibility evaluation model based on information method was constructed by combining elevation,slope,aspect,roughness,stratigraphic lithology,distance from fault,distance from river,distance from road and NDVI evaluation factors.The natural breakpoint method was used to divide the area into non-prone area,lowprone area,medium-prone area and high-prone area.Finally,the ROC curve was used to test the evaluation results,and the AUC value was 0.867.