Research On Intelligent Recognition Of Optical Remote Sensing Image Of Collapse And Landslide Disaster

Bijie City is located in the southwest of Guizhou Province.With complex lithology and geological structure and frequent human engineering activities,it is a frequent occurrence area of geological disasters,among which collapse and landslide disaster is the most significant.On the basis of summarizing the data of the collapse and landslide disasters in Bijie City,this paper takes the optical remote sensing image of the collapse and landslide disaster in the study area as the data source,establishes the remote sensing characteristics of the landslide and landslide respectively,and establishes the data set of the collapse and landslide disaster.Then,the development and distribution information of collapse and landslide disaster at wide scale was identified by YOLOv3 model.Finally,the U-Vi T model is used to automatically identify the disaster range of collapse and landslide disaster.The main research results are as follows:(1)Regional geological conditions and distribution law of collapse and landslide disaster in Bijie City.Based on DEM data,meteorological information,1:1 million geological map and Google Earth optical remote sensing images,it was concluded that the collapse and landslide disasters in the study area mostly occur in the low and middle mountains,the slope was 0 ～ 25° and the fluctuation was below 150°.(2)Remote sensing characteristics of collapse and landslide disaster.The image characteristics of collapse disaster and collapse and landslide disaster in optical remote sensing images are analyzed based on morphological characteristics,texture characteristics,spectral characteristics and structural characteristics.The results showed that the boundary shape of collapse and landslide disasters was changeable,in collapse and landslide disasters,joints and cracks are more common at the rear edge of slope,scratches and scratches were more common in the slip area,and gravel and solitary stone were more common in the accumulation area,which enhanced the complexity of image features in the accumulation area.(3)Multiple target detection models were compared and analyzed to realize the automatic identification of collapse and landslide disaster distribution on a wide scale.Based on the open data set of collapse and landslide disaster in Bijie City and the Google Earth remote sensing images of Bijie City collected in 2022 as source data,this paper completes the expansion of data diversity through data enhancement operations.Qualitative analysis was made on the identification effect of four models Faster R-CNN,SSD,YOLOv3 and YOLOv6 in optical remote sensing images of collapse and landslide disaster at different scales.Four evaluation indexes,Recall,Precision,F1 Score and m AP,were used to quantitatively analyze the identification accuracy of the model.The results showed that the YOLOv3 model had a good identification effect on the collapse and landslide disaster in Bijie City.The above four evaluation indexes were 0.6086,0.9143,0.73 and 0.7246,respectively.Finally,the YOLOv3 model was used to automatically identify the development and distribution of wide-scale collapse and landslide disasters in Zhijin County,Bijie City,and finally obtained the specific locations and spatial distribution characteristics of 152 collapse and landslide disasters in this region,which were mainly distributed in Jinfeng Street,Zhongzhai Township,Niuchang Town and other areas.Therefore,the YOLOv3 model has a good identification effect for the collapse and landslide disaster in Bijie City,and has the automatic identification ability for the development and distribution location of the collapse and landslide disaster.(4)The U-Vi T model was constructed based on U-Net and Vi T models to realize the identification of the disaster range caused by the single scale collapse and landslide disaster.Based on the source data used in the object detection model,the data set is augmented by data enhancement operations.The U-Vi T model is improved by using Vi T and Conv Transpose2 d convolutional neural network.By comparing and analyzing the influence of six learning rate variation algorithms on the model identification of the loss value of sliding disaster,it is concluded that Step LR algorithm has the lowest loss value in the identification of sliding disaster.Through qualitative analysis of the recognition effect of Vi T model and quantitative analysis of the recognition accuracy of U-Vi T and U-Net model in pixel accuracy(PA),class average pixel accuracy(MPA)and average intersection ratio(MIo U),it is obtained that U-Vi T model has higher recognition longitude than U-Net model.It can significantly improve the recognition performance of optical remote sensing image of sliding disaster.Finally,the U-Vi T model is used to realize the automatic identification of the collapse and landslide disaster area of Jinfeng Street at the single scale,and the relative errors of the identification results are all less than 0.11.Therefore,the improved U-Vi T model has good recognition performance,which is helpful to improve the level of intelligent recognition of the disaster range caused by sliding disaster.