Research On Automatic Extraction Of Remote Sensing Images Of Co-seismic Landslides

Co-seismic landslides often cause road damage,river blockage,house burial and bridge collapse,resulting in the failure of emergency rescue and on-site investigation,which will seriously affect life rescue and earthquake disaster assessment.Therefore,the rapid and accurate acquisition of the location,distribution,size and buried buildings,roads,vegetation and other related information of earthquake landslides is of great significance to guide earthquake emergency rescue,disaster assessment and post-disaster reconstruction.With the rapid development of satellite remote sensing and spatial information technology,remote sensing image data has the characteristics of wide coverage,short acquisition time,low cost and large amount of data,which can provide data basis for seismic landslide identification in the whole earthquake area.After the earthquake,the rapid identification and mapping of co-seismic landslides in the whole earthquake area is also an important basis for understanding the development degree,distribution law and risk assessment of co-seismic landslides.Nowadays,with computer technology developing,it shows great advantages in earthquake landslide identification through the expert experience of system quantitative simulation,and have high efficiency emergency response ability based on advanced technology and algorithms.with the rise of artificial intelligence,especially deep learning,machine learning methods have been used in earthquake landslide identification.Relevant researchers have achieved good results in small-scale and single-environment experimental research.However,in the face of the whole earthquake area with large-scale and complex environment,the correct rate of extracting co-seismic landslides is low,and there is no ideal method.Based on the Planet satellite remote sensing images of 3m resolution before and after the M_w6.6 earthquake in the central and eastern part of Hokkaido,Japan on September 6,2018,this paper studies the automatic identification technology of earthquake landslides in complex environment in the whole earthquake area dominated by remote sensing images.Based on the Planet true color images before and after the earthquake,combined with the three-dimensional visualization of Google Earth,the landslides induced by the M_w6.6 earthquake in Hokkaido,Japan on September 6,2018were interpreted by visual interpretation.A total of 9315 co-seismic landslides were interpreted in the study area,with a total area of 30.99 square kilometers.Ten groups of automatic extraction experiments of seismic landslides are carried out on Natural True Color Images(RGB),Standard False Color Images(FCS),Band Ratio Operation Images(BR)and Normalized Difference Vegetation Index Operation Images(NDVI)by Image Difference Method,Maximum Likelihood Method,Sample-based Object-oriented Classification Method and Deep Learning Method,and the extraction results are evaluated and compared with the accuracy of confusion matrix.It is considered that under the premise of ensuring the high extraction rate and accuracy of real earthquake landslides,in order to meet the timeliness requirements of earthquake emergency rescue,depth learning methods based on natural true color images and standard false color images can be selected.in the follow-up research,we can continue to improve the automatic extraction model of seismic landslides,and can try to extract large-scale landslides caused by different environments and rainfall.The object-oriented classification method based on natural true color image and standard false color image,although the time cost is high,but compared with manual visual interpretation,the time cost and labor cost are lower,and the accuracy of earthquake landslide extraction results is also better.it can be used as the basis of follow-up verification work.Although the extraction time of maximum likelihood method based on traditional supervised classification is fast,it is of little practical significance for post-earthquake emergency relief.However,the image difference method based on band ratio operation image and normalized vegetation index operation image is not recommended because of its high requirements for images and it is not easy to meet the conditions in earthquake emergency,so it is not recommended to use these two methods when there are no other better methods and need to quickly evaluate the disaster situation,but the reference significance of earthquake landslide extraction results is limited.