Research On Extraction Methods Of Damaged Buildings After Earthquake Based On Optical Remote Sensing

It is often unable to get better results that extracting information from damaged buildings after earthquakes using remote sensing technology.Most of the extraction characteristics of damaged buildings are based on artificial experience and lack of systematic generalization.In addition,the ground objects should have an optimal expression scale in remote sensing images based on previous studies.Only when the objects are studied under the best remote sensing scale can they have a relatively superior result.In view of the problems in remote sensing extraction of damaged buildings,the following aspects are mainly studied in this paper using object oriented method and taking optical remote sensing image as the data source based the support of ENVI,ArcGIS and e-Cognition software platform.(1)The damaged building information is extracted by using the object-oriented method after preprocessing the remote sensing data of the earthquake cases.Object oriented methods include segmenting images by using ESP(Estimate Scale Parameter)and classifying image pairs by using machine learning methods(classification and decision tree(CART),support vector machine(SVM),Random Trees)and the method based on optimal feature space.Then,we compare the four kinds of damaged building extraction methods through the accuracy evaluation,and analyze the relative merits of several extraction methods.(2)The object oriented machine learning method and the extraction method based on the optimal feature space are used to extract the damaged buildings in the remote sensing images of different data sources,different research areas and different seismic cases based on the first part.The rule sets of the damaged building extraction are summarized.(3)The damaged building information is extracted in different spatial resolution images obtained from resampling.And the best observation scale of damaged building is explored.The results show that(1)Random forest method has the best extraction results in several machine learning methods.The classification results based on the optimal feature space extraction method are also good,and the overall accuracy of the classification is even better than that of random forest method.(2)The optimal feature space of the damaged building changes with the change of the study area and the data source,and its universality is poor.Some characteristic parameters(GLCM Homogeneity ? Border index ? Asymmetry ? Density)remain unchanged in the experimental images,showing a strong generality.Combining CART method,the rule set of the damaged building extraction is finally obtained,including 12 characteristics: Mean of R,Mean of B,Ratio of R,Ratio of B,Brightness,Shape index,Asymmetry,Density,Border index,GLCM Homogeneity,GLCM Mean,GLCM Contrast.(3)The Hellden and Short indexes are used to evaluate the accuracy.The change trend of the damaged building extraction effect is obtained with the change of the observation scale,and the best observation scale is between 1m-1.2m.In this paper,the method of information extraction for damaged buildings is computed automatically by computer no matter in the optimal segmentation scale selection or in the optimal feature space construction,which reduces the interference of human subjective factors.In addition,the rules set and the best observation scale established in this paper all have good popularization value.They can be used for reference in future research or application to improve the accuracy and efficiency of information extraction.