Research On Earthquake-damaged Building Detection Based On High-resolution Remote Sensing Image

Detection of earthquake-damaged buildings based on high-resolution remote sensing images is of great significance for emergency rescue and post-disaster reconstruction.At present,earthquake-damaged buildings detection methods can be divided into classical machine learning and deep learning methods.The former uses user-defined low-level and middle-level semantic features and is suitable for small sample applications.The latter can automatically extract discriminative and representative abstract features,and can usually obtain more accurate detection results under sufficient training samples.Therefore,the two methods have their own advantages in practical application,but at the same time they face different challenges.Therefore,this paper studies the difficulties faced by the two detection methods respectively.The specific research contents are as follows:(1)Based on classical machine learning support vector machine classifier,this paper proposes a method of earthquake-damaged buildings detection based on sparse dictionary.The lack of spatial context features and information redundancy in multi-feature visual dictionaries are discussed.Firstly,the earthquake-damaged buildings are depicted from multiple angles combined with spectral,textural and geometric morphological features;secondly,the spatial context information is further introduced by constructing the same and different pairs of words,so as to construct a multi-feature initial visual dictionary;on this basis,K-singular value decomposition algorithm is used to perform sparse representation of the visual dictionary to reduce redundant information as much as possible;finally,the final detection result is obtained by support vector machine.The results of multiple post-earthquake image experiments show that the overall accuracy of the proposed method can reach over 85%,and it is significantly superior to the comparison method in visual interpretation and quantitative analysis.(2)Based on UNet in deep learning,this paper proposes a seismic damage building detection method based on mutual attention module and cost sensitive iterative loss.Firstly,aiming at the problem that UNet ignores the semantic difference between high-level and low-level features,a mutual attention module is designed in the skip connections used in UNet architecture to enhance the capability of feature expression by sufficiently using high-level and low-level feature information.On this basis,aiming at the problem that the general loss function will cause the network to pay less attention to the earthquake-damaged building samples,a novel cost sensitive iterative loss is designed in an attempt to compel the proposed model to more focus on the learning of earthquake-damaged building feature samples.To verify the performance of the proposed method,datasets for detections of post-earthquake images for the collapsed buildings in the Yushu earthquake and Haiti earthquake are set.Experimental results indicate that embedding mutual attention module and cost sensitive iterative loss is conducive to significantly improving detection precision of earthquake-damaged buildings.This finding proves that the proposed method is feasible and effective.(3)The influence of the number of training samples on the overall accuracy and the comprehensive performance of the two methods are compared and analyzed.The results show that the method in chapter 3 has more advantages in the number of training samples,hardware requirements,running time and object-level recognition results.The method in chapter 4 has a significant advantage of detection accuracy under sufficient sample size.To sum up,this paper provides two efficient solutions for the detection of earthquake-damaged buildings under different sample conditions,and the research results are conducive to promoting the application of remote sensing technology in the field of post-earthquake emergency response and post-earthquake reconstruction.