Damaged Building Extraction From Remote Sensing Images Based On Adversarial Adaptation Networks

Damaged buildings are the key areas of emergency rescue after earthquakes,and an efficient and accurate assessment of them can provide strong information support for rescue operations.Remote sensing images have the advantages of large-scale earth observation,fast acquisition speed and short period,and have become one of the main methods for evaluating damaged buildings.With the continuous development of remote sensing technology,highresolution remote sensing images have richer features,while traditional artificial feature design methods cannot fully learn rich image features.The method based on deep learning has huge advantages in extracting highly abstract features,and gradually replaces the traditional method of extracting damaged buildings in the task of extracting damaged buildings,becoming the mainstream.However,training models based on deep learning methods requires a large number of high-quality labeled samples,which contradicts the time urgency of emergency rescue after earthquakes.A common solution to this problem of lack of labeled samples is to adopt a domain adaptive learning method to apply the features from existing damaged building samples to new post-earthquake images without labeled samples after transfer learning.However,the current research objects of existing domain adaptive learning methods are not the damaged buildings in remote sensing images,and the existing computer vision domain adaptive methods are directly applied to the task of extracting damaged buildings from remote sensing images,which lacks the pertinence of the task.Therefore,on the basis of adversarial domain adaptation,this thesis divides the damaged building extraction based on adversarial domain adaptation into scene classification and semantic segmentation from two directions: direct use of building vector information and indirect extraction of building location information.From this perspective,the research on the extraction of damaged buildings based on adversarial domain adaptation is carried out,and the improved design of the adversarial domain adaptive network for damaged buildings is further discussed,so that damaged buildings can be more accurately extracted from postearthquake remote sensing images.The main work and conclusions of this thesis are summarized as follows:(1)For the scene classification model that directly utilizes building vector information,an adversarial domain adaptive method based on dynamic alignment of local features and sample equalization damage is proposed.In this method,the scene samples and annotation features of the source domain are transferred to the target domain,and the following three modules are added to the traditional adversarial domain adaptive network: 1)First,a local feature alignment module based on conditional probability distribution is added to solve the problem of class 2)An adaptive dynamic balance weight is used between the global feature and the local feature,and the weight learns the importance of the global feature and the local feature autonomously during training,so as to improve the application of the method in different target domains.3)The sample equalization coefficient is added to the classifier of the adversarial domain adaptive network to improve the long-tail phenomenon in training.According to the experimental results and comparative analysis,the following important conclusions can be drawn: 1)The local feature alignment module significantly improves the F1 score while improving the overall accuracy,indicating that the method does not reduce the recognition accuracy of intact buildings.2)The prediction effect of the selected typical semantic segmentation model for the edge is still far from the ideal result,especially for dense changing areas,these models can easily combine multiple changing areas.The prediction is a region of change,which shows that the edge prediction ability of the model needs to be improved.The dynamic balance weight adaptively obtains the final weight coefficient in different target domain data,and can improve the recognition accuracy of the target domain,indicating the effectiveness of the dynamic balance coefficient;3)The sample balance coefficient improves the final weight as a whole.The accuracy of building recognition is damaged,so that the network training will not pay too much attention to the samples of buildings with intact heads.(2)For the semantic segmentation model that indirectly extracts the location information of buildings,an adversarial domain adaptation method based on inter-domain-intra-domain collaboration is proposed.In this method,the semantic segmentation samples and annotation features of the source domain are transferred to the target domain,the feature extraction layer is transformed into a twin extraction layer in the traditional adversarial domain adaptive network,and pre-disaster and post-disaster remote sensing images are input simultaneously.The sharing reduces the extraction noise caused by the deviation of the two-phase images,and first reduces the overall difference between domains through the adversarial loss;then,the pseudo-labels of the target domain are sorted by the image entropy value to classify the easy transfer class and the hard transfer class.Domain adaptation is performed again within the domain to improve the problem of poor migration of some samples caused by the intra-class heterogeneity of damaged building samples.According to the experimental results and comparative analysis,the following important conclusions can be drawn: 1)The adversarial domain adaptive network based on the Siamese network improves the extraction effect of severely damaged buildings and improves the overall extraction accuracy;2)The selected typical semantic segmentation The prediction effect of the model on the edge is still a certain distance from the ideal result.The domain adaptation based on the minimization of entropy value further optimizes the domain adaptation effect and improves the extraction results of some target domain samples with poor results.