Research On Identification Of Building Damage Information From Remote Sensing Image Based On Deep Learning

Once an earthquake,especially a large earthquake occurs,it will cause great disasters to human society.After the earthquake,buildings as one of an important disaster exposure,the collapse of which would cause casualties and economic losses.When an earthquake occurs,the roads,communications and other facilities are usually damaged in the disaster areas,so it is usually difficult to know about the situation of disasters area.At this time,the acquirement of information about the disaster area is the primary work which is important for the decision making to assign relief forces to release the losses.The disaster information can be acquired quickly through remote sensing technology which can quickly and extensively acquire the image of the disaster area.But for the remote sensing image interpretation,the traditional pixel-based and object-oriented methods have some problems and shortcomings.The precision of pixel-based method is low and the interpretation results usually company with "salt and pepper" noise.As for object-oriented method,image segmentation usually restricts the accuracy of image interpretation.While with the development of deep learning technology represented by neural network,it has achieved good application effect in the field of image processing.Therefore,deep learning method is applied to the identification of building seismic damage in order to improve the automation degree.The primary works of this paper include the following aspects:(1)Traditional methods used to earthquake damage extraction of buildings from remote sensing image are analyzed.Seismic damage extraction of buildings by pixel-based method is naive which only use spectral information of the image and can't make full use of the context information.As a result,the extraction accuracy is low and "salt and pepper " noise will be generated.Object-oriented technology,an extension of the pixel-based technology,is widely used to extract damage information.To a certain extent,Object-oriented technology overcomes pixel-based technology's shortcomings that can't make full use of context information.The first step of Object-oriented technology is segmentation,but the image patches acquired bysegmentation often cannot accurately represent the object,and this phenomenon affected the accuracy of the method.(2)The development status of deep learning technology in the field of image processing,especially in the field of image semantic segmentation,is researched.The application of neural network in the field of image semantic segmentation can be summarized as the following methods:(a)The method based on the traditional convolutional neural network is the na?ve method which processing speed is slow.(b)The fully convolutional neural network method is a development of the traditional convolutional neural network.It improves the efficiency of image semantic segmentation to a certain extent,but its ability to recognize object boundaries is poor.(c)The neural network based on encoder and decoder structure improves the recognition accuracy of object boundary in some extent.(3)Based on the comparison of various methods,this paper chooses the fully convolutional neural network and the neural network based on encoder and decoder to extract building damage information.The 7.1-magnitude yushu-earthquake in 2010 was selected as an analysis case,and the research area was part of the urban area of yushu county.Based on the aerial image of the research area,samples for building damage identification are selected.(4)The seismic damage buildings is identified based on the fully convolutional neural network.The fully convolutional neural network mainly includes convolutional layers,pooling layers and up-sampling layers,and it also adopts the strategy of skip-layers.393 images with size of 500×500 pixels were used to train the network,and another 34 images of 500×500 pixels were used to test the accuracy of the network.The test results show that the overall pixel accuracy of building seismic damage extraction can reach 82%,and the Kappa index is 62%,which indicates that this method has the ability to extract building seismic damage information.At the same time,the existing problems of the network structure are analyzed.(5)Deeplabv3+ model is used to identify buildings seismic damage information.Deeplabv3+ model includes atrous convolution,depthwise separable convolution,modified Aligned Xception model etc.The same training and testing data as the fullyconvolutional neural network are used to train and test Deeplabv3+ model.The result shows that this model has a good result in the extraction of building seismic damage information.The overall accuracy is 84%.Kappa index is 0.79.The extraction accuracy of the background,collapsed buildings and uncollapsed buildings is 0.909,0.833 and 0.904 respectively.The influence of different number of samples on the accuracy is preliminarily researched.The result shows that although the number of samples has some influence on the accuracy,a little number of samples used to train the module but it still reach a high accuracy(the overall accuracy is higher than 80%).