Research On Earthquake-damaged Buildings Detection Method Based On Post-earthquake High-resolution Remote Sensing Images

Earthquake is a serious natural disaster.Once occurred,it will cause great damage to the safety of human's life and property,and the damage of buildings is an important basis to judge casualties and property losses.With the increasing coverage and image resolution of remote sensing images,it has become one of the important means to detect earthquake-damaged buildings by using high-resolution remote sensing images after earthquake.At present,the methods of earthquake-damaged buildings detection based on post-seismic high-resolution remote sensing images are mainly divided into the traditional classification model method and the deep learning method.The former has the advantages of fast algorithm implementation and good accuracy on small datasets.However,the feature sets designed and extracted by hand may have feature redundancy and feature conflict.How to extract the most representative feature sets is a difficult problem faced by this kind of method.The latter mainly realizes "end-to-end" classification through deep convolutional neural network,but the convolution-based operation usually only uses the spatial context information and ignores the category context information in the whole image.In addition,semantic segmentation requires predicting categories for each pixel,while the current method treats each pixel equally,resulting in limited accuracy of segmentation boundaries.Therefore,this paper studies the methods of earthquake-damaged buildings detection based on traditional classification model and the deep learning respectively.The specific research results are as follows:(1)In the method of earthquake-damaged buildings detection based on traditional classification model,a method based on feature and decision tree optimization random forest is proposed.Firstly,the high-resolution remote sensing image is segmented by WJSEG algorithm,and the set of potential building objects is obtained by using geometric morphological features.Secondly,the number of decision trees is selected adaptively based on the proposed classification accuracy fluctuation curve.On this basis,under the guidance of feature importance index,three categories of features,namely spectrum,texture and geometric morphology,were screened to obtain a representative set of seismic damage features.Finally,the proposed random forest classification model is used to further classify the potential objective buildings into intact buildings,seismic damaged buildings,ruins and others.The results show that the overall accuracy of the proposed method can reach more than85% and the error detection rate is less than 6% in the experiments of multiple groups of post-earthquake high-resolution remote sensing images from different regions and different sensors,which can provide critical and reliable supportive information for post-earthquake emergency response and post-earthquake reconstruction.(2)In the method of earthquake-damaged buildings detection based on deep learning,a semantic segmentation method of post-earthquake remote sensing images based on object context and boundary enhanced loss function is proposed.On the basis of Deep Labv3+ and UNet respectively,the object context attention module is added to improve the feature learning ability of the model from both the spatial context and the object context.In addition,aiming at the difficulty of boundary segmentation caused by the mixing of ground objects in complex post-earthquake remote sensing scenes,a boundary index based on pixel spatial position information is proposed to improve the Focal Loss,so that the model can pay more attention to the learning of boundary pixels.It is founded that in the classification of post-earthquake datasets annotated by hand,UNet has multiple skip-connections which are beneficial to the recovery of spatial details,so it has a higher performance than Deep Labv3+.The overall accuracy of Deep Labv3+ and UNet with access to the object context module is 0.2%?0.8% higher than that without access,and the average Io U is 2%?3% higher than that without access,which verifies the effectiveness of feature enhancement using the object context attention module.Among them,the method of connecting OCR module in series in Deep Labv3+ is more accurate than that of connecting OCR module in parallel.In addition,the proposed bounding enhancement loss function BE Loss is used in Deep Labv3+ and UNet respectively,and compared with the other two loss functions,the results show that the overall accuracy of the proposed loss function is the highest.