The Research On Building Change Detection Method Of Earthquake Damage Remote Sensing Image Based On Deep Learning

Earthquake disasters are often relatively sudden and destructive,and are one of the natural disasters that cause the most economic losses and casualties in the world.After the earthquake disaster,timely and accurate disaster loss assessment of the affected areas is of great significance for formulating emergency rescue plans and realizing post-disaster rescue and reconstruction.With the continuous development of remote sensing technology and image processing technology,data analysis of earth observation remote sensing images has become an important tool for detailed investigation and post-disaster reconstruction.The representative features in high-resolution optical remote sensing images—the collapse and damage of buildings is a favorable means to judge the loss of earthquake damage,so the change detection of buildings based on dual-temporal remote sensing images can be better explored changes in the damage of earthquake-damaged buildings lay the foundation for earthquake disaster monitoring and damage assessment.Since the change detection method based on deep learning can automatically learn the change features in remote sensing images of different phases,by comparing the changes of ground objects in the dual-temporal images before and after,the representative change features can be effectively extracted,and the interference factors such as false changes can be reduced.Therefore,deep learning technology is gradually becoming popular in the field of remote sensing image change detection and analysis.On the basis of fully researching and analyzing the existing excellent methods,this paper summarizes some problems that still exist in the current methods,t he research content of this paper mainly includes the following aspects:(1)Construct a multi-category earthquake-damaged building change detection dataset.Aiming at the current lack of data sets related to multi-category earthquake damage change detection of buildings,this paper uses the labelme labeling tool to analyze the earthquake damage based on the high-resolution remote sensing image data of Indonesia’s disaster-affected area in two phases and according to the x BD disaster data labeling evaluation standard.Annotate building changes,and combine Python language and image processing libraries to process the JSON files generated by the annotations to generate binary annotation image data.In view of the disadvantages of limited data volume,this paper also expands the number and diversity of data set samples through geometric transformation data enhancement.In addition,combined with the open-source binary classification change detection dataset LEVIR-CD,a set of multi-classification seismic damage building change detection dataset(MSMC Dataset)was jointly constructed.(2)Construct a multi-classification change detection method for buildings in remote sensing images of earthquake damage based on a multi-level attention mechanism(MLANet).This paper designs a new multi-level spatial attention mechanism(MLAM)by combining spatial attention mechanism and channel attention mechanism,and proposes an integrated Siamese network and UNet++ network It is the encoding-decoding structure of the backbone network,and MLAM is embedded in the decoding structure to form a new multi-classification remote sensing image change detection algorithm(MLANet).The model uses a twin structure to activate the same position in the feature map of the dual-temporal remote sensing image by sharing weights.And use the UNet++ structure with dense skip connections to fuse multi-level features to provide richer neighborhood information.The designed MLAM effectively alleviates the problems of insufficient natural aggregation and spatial mismatch in feature information fusion,highlights the effectiveness of information at each feature layer,and promotes the fusion of detailed features.Compared with other state-of-the-art change detection network models,MLANet effectively improves the accuracy of building change detection.(3)Construct a multi-classification change detection method for buildings in earthquake damage remote sensing images based on lightweight adjustment modules(LRBNet).Aiming at the problem that the deep learning network consumes a lot of computing resources although it has better recognition performance,this paper proposes a new lightweight feature extraction module(LCM)to replace the traditional convolution operation mode,which significantly reduces the computational complexity.A lightweight residual structure(LRB)based on the LCM is also designed,and LRB is applied to the overall backbone network architecture of the fusion of twin networks and UNet++ to complete the feature extraction process of the overall network architecture,which greatly reduces the calculation amount of the model and realizes the rapid positioning of the disaster-affected area of the building.In addition,in order to refine and summarize the change semantics of multi-level earthquake-damaged buildings and suppress similar semantic gaps and location deviations,a designed multi-level damage feature aggregation module(MDAAM)is embedded in the last stage of the network.MDAAM can focus on the analysis of important information in the space domain and channel domain,thereby improving the performance of building change detection and finding an effective balance between recognition accuracy and efficiency.