Research On Building Change Detection Method Based On High Resolution Remote Sensing Image

At present,our country is in the promotion stage of urbanization.It is of great significance to rapidly and accurately realize building change detection for land space planning implementation,geographical conditions monitoring and city dynamic simulation.For remote sensing image building change detection task,manual interpretation method is time-consuming and arduous,and is not suitable for large-scale building change detection.However,many remote sensing image building change detection methods based on machine learning technology or deep learning technology are generally unable to effectively use global spatial context information.It is easy to cause misclassification problems caused by inconsistent scene semantics.In addition,the problem of pseudo changes of high-rise buildings caused by the difference of incidence Angle of two remote sensing images has not been solved well.In view of the above shortcomings,this paper intends to adopt deep learning technology to carry out research on building change detection methods of high-resolution remote sensing images based on deep learning technology from two aspects:pixel-level building change detection method and scene-pixel level building change detection method:(1)Pixel level building change detection method.In order to solve the problem that global context information is difficult to capture in remote sensing image change detection task,this paper uses the ideas of skip connection and twinning structure for reference,as well as Transformer structure in natural language processing field.A pixellevel building change detection model based on five-level feature extraction structure,feature fusion structure,twin structure and Transformer structure is proposed.The model encoder adopts a hybrid CNN-Transformer structure to further capture the global context information of remote sensing image for the encoding features obtained by CNN architecture through the self-attention mechanism of Transformer structure,which enhances the long-distance context modeling ability of the model for the change detection task of pixel-level remote sensing image.The experimental results show that compared with the existing classical pixel-level building change detection model,the proposed model can obtain better building change detection effect.The overall accuracy of the proposed model is 99.07%in the LEVIR-CD dataset and 98.38%in the CDD dataset.(2)Scene-pixel level building change detection method.The difference of incidence Angle is a common problem in two-phase remote sensing images.If the pixellevel building change detection method is used to detect building change in such images,the pseudo-change results of high-rise building change detection are likely to appear.To solve this problem,this paper proposes a scene-pixel level building change detection method,which includes building extraction stage and building change detection stage.Firstly,the convolutional neural network in the building extraction stage is used to extract buildings in the two phases of remote sensing images.In this stage,buildings in the remote sensing images can be extracted and non-buildings can be filtered.Then,the extracted buildings are input into the building change detection and discrimination network used in the building change detection stage,and the buildings in the two phases of the image are judged to change or not.In addition,in order to solve the problem of insufficient real class samples of building changes,pseudo samples are used as a means of data enhancement.Experimental results show that the scene-pixel level building change detection method can effectively alleviate the problem of high-rise building migration pseudo-change measurement caused by imaging incident Angle difference.The scene-pixel level building change detection method can obtain 97.05%scene-level overall accuracy and 87.64%Scene-pixel level overall accuracy in Zhuhai data set.