Change Detection Based On Semantic Segmentation For High-resolution Remote Sensing Image

With the continuous progress of technology,the resolution of remote sensing image is gradually improved.These high-resolution remote sensing images play an important role in production and national defense,such as satellite navigation,disaster early warning,land evolution analysis,etc.Among them,change detection is an important application of high-resolution remote sensing image.However,the traditional change detection task mainly focuses on whether a certain area has changed,which belongs to the binary classification problem,and it seldom involves the specific information of the change type,that is,what kind of ground object changes into another,which limits its application.In order to solve this problem,this paper deeply combines the semantic information with the traditional change detection task,and improves the traditional change detection task by means of semantic segmentation,so as to obtain the specific information about the change type.Specifically,this thesis carries out the following work.Firstly,the context aggregation network is designed for the semantic segmentation of remote sensing image.The network consists of context fusion module,attention mixing module and residual convolution module.The context fusion module is used to aggregate the context information of different receptive fields and introduce global information.Attention mixing module uses channel attention mechanism to combine multi-level features,and selectively emphasizes more differentiated features.The residual convolution module is used to refine the features of all levels.This thesis evaluates the proposed network on the open ISPRS Vaihingen and Potsdam datasets.The experimental results show that the performance of the model is better than existing best semantic segmentation models at that time.Secondly,in order to obtain the specific information about change types,this paper has studied semantic change pattern analysis(SCPA)based on semantic segmentation.SCPA can make full use of semantic information and analyze change types,which is a multi-class classification problem.Specifically,for a pair of registered images,the output of this task is a multi class classification result at the pixel level,and each class label represents the change type of the corresponding pixel pair.The change type here is defined by the pixel semantic labels in the source image and the target image,that is,the pixel changes from one type of object to another.Therefore,this task can accurately analyze the changes of each pixel and give the specific types of changes,which is very helpful for the subsequent analysis and application.Finally,a data set is constructed for the proposed SCPA task by manually marking pixel by pixel.This data set is based on a pair of registered high-resolution satellite images in Wuhan city.It is called SCPA Wuhan City(SCPA-WC)data set.It is the first accurate remote sensing image data set containing semantic information labels.On this basis,the SCPA task is implemented by semantic segmentation.The SCPA-WC data set is fully tested and the corresponding performance benchmark is established.The experimental results show that the SCPA task is very challenging,and the existing methods are far from satisfying,which is a very worthy field to explore.