A SAR Image Change Detection Network Based On Sample Imbalance Learning And Structural Semantic Information

Synthetic Aperture Rader(SAR)is a high-resolution imaging radar that can be used in all-weather and all-weather applications.It is widely used in remote sensing.SAR image change detection is obtained by comparing and analyzing the same region and at different time.The SAR image is used to determine the change of the ground object,and the qualitative and quantitative description of the change.It has a wide range of applications in disaster analysis,agricultural surveys,and resource monitoring.The richness of SAR images provides data support for change detection,but due to the imaging principle of SAR images,there is speckle noise attached to the SAR image,and it also cause geometric distortion and radiation distortion of the image.Research on the topic of change detection has increased the difficulty.This thesis first introduces the general flow and classic methods of SAR image change detection under the traditional framework,and introduces the application of deep learning in the field of change detection too.Based on the analysis of the advantages and disadvantages of existing methods,the accuracy is higher and more efficient methods for change detection are explored,the current research results are as follows:The imbalanced problem in SAR image change detection is studied.A method of SAR image change detection based on sample imbalance learning and multi-layer principal component analysis is proposed.This method solves the imbalance problem in the change detection by the method of morphological-based sample selection;then the multi-layer principal component analysis network PCA-Net is used as the feature extractor,in which the convolution kernel parameter is the feature vectors of PCA by the training sample.So the kernel parameter is composed without updating the parameters by backpropagation,which greatly improves the efficiency of change detection.Finally,a linear SVM is used for classification to obtain the final change detection result.The advantage of this method is that it is simple and effective.It has neither the cumbersome steps of the traditional method nor the back propagation to update the parameters,and solves the sample imbalance problem and achieves good results.Aiming at the problem that most methods are sensitive to difference graphs,a SAR image change detection method based on cascaded deep semantic forest is proposed.The purpose of this method is to obtain a clear change region structure from the difference map of noise pollution..The cascading depth semantic forest proposed by the method is composed of a plurality of sub-modules,wherein each sub-module comprises two parts,the first part is a feature extractor for extracting the representation features in the input image,and the second part is a deep forest,Learning the semantic context information of the change region,and then using the representation features of the semantic context information to update the deep semantic forest,the learning representation feature and the update depth semantic forest will alternate.It can be seen from the output of each sub-module in the experiment that the change probability map is more and more clear,which proves the effectiveness of the method by adding semantic context information.A SAR image change detection method based on pyramid pooling model is proposed.The method borrows the idea of full convolution,divides the image into several blocks,obtains the predicted value of each pixel in each image block through the pyramid pooling model,and finally uses the voting method to obtain the final change detection result.In the previous method,a single pixel tag is used.In this method,the association of the tags in the neighborhood is used,and the network training using the tags of all the pixels in one image block can effectively reduce the influence of noise,and the training method enables the chapter to be the same.Learning different types of changes in multiple sets of data in a network and directly testing a new set of unlabeled data.