Research On Change Detection Method In Single-Polarization SAR Imagery

Using remote sensing image for change detection can effectively reduce the time and labor costs of man-made field surveying and mapping,improve the detection efficiency.It is a research application that has a key impact on national economy,people’s livelihood and homeland security.In different types of remote sensing images,as SAR(Synthetic Aperture Radar)is an all-time and all-weather technology,which can work normly in all light and climate conditions makes using SAR images for change detection has attracted much attention in the fields of post-earthquake loss estimation,flood range detection,urbanization research,and coastline extraction and so on.Limited to a unique imaging mechanism,SAR images will be affected by inherent coherent speckle noise,making it difficult to interpret images manually,which has prompted SAR imagery change detection research towards the direction of unsupervised and unmanned intervention.This paper takes single-polarized SAR image as the research object,and on the basis of combining the traditional SAR imagery change detection algorithm and deep learning technology,In-depth study of unsupervised SAR image change detection algorithms for different tasks.The research content of this paper mainly includes the following aspects:(1)Aiming at the task of binary SAR imagery change detection,a binary SAR imagery change detection algorithm based on difference-enhanced difference image segmentation and dual-channel convolutional neural network is proposed.The algorithm first proposes a difference image generation algorithm with difference enhancement effect and on this basis,develops an efficient threshold segmentation algorithm as a sample volume control factor;then improves the hierarchical FCM algorithm realizes the optimal selection of the sample set.Finally,a weight-adaptive dual-channel convolutional neural network is constructed to solve the problem of unbalanced sample classes when unsupervised sample selection.Experiments on the Gao-Fen 3 SAR images in Beijing show that the method can deal with the imbalance between clusters in the sample set better and achieve good results in change detection.(2)Aiming at the task of binary SAR imagery change detection,another SAR imagery change detection algorithm based on visual saliency and adaptive Focal loss convolutional neural network is proposed.Aiming at the problem that the existing difference images are sensitive to the difference in low-scattering areas and do not have the ability to focus on the areas of potential changes,a method for generating a difference image with multi-dimensional difference feature saliency detection with strong difference expression capabilities is proposed.Based on this,taking into account the characteristics of SAR images that are seriously affected by noise,a multi-level clustering algorithm that considers neighborhood information has been developed for reliable sample selection.Finally,in order to deal with the imbalance between intraclass and inter-class in the sample set,a dual-channel convolutional neural network based on an adaptive Focal loss loss function is used for sample learning and result classification.Experiments on the SAR image of Gao-fen 3 in Beijing show that the method can accurately identify the changed areas.(3)For ternary SAR imagery change detection,we propose a method based on the normalized maximal between-class variance and GKIT to segment a difference image by applying dual thresholds in SAR change detection.First,the normalized maximal between-class variance values of two sides surrounding the peak in the histogram are taken as the degrees of the overlapping gray level,and then the thresholds selection sequence and the candidate intervals are confirmed.Second,the side at which the gray level lightly overlaps is segmented by GKIT,and the threshold and the fitting function of the unchanged class are obtained.Third,the fitting function of the unchanged class is used to replace the corresponding part in the origin histogram to form a new histogram that is subsequently segmented to obtain the threshold in the second candidate interval.Finally,the two thresholds are applied on the difference image to obtain the final change result.The experiment on Gao-Fen 3 images reveals that the performance on our proposed method outperforms D-GKIT and can deal well with the overlapping gray level overlapped in the histogram of the difference image.The confusion matrix of the results for various local areas in the change image also shows that the proposed method has been slightly influenced by the overlapping gray level overlapped and obtains generally good results.(4)In terms of multi-class SAR image change detection,for the problem of multiclass change detection that is rarely involved in the current single-polarization SAR image change detection,this paper proposes an unsupervised multi-class SAR image change based on visual perception and PCANet Detection method.The method first uses the significant texture differences to determine the initial classification of pixels between two temporal images;then constructs a three-dimensional space,and uses the Mean-shift algorithm and visual perception to complete the description of the pixel space state;Combined with the initial classification results,a multi-class pixel annotation for model training was constructed.Finally,PCANet was used to learn and predict the input,and the purpose of unsupervised multi-class change detection of single-polarized SAR images was achieved.Experiments with the SAR image data of Gao-Fen 3 in Beijing show that the method in this paper can identify the changing area and identify the changing content,then the purpose of multi-class change detection was achieved.