| Synthetic Aperture Radar(SAR)image change detection is a technique used to detect changes in the same location between two SAR images captured at different times.It has been widely applied in estimating vegetation coverage and other areas.Compared to traditional algorithms,deep neural networks are better equipped to handle the challenges of unsupervised SAR image analysis with limited training samples.Therefore,this paper focuses on studying unsupervised SAR image change detection using deep neural networks.Although existing deep neural networks can effectively perform detection,they suffer from issues such as low-quality pseudo labels,inadequate extraction of image features by the network,and significant differences in data distributions,which inevitably affect detection accuracy.To address these problems,this paper presents the following main contributions:In the method that utilizes the fusion of difference images and Hierarchical Fuzzy C-Means(HFCM)for pseudo label generation,where the low quality of pseudo labels affects change detection accuracy,this paper proposes a method that utilizes the Dempster-Shafer(DS)evidence fusion theory to generate pseudo labels.The proposed method combines the advantages of three different difference images(log ratio-fuzzy membership difference image,mean log ratio-fuzzy membership difference image,and neighborhood mean log ratio difference image)through DS evidence fusion theory,thereby theoretically improving the quality of pseudo labels.Experimental results on multiple datasets demonstrate that the proposed algorithm achieves improved accuracy compared to existing algorithms.In the problem of insufficient feature utilization and failure to consider data distribution differences in the Dual Domain Network(DDNet),leading to poor network performance and subsequently affecting change detection accuracy,this paper improves DDNet by introducing the Wavelet Texture Convolutional Neural Network(WT-CNN)and proposes an unsupervised SAR image change detection algorithm based on log ratio difference image,HFCM,and WTCNN.Specifically,WT-CNN incorporates wavelet high-frequency texture details to more comprehensively extract deep texture features from images.Additionally,the proposed method employs cyclic training,where eligible test samples are iteratively added to the training set,reducing the adverse effects caused by data distribution differences.Comparative experimental results demonstrate that the proposed method achieves varying degrees of accuracy improvement on multiple datasets compared to existing algorithms.By combining the DS evidence fusion theory for pseudo label generation,WT-CNN,and cyclic training,this paper proposes an image change detection algorithm based on DS evidence fusion theory and WT-CNN.Firstly,the algorithm generates high-quality pseudo labels based on DS evidence fusion theory.Secondly,it utilizes WT-CNN to extract deep texture features from images and applies cyclic training to improve network performance.Comparative experimental results demonstrate that the proposed method achieves varying degrees of accuracy improvement on multiple datasets compared to existing algorithms. |
