Research On SAR Image Change Detection Based On Superpixel Segmentation And Siamese Networ

Remote sensing change detection(CD)technology plays a crucial role in various applications including urban planning,environmental monitoring,agricultural investigation,and disaster assessment,as it has the capability to continuously monitor large-scale changes on the Earth’s surface.Synthetic Aperture Radar(SAR)is an active remote sensing technology that is less affected by factors such as illumination and atmospheric conditions.This attribute allows SAR to work continuously and flexibly with all-day and all-weather imaging,thus positioning it as an important data source for CD.Currently,SAR image CD techniques mainly involve knowledge-driven and datadriven methods.The knowledge-driven methods are primarily based on mathematical physics model with strong interpretability,but lacks robust feature mining capabilities.However,data-driven methods can effectively extract deep-level features,but often require a large number of samples for model training,which is difficult to obtain for SAR data.To solve the above problems,this paper investigates deep learning-based change detection(DLCD)for SAR images,and proposes an unsupervised SAR image CD method that combines superpixel segmentation and Siamese network.The research conducted in this paper is summarized as follows:(1)This study explores DLCD techniques used for SAR images.Through comparative analysis of classical neural networks in Deep Learning and models widely used in DLCD,it was found that fully convolutional semantic segmentation networks are more suitable for CD tasks.Furthermore,several representative CD models based on fully convolutional networks in the DLCD field were compared and extended for SAR data to evaluate their performance.Experimental results show that compared to single-stream networks,the Siamese network architecture can independently process the multi-temporal data to fully mine the multi-temporal features,thus being more advantageous in CD.In addition,adding high-quality explicit difference guidance information to the model can help improve the accuracy of CD.(2)A novel end-to-end unsupervised CD method combining self-adaptive superpixel segmentation and Siamese network for SAR images is proposed.Firstly,an unsupervised pre-task was used to select reliable training samples.Then,a superpixel segmentation network is introduced and connected with a Siamese network to construct the CD model.A joint loss function containing task-adaptive loss is defined to train the superpixel segmentation network and Siamese network end-to-end until the global optimal parameters are obtained.Finally,a binary change map is output.Combined superpixel segmentation effectively suppresses speckle noise and enhances the semantic perception ability of the model.In addition,the backpropagation of the taskspecific loss function promotes the adaptive adjustment of the superpixel.The design of the Siamese structure and the adaptive adjustment of the superpixel ensure the consistency of the superpixel segmentation in the unchanged areas of multi-temporal data,while the segmentation in the changed areas is tightly adhered to the change boundary.Several public SAR CD datasets are used to verify the effectiveness of the proposed method.Compared with seven other advanced DLCD methods,the proposed method achieved the highest accuracy in OA,F1-score,and Kappa,and also showed superiority in suppressing speckle noise,refining the change boundary,and improving the detection accuracy of small area changes.(3)This paper explores the ability of CD models to detect changes based on transfer learning,which is closely related to the generalization performance of the model,and has received little attention and discussion in previous research.The transfer learning experiment between homologous and heterogeneous data(optical and SAR)is designed to explore the ability of the CD model to detect changes.Simultaneously,the superior performance of the proposed method combined with superpixel segmentation is further verified,which has a strong ability to detect changes and a good generalization performance.This study also found that adding high-quality explicit difference guidance to the model effectively enhances the ability of CD model to detect changes.Finally,the importance of focusing on the ability of CD models to detect changes in research on sample-free or small-sample CD tasks using SAR images is analyzed and summarized.