SAR Image Change Detection Based On Siamese Capsule Network

Synthetic Aperture Radar(SAR)is an active microwave remote sensing technology with the characteristics of all-day,all-weather,strong penetration,and short revisit period.SAR image change detection refers to the analysis and processing of two or more SAR images collected from the same area at different times,so as to identify the changed areas.The continuous development of SAR technology has provided rich data support for change detection research.In recent years,deep learning has shown obvious advantages in the fields of computer vision and remote sensing image processing,thus providing a new research direction for SAR image change detection.To this end,based on siamese network and capsule network,this thesis focuses on the characteristics of SAR images and constructs two effective SAR image change detection algorithms.The main research work is summarised as follows:Existing SAR image change detection algorithms generally suffer from the problem of being susceptible to coherent speckle noise.In addition,the special multichannel structure of siamese network gives it the ability to handle multiple input data;however,there is relatively little research on siamese network-based SAR image change detection algorithms.In this thesis,by combining siamese network and attention mechanism,a dual attention siamese network(DCA＿SNet)is proposed for SAR image change detection.Firstly,a multiscale siamese network based on dual-channel attention structure is constructed to dynamically learn the scale information of land covers in SAR images,obtaining the effective deep features that can identify their change information.After that,the capsule network is introduced as the classifier of the whole model to improve its ability to classify the changed and unchanged regions in SAR images.Experimental results demonstrate the effectiveness of the proposed DCA＿SNet model.By considering the advantages of capsule network in mining spatial location and attribute information of data,this thesis further utilizes capsule network to construct a U-Net siamese capsule network(U＿SNet)for SAR image change detection.Firstly,due to the end-to-end and high-precision classification characteristics of U-Net and the problem that the existing change detection algorithms fail to accurately identify the small changes in SAR images,the siamese capsule network based on U-Net is firstly designed,which can make full use of the semantic and spatial hierarchical information of SAR images by constructing the encoder-decoder and skip-connection structures through convolutional capsule and deconvolutional capsule.Then,another capsule network is designed to obtain the changed and unchanged results of SAR images.Experiments on three real SAR image datasets show that the proposed model has better change detection performance than other algorithms.