Research On Multi-temporal SAR Image Change Detection Based On Clustering And Deep Learning

The research of remote sensing image change detection technology is a crucial topic in the field of image processing,and the purpose is to process and analyze two images acquired over the same area at different times,so as to judge whether the area has changed and locate the changed area.Synthetic Aperture Radar(SAR)has own uniqueness,the imaging process is not affected by poor environments,and it can also image under adverse conditions.Meanwhile,SAR has the ability to continuously collect ground object information at all times and all days.So,it is widely used in many applications,such as urban planning,disaster assessment,environmental protection,and military reconnaissance.In these applications,multi-temporal SAR image change detection technology is one of the key technologies in SAR image research.So it has attracted the attention of more and more scholars.However,due to the special imaging principle of SAR,inherent speckle noise will be introduced during the imaging process,and the ground object background is extremely complex,which brings many difficulties to change detection.In response to the above problems,this thesis carried out in-depth research on multi-temporal SAR image change detection technology,and achieved a series of results as follows.(1)Aiming at the interference caused by strong speckle noise and the existing approaches do not consider the spatial semantic proximity relationship,a SAR image change detection approach based on saliency and FINCH(First Integer Neighbor Clustering Hierarchy)clustering is proposed.First,the log ratio operator and the subtraction operator are used for obtain two difference images,and then the weight parameters are set for the two difference images to generate the fusion difference image,so as to preserve the local consistency and edge detail information of the image.Subsequently,the saliency detection is performed on the two difference images,and the common salient region is mapped to the fusion difference image to obtain the salient region of the fusion difference image,thereby eliminating the noise interference in background region.Finally,taking the spatial semantic proximity relationship into account,FINCH clustering is performed on the saliency regions of the fusion difference image to further obtain the final change map.Experimental result shows that this approach is more robust to speckle noise and has better change detection accuracy.(2)Aiming at the problem of the poor accuracy of the clustering approach to distinguish between the change and the unchanged pixels,a SAR image change detection approach based on multi-difference images Convolutional-Wavelet Neural Network(CWNN)fusion is proposed.First,the log ratio and the subtraction operator are used for obtain two difference images.Second,the saliency detection is performed on the two difference images respectively,the salient region is extracted.And pre-classify the salient regions respectively,divide them into change,intermediate and unchanged classes.Then,select suitable training samples from the change and unchanged classes of the two difference images,use them for CWNN training respectively,and further classify the intermediate classes through the trained network to obtain the change detection results.Finally,the change information of multiple approaches is fused through voting to obtain the final change map.Experimental result shows that this approach can greatly improve the accuracy and anti-noise performance of change detection while learning deep features of the image.