Research On Synthetic Aperture Radar Image Change Detection Methods

Change detection(CD)is an important research direction of remote sensing image interpretation,and is widely used in military and civilian fields,such as military survey,disaster monitoring and evaluation,urban construction planning,and agricultural detection.Synthetic aperture radar(SAR)has the advantages of all-weather and all-day operation due to its special microwave imaging mechanism.Therefore,it can be imaged in extremely harsh environment,and will not be affected by clouds and other interferences.Due to its advantages,more and more scholars and scientists begin to pay attention to SAR image change detection.However,the speckle noise caused by microwave imaging greatly affects the accuracy of change detection results.The complex ground object background also brings great challenges to the detection.In view of the above problems,this paper has made an in-depth study on the traditional methods and deep learning methods of change detection in multi-temporal SAR images.The following is the specific work of this paper.(1)Aiming at the problem that SAR images are seriously affected by speckle noise and the existing algorithms have incomplete information utilization in difference image construction,a SAR image change detection method based on saliency ratio-mean ratio with multi-scale morphological reconstruction fuzzy c-means(SRMR-MSMRFCM)is proposed.Firstly,traditional ratio difference image and mean ratio difference image are fused by multiplication to preliminarily eliminate image noise and enhance the characteristics of changed and unchanged areas.Secondly,saliency detection and binarization are performed on the fusion difference image to obtain roughly changed unchanged areas.Thirdly,according to the saliency image,the difference image of the corresponding areas is processed by region.The multi-scale morphological reconstruction is performed for the changed area,and the multi-scale image is obtained by bilinear interpolation method.Single-scale morphological reconstruction is performed on the unchanged area.Finally,by introducing morphological information into fuzzy c-means(FCM),the final detection results are obtained.Experimental results show that this method can suppress speckle noise in SAR images and achieve good accuracy.(2)Aiming at the problem of serious false label errors of training samples and incomplete information extraction in the deep learning method of SAR image change detection,a SAR image change detection method based on adaptive attention and convolution fusion network(AACFNet)under refined samples is proposed.Firstly,the fusion difference image is obtained by the multiplication operator and the binary saliency image is obtained.Secondly,the difference image is gamma corrected,and each pixel of the difference map is pre-divided into change class,intermediate class and unchanged class through FCM.Among them,changed and unchanged classes are used as training samples and labels for network training.Thirdly,the labels corresponding to non-saliency areas are all set as unchanged class to improve the accuracy of the sample.The training samples with obvious errors are further removed through threshold screening.Finally,the original images and the patches corresponding to the difference image before and after the change are sent to AACFNet for training as three channels,respectively.And the final detection result is output.The proposed AACFNet adaptively combines the features of self-attention and convolution extraction.In addition,dynamic convolution is used to better extract image features and fuse information of multiple scales.Experiments show that this method improves the accuracy of sample labels,and neural network can obtain more comprehensive and stable feature information.