Research On SAR Image Change Detection Based On Feature Fusion

Synthetic aperture radar(SAR)which is not sensitive to sun illuminations and allweather condition and contains abundant surface information has become important data resources of remote sensing image interpretation.SAR image change detection aims at identifying change information in images obtained at the same scene but different times.It is in great demands in military and civil activities.However,the unique imaging mechanism of SAR leads to image distortion and products a large amount of speckles,which weakens the performance of image change detection algorithms.Therefore,studying the SAR image change detection algorithms have important theoretical significance and practical value.The traditional algorithms for SAR image change detection mainly use single difference image features to obtain change information,which makes it difficult to achieve high-precision change detection results in complex environments.Therefore,the dissertation mainly studies the SAR image change detection algorithms based on feature fusion.Two SAR change detection algorithms based on feature fusion are proposed for enhancing the anti-noise performance,detail retention ability and achieving the effective feature fusion.The main contents of the dissertation are as follows:(1)It is found that pixel classification is sensitive to speckle noise in SAR image change detection task.Therefore,the SAR image change detection algorithm based on image fusion and level set is proposed.In the proposed method,the intensity and texture difference images are obtained first.And the fusion difference image is built on the basis of stationary wavelet transform and local variance criterion.The fusion scheme proposed provides strong noise immunity as well as good preservation of edge locations of changed areas.Then,the KI threshold is implemented to pre-classify the difference map to obtain the local prior information of the fusion difference image.Finally,the local prior information is constructed into adaptive kernel function to introduce energy function of level set,which significantly improves the anti-coherent noise performance of the algorithm.Experimental results on real images demonstrate the effectiveness of the proposed algorithm and illustrate that it has both strong noise immunity and good preservation of edge locations of changed areas.(2)It is known that the detail information of difference image is difficult to be fully utilized in real tasks.Therefore,a multi-scale depth features fusion is proposed for SAR image change detection.The algorithm includes the following three steps.Firstly,stationary wavelet transform is used to decompose difference image into multi-scale image,which are independently reconstructed into difference images with good anti-noise performance and detail retention performance.Then,a fuzzy c-clustering algorithm that can realize multi-image classification is carried to obtain pseudo-labels for feature classifiers.And the reliable training samples are selected from the multi-scale difference images with voting rules.Based on the results of sample selection,the convolutional neural network is used to fuse the multi-scale depth features to realize the classification of the changed and the unchanged areas,which can further improve the ability of the algorithm to distinguish between samples,anti-noise performance and effective description of the integrity of changed regions.Experiments on real SAR image change detection demonstrate that the algorithm can greatly improve the accuracy and anti-noise performance.