| Human activities and frequent natural disasters have accelerated the change of the earth’s surface,and remote sensing image change detection has become the main solution to effectively detect large-scale surface changes.Change detection is to analyze the remote sensing images of the same area obtained in different periods to determine the man-made or natural changes on the surface.Change detection has been applied in many important fields,such as environmental protection,disaster assessment and military reconnaissance.With the continuous progress of earth observation technology,it is more convenient to obtain satellite remote sensing images with different resolutions and different types.Synthetic aperture radar(SAR)image has attracted extensive attention because of its all-weather,all-time working ability and strong penetration.However,the existence of speckle noise during imaging will seriously impact on the accuracy of change detection results.Therefore,how to effectively remove noise and correctly identify the changing area is the goal that needs to be paid attention to and solved.With the development of computer technology,the technology related to computational intelligence has been greatly developed and plays a significant role in image change detection.Based on the existing change detection technology and computational intelligence theory,this paper studies the change detection of SAR image.The specific innovative research work included in this paper is described as follows:(1)A decomposition-based bi-objective fuzzy clustering method for change detection in SAR Images is proposed.Change detection for SAR images is formulated as a bi-objectives fuzzy clustering problem from the aspects of preserving the detail and removing the noise.The log-mean ratio method is used to generate the first difference image to improve the extraction ability of the original detail information.The second difference image is obtained by combining the homomorphic filtering and saliency detection,which effectively removes the speckle noise.Then the fuzzy clustering objective functions are constructed for two difference images to classify the pixels under different requirements.Finally,a new membership updating method is proposed to comprehensively analyze the change detection results under different objectives,which balance the influence of the two objectives on the final detection result.The proposed method improves the insufficient analysis of difference image in the existing methods and improves the robustness of the algorithm to noise.The experimental results on real SAR datasets show that the proposed method has excellent detection performance.(2)A fuzzy clustering-based self-supervised neural network change detection method is proposed.Firstly,the fuzzy local information clustering algorithm is used as a pretext task to obtain the pseudo label of each pixel.The pixels are pre-classified into three categories: changed,unchanged and uncertain classes through the membership degree threshold.Then,the pixel blocks with the same window size at the same position in the two SAR images are spliced,and the spliced pixel blocks with pseudo labels of change and invariance are selected as the initial training samples.Data augmentation is used to expand the training data and enhance its generalization ability.Finally,the convolution neural network is used to construct the training model.The uncertain categories are further classified by extracting the characteristics of changed class and unchanged class.The final change detection results are obtained by refining the classification results of the neural network.The proposed method effectively improves the ability of image feature extraction,reduces the dependence on prior information,and has strong adaptability. |
