Research On SAR Image Change Detection Method Based On Improved Difference Operator And Extreme Learning Machine

SAR image change detection is a technology that processes multi-temporal SAR images of a certain region or object,then evaluates whether the target is change and obtains the characteristics of the change.Since the electromagnetic wave emitted by SAR doesn't require a light source,it can penetrate clouds,fog and water vapor without distortion,and isn't affected by complex climates such as light and temperature,SAR can work in any climate,which makes the SAR image obtained through the SAR system naturally become the core data of change detection and is applied on a large scale.With the gradual maturity of technology,SAR image change detection has been commonly applied to farmland planning,natural disaster detection,geographic survey,water area monitoring,and military reconnaissance.The difference map acquisition and the difference map analysis are two light steps in the common process of SAR image change detection.The difference map with good quality is the prerequisite for obtaining high-precision change detection results,and the efficient difference map analysis method is the top priority to obtain high-precision change detection results.This thesis combines existing problems to research the difference map acquisition method and difference map analysis method in the change detection process,and proposes two SAR image change detection algorithms.The summary is as follows:(1)In order to solve the problems of change detection results being susceptible to the multiplicative noise generated by coherent accumulation,the small application range of the traditional single difference map acquisition method,and the single type of detection change,this thesis proposes change detection in SAR images on account of guided image filtering and improved difference operator.The algorithm uses single-channel guided image filtering to denoise the SAR image,and suppresses coherent speckles while ensuring the edge texture details;then combining the advantages of the neighborhood ratio operator and the improved logarithmic ratio operator,a method for obtaining difference map that can effectively suppress noise and unchanged information,enhance low pixel intensity,detect various types of changes and have higher contrast;finally obtain the final change detection through fuzzy local c-means(FLICM)method.The experimental data prove that the proposed difference operator can obtain a difference map that reflects the change of ground features and has good noise resistance,and the accuracy and consistency of the proposed detection algorithm are better.(2)In order to solve the problems of not extracting and not fully utilizing the feature information for instance gray scale and texture of difference map,and the detection precision can be improved,this thesis proposes change detection in SAR images based on Gabor and Extreme Learning Machine(ELM).The algorithm obtains Gabor multi-scale and multi-directional feature information from the difference image based on the improved difference operator and guided image filtering,and extracts pre-selected pixels from the Gabor feature vector of the pixel through the hierarchical FCM clustering method(more likely Change pixels and invariant pixels)and reclassified pixels;then use preselected pixels to find the pixels in the same position of the initial image and combine the neighbor pixels to form a sample to train ELM,and then rely on the ELM classifier to perform the final classification of the reclassified pixels;finally combined with pre-selected pixels to gain the ultimate change detection result.Data of experiment confirm that the algorithm makes full use of the grayscale and texture information of the difference map to further enhance the robustness and improve the detection accuracy.