Research On Spatial Neighborhood Representation And Dual-branch Deep Network Model For Remote Sensing Image Change Detection

Remote sensing image change detection is of great significance to social development and environmental protection,and has become a research hotspot at home and abroad.As the resolution of remote sensing images increases,the images show many new characteristics,which brings many new challenges to change detection.Given that the existing pixel-level,object-level,and deep learning-based change detection methods have their own characteristics and limitations,targeted further research is of great significance for change detection.To this end,based on the summary and analysis of the existing change detection methods,this paper studies how effectively use the various information in the image to optimize the detection results from the pixel-level change detection and the change detection based on deep learning.The main research contents of this paper are as follows:(1)Analyze the characteristics of the change detection method.This article analyzes the characteristics of object-level and pixel-level change detection methods,and summarizes the generation method of neighborhood information and its application in traditional change detection methods(pixel-level and object-level change detection methods).It clarifies the effectiveness of neighborhood information in optimizing detection results,laying a foundation for exploring new pixel-level change detection methods.At the same time,by analyzing the relationship between deep learning and change detection,summarize the characteristics of the change detection model based on deep learning,and provide suggestions for constructing new network models.(2)The construction combines spatial neighborhood expression and structural feature change detection.This method first introduces a new similarity measure-matching error,and then combines it with neighborhood-related images that are sensitive to spectral differences to provide more robust neighborhood information.At the same time,structural features that are robust to spectral differences are introduced to make full use of the various information of the image.Finally,the above three features are used as classification attributes to obtain the final binary change map through decision tree classification and Markov optimization.The results showed that this method weakened the "salt and pepper phenomenon" in the test results,and obtained better test results.(3)This paper proposes a change detection method based on the body and edge dual branch network.This method is based on the characteristics of strong similarity between the pixels in the main part of the object and weak similarity in the edge part.The dual-branch structure is used to optimize the main body and edge of the feature,and by introducing the idea of multiple supervision,the calculated main body and edge,The loss between the prediction result and the tag is transmitted to the network,the network parameters are updated,and the accurate optimization of the main body and edge of the feature is realized.The whole detection model is mainly composed of three parts: feature extraction,feature decomposition,and feature optimization and reorganization.The results show that the method can accurately identify the change boundary.