Research On Building Change Detection Based On Siamese Network

The investigation and monitoring of natural resources is the basis of realizing the planning,utilization,protection and restoration of natural resources.The change detection of buildings has always been an important monitoring object and content in the field of natural resources investigation and monitoring.The change and development of buildings are closely related to human economic development and social life.Therefore,it is of great significance to obtain building change data quickly and efficiently for urban development,spatial planning,population and industrial distribution.At present,building change detection mainly relies on manual visual interpretation of change information,and its efficiency and accuracy greatly limit the extraction of building change information.In recent years,deep learning technology is widely used in remote sensing image automatic recognition and extraction.Research shows that the change detection algorithm based on deep learning can effectively identify the change region.Therefore,this paper studies the application of deep learning technology in building change detection,and the main research contents include the following three aspects(1)The construction of Spatial – Temporary Attention Neural network(stanet)is deeply studied.The network is mainly constructed by Siamese Network,which can detect the change of two periods of remote sensing images.The feature extraction layer of the network adopts the deep residual network(RESNET).Before comparing the differences between the two feature vectors,the proposed features are enhanced by self attention mechanism and pyramid spatiotemporal attention,The whole network can improve the network prediction ability by calculating the weight of different spatialtemporal characteristics and the spatial-temporal dependence of different scales.(2)Combing the existing requirements of natural resource change detection,this paper designs and integrates multi space operators to process the change detection and prediction results,mainly including: pattern vectorization,pattern simplification and pattern regularization operators.(3)Based on the aerial remote sensing data of a district in Jiangsu Province in2016 and 2019,a set of building change detection data set is made,and a set of building change detection model is trained on the basis of the data set.The optimal F1 coefficient of the model is 0.787,and the MIo U coefficient is 0.764.This paper studies the building change detection method for natural resources survey and monitoring,designs and implements the building change detection application system.The main research work includes the research of Siamese network change detection method integrating multi spatial operators and the design and implementation of building change detection system.The innovation of this paper is to improve the input and output part of the Siamese network model,integrate the building change detection sample enhancement method and multi space operator,and develop the traditional "pixel level" image change detection to "spot level" image change detection,which provides a method reference for the further development of intelligent monitoring application of natural resources big data.