A Research On Change Detection Method Of Remote Sensing Image Based On Generative Adversarial Network

Remote sensing image change detection is to compare and analyze the images of different periods in the same area,and obtain the characteristic differences of the state before and after the surface.It is widely used in urban planning,disaster assessment and environmental monitoring.With the advancement of technology and the deepening of application,the number and types of remote sensing images are increasing.Traditional change detection methods have been unable to accurately and efficiently process the change information of complex images.In recent years,the powerful autonomous learning ability and representation ability of deep learning technology have also brought new ideas to the research of change detection methods.Among them,although the commonly used convolutional neural network has achieved good application results,this model is a discriminant model,and the accuracy of the algorithm depends heavily on the label data,which requires a lot of manpower and material resources to make label samples.At the same time,because change detection requires images of different periods,the collected data will inevitably be affected by sensors and the external environment,resulting in a great difference in the imaging style of the image,thus affecting the accuracy of change detection.In view of the above problems,this paper improves the remote sensing image change detection method based on generative adversarial network.The generative adversarial network belongs to the generative model,which mainly learns the feature distribution between data,which can reduce the dependence on the label data set,and the adversarial learning can transfer images from different image domains to the same image domain through style transformation.Therefore,this paper will study the change detection method of remote sensing image from the following aspects.(1)In order to deeply understand the field of remote sensing image change detection,this paper combs and summarizes the process and development process of change detection,and classifies the current change detection methods based on the research status at home and abroad.The characteristics of various methods and the current problems are analyzed and summarized.Two improved remote sensing image change detection models based on generative adversarial networks are proposed.(2)This chapter proposes an improved semi-supervised feature extraction remote sensing image change detection model based on generative adversarial networks.Aiming at the problem that the change detection method of supervised learning method depends heavily on the label image with change information,and the production of label map requires a lot of manpower and material resources,and affects the detection efficiency,a semi-supervised change detection model based on generative adversarial network is designed,which uses U~2-Net to realize feature extraction,so as to reduce the dependence on label data set and ensure the ability of the model to extract change information.The experimental results show that the method can also ensure the detection performance of the model with a small number of ground truth images.(3)A remote sensing image change detection model based on unsupervised learning of generative adversarial networks is proposed.This method uses the improved image style transfer network FAM-Cycle GAN to preprocess the image,so as to reduce the influence of image style difference on change detection.In order to get rid of the dependence on the label data set,based on the semi-supervised learning change detection model,the traditional change detection method PCAK-means is used to obtain the initial detection result map,and then the contour coefficient is used to select the pseudo label data for training the change detection network.Through comparative experiments on public datasets,it is verified that the FAM-Cycle GAN network helps to improve the performance of change detection.The appropriate contour coefficient parameters are selected,and the pseudo-label images are selected to train the change detection model.The experimental results show that the method has good detection performance..