Change Detection Technology Of Heterogeneous Remote Sensing Images Based On Style Transfer

Remote sensing image change detection is an important means for earth observation.Its goal is to determine the changing area in the same geographical area through observation and analysis of two remote sensing images acquired at different times.It has been widely used in many earth observation scenes such as environmental-climate monitoring,urban planning research,and natural disaster assessment.Change detection techniques can be divided into homogenous image change detection and heterogeneous image change detection according to whether the image source is the same.Since the homogenous image change detection technique is only applied to a specific single type of remote sensing images,it does not applicable to other types of remote sensing images.Moreover,in some emergencies（such as typhoons and earthquakes,etc.）,due to the limitations of weather conditions,it is difficult for us to obtain homogenous images before and after the occurrence of the event in time,which has great limitations in the practical application of the homogenous image change detection technique.Therefore,how to use different sources of image information for change detection has very important practical significance.Different acquisition sources mean that heterogeneous images have different data domains,different statistical characteristics,and different image representations.Combining heterogeneous image information on the one hand makes full use of the functional characteristics of different sensors,but on the other hand,presents additional technical challenges.Therefore,with style transfer as the core idea,this thesis carries out research on change detection methods for heterogeneous remote sensing images,providing an effective scheme for the application of remote sensing to earth observation.The main research contents of this thesis are as follows:（1）To solve the problem of data discrepancy in the change detection from heterogeneous images,this thesis introduces the image style transfer to perform a homogenous transformation to eliminate the difference in style form and statistical characteristics of heterogeneous images.Since most existing style transfer methods only constrain the distribution of images in feature space to achieve style transfer,directly applying these methods to remote sensing images can not satisfy the homology property,while style transfer using adversarial networks can effectively learn the data distribution of images,but its training process requires sufficient image data.To address these problems,an improved image style transfer method is proposed.Based on the neural style transfer,combined with the adversarial networks,the distribution of images in the feature space and data space is constrained simultaneously in the process of style transfer,and the style transfer of remote sensing images can be realized with only two images.The experimental results on change detection show that the proposed method can effectively eliminate the differences between heterogeneous images and thus improve the accuracy of subsequent change detection.（2）Most of the existing change detection methods are usually implemented with the help of certain label information,which limits the application of these methods due to the human and material resources required to obtain the label information,so how to achieve unsupervised change detection is a challenging and practical problem.Aiming at this requirement,this thesis proposes a heterogenous image change detection method based on bidirectional domain transfer.Firstly,the convolution denoising autoencoder is used to extract the domain feature information of heterogeneous images,and the cycle consistency idea is used to carry out the bidirectional transfer between domain features.The relationship between different domains is mined by cross-domain fusion,and the difference map is obtained by combining the different information on the two domains.On this basis,the clustering algorithm and sample selection strategy are combined to perform change analysis on the difference map,to obtain reliable training samples and further improve the performance of change detection by training a robust classifier.The experimental results show that the validity of the proposed method is verified by obtaining satisfactory detection results on several datasets.