Change Detection In Remote Sensing Images Based On Difference Representation Learning

The rapid development of earth observation technology in recent years has made it easier to obtain high-quality remote sensing data.Change detection and interpretation analysis based on these massive data has become an important means to dynamically monitor the earth’s environmental changes.Change detection has been widely used in many fields such as environmental monitoring,disaster assessment,urban expansion,and resource management.The key to change detection is how to better represent the difference so as to highlight the difference.Therefore,combining the data characteristics of different remote sensing images,this thesis focuses on how to make full use of the data to mine the differential feature representation that is conducive to the change detection task.The specific content includes the following aspects:(1)Detecting various types of changes in ground objects from multi-temporal remote sensing images enables more comprehensive monitoring of the surface,which requires good discrimination of different ground objects.For this task,this thesis proposes a multiple change detection method based on the generative adversarial network.It uses a generative adversarial network to perform unsupervised distribution fitting on bi-temporal images,and the trained discriminator network can be used to extract feature representations that can fully represent the discriminative features of the image itself.Finally,by combining classifiers and fine-tuning training with a few samples,end-to-end multi-class change detection can be performed.Experimental results show that the discriminative features extracted by the method can better distinguish different ground objects and improve detection performance.(2)Many object-based change detection techniques only consider the characteristics of objects themselves,ignoring the correlation between objects.To solve this problem,this thesis proposes a change detection method based on a multi-scale graph attention network.It uses the superpixel segmentation technique to divide the homogeneous region and constructs the superpixel graphs according to the adjacency relationship between the superpixel nodes.Then,the graph attention network is used to synthesize the different contributions of the adjacent nodes to obtain the feature representation of each node.Finally,by classifying the node features,the change category of the coverage area of each node can be obtained.Experimental results show that the proposed method can better process irregular object data and obtain better feature representation,thus improving detection efficiency and accuracy.(3)For hyperspectral image change detection,in addition to extracting effective difference feature representations,it is also necessary to use spatial-spectral features more accurately.To address these issues,this thesis proposes a principal component-guided self-supervised feature extraction and spatial-spectral attention joint network.First,a self-supervised mapping from the difference image patches to the principal components of the central pixels of these patches is established,thereby extracting the main spatial features of the differences.Then,it introduces adaptive attention weighting factors for the spatial-spectral joint features of each pixel.Experiments show that the proposed method can capture the main difference feature representations,and adaptively adjust the joint ratio of spatial and spectral features to improve the detection performance.(4)Due to a large number of spectral bands in hyperspectral images,hyperspectral imagebased change detection methods are susceptible to interference from a large amount of irrelevant or noisy spectral and spatial information.To address these issues,this thesis proposes a spectral and Gaussian spatial attention network.It can adaptively enhance useful information for detecting changes and suppress irrelevant noise and redundant information through learning from both spectral and spatial perspectives.Especially,the closer the pixels are,the higher their similarity or correlation,so this thesis designs a spatial attention mechanism based on Gaussian distribution.It builds a Gaussian distribution based on the learned distribution parameters and controls the correlation between pixels by sampling the distribution.Experiments show that the proposed network can capture changed regions more accurately while reducing the sensitivity of patch-based methods to patch size.(5)To learn efficient visual representations that are beneficial for fine-grained change detection without annotations,this thesis proposes a self-supervised global-local contrastive learning framework.The method first performs data augmentation on bi-temporal images,followed by a change detection backbone network as a feature extractor.Then,image-level and pixel-level instance discrimination tasks are performed on the extracted features through the global and local contrastive learning modules,respectively,which can improve the discriminativeness of different instances from global and local perspectives.The experimental results show that the model pre-trained by the proposed framework can more accurately distinguish different objects and improve the performance of fine-grained change detection tasks.