High-resolution Remote Sensing Images Change Detection Based On Domain Adaptation

Change detection(CD)aims to identify significant differences in geographical elements between bi-temporal images of the same geographic area.It takes registered bi-temporal images as the input and outputs pixel-wise change maps,which has wide applications in urbanization monitoring,resource and environment monitoring,disaster assessment,etc.Recently,many excellent methods have been proposed due to the easy acquisition of high-resolution remote sensing images and the rapid development of deep learning.Because of the powerful ability of automatic high-level feature extraction,they have demonstrated superior performance and robustness to conventional methods.This fundamental but important remote sensing task has gradually become an active topic in the computer vision community.Aiming at the problem of domain shift existing in change detection tasks,this paper studies the cross-domain change detection method of high-resolution remote sensing image based on domain adaptation from the perspectives of supervised learning and unsupervised learning,which mainly includes the following two aspects of research content and innovation.On the one hand,since the bitemporal images input into the change detection network is acquired at different times,the appearance of the geographical elements changes with external factors such as sensors,atmospheric conditions,lighting conditions,and seasons,bringing severe domain shift between bitemporal images.However,most existing deep learning-based works try to elaborately design complicated neural networks with powerful feature representations,but ignore the pseudo-changes brought from the domain shift,resulting in suboptimal results.To solve this problem,this paper combines image adaptation and feature adaptation to propose an end-to-end cross-domain change detection framework.The image adaptation module exploits generative adversarial learning with cycleconsistency constraints to perform cross-domain style transformation,which effectively narrows the domain gap in a two-side generation fashion.In the Feature Adaptation Module,this paper further reduces the domain difference by extracting domain-invariant features to align different feature distributions on feature space.On the other hand,because that change detection is a pixel-level segmentation task,it usually takes a lot of effort and time to obtain accurate ground truth,and it may not be possible to obtain enough label data for training deep networks in practical applications.Moreover,due to external factors such as sensors,taken conditions,lighting conditions,and seasons,there are often large differences in domain distribution between two different datasets,and directly applying the trained model to new data will inevitably bring about a significant decrease in detection performance.To solve this problem,this paper proposes a cross-domain change detection framework based on unsupervised domain adaptation,introduces copy-paste strategy to construct hybrid intermediate domains of source domain and target domain,and uses mean teacher framework to align domain features and improve the detection effect of the network on the target domain.For supervised cross-domain detection framework,extensive experiments and analyses based on three different change detection methods on CDD and WHU building datasets demonstrate the effectiveness and generalizability of our proposed framework.The framework of this paper can be combined with any change detection network to improve its ability to handle cross-domain problems and enhance the detection performance.Notably,our framework pushes several representative baseline models up to new State-Of-The-Art records,achieving 97.34% and 92.36% on the CDD and WHU building datasets,respectively.For unsupervised cross-domain detection framework,in this paper,experiments are conducted on two unsupervised cross-domain tasks,WHU building->LEVIR-CD and LEVIR-CD->WHU building,based on three different change detection methods,respectively,and the experimental results demonstrate that the framework in this paper can be combined with arbitrary change detection networks to improve the ability of the network to handle unsupervised cross-domain change detection problems.