A Study On High-Resolution Remote Sensing Image Change Detection Method Based On Dual-perspective Change Context And Semi-supervised Learning

Remote sensing image change detection refers to the analysis of land cover change information by multiple observations of a specific area.Change detection has a wide range of applications in land use management,urban expansion,disaster analysis,and other fields.With the rapid development of Earth observation technology,the ability to obtain high-quality images of the same area using high-resolution sensors has significantly improved,providing data support for change detection technology.The emergence of deep learning technology has further accelerated the development of change detection research.Currently,deep learning-based change detection methods have achieved the best performance in many open datasets for change detection.However,most deep learning-based methods rely on a large number of change labels,and the high cost of annotating training samples seriously hinders the application of change detection in practical scenarios.This paper conducts in-depth research on highprecision change information interpretation and change detection methods under limited annotated sample conditions,based on existing research.The main research contents include the following two aspects:(1)To address the insufficient mining of temporal information and the complexity of contextual information in high-resolution remote sensing images in change detection tasks,this paper proposes a change detection method based on a dual-perspective change context network.The method emphasizes the extraction and optimization of change features through bi-temporal feature fusion and context modeling.First,a new dual-perspective fusion module is proposed to extract change features,which explores change information from the perspective of each time phase and improves sensitivity to change-related information by calculating the change attention between the two time phases.Then,a change context module is proposed,which is a relational context method that considers the correlation between change pixels and corresponding change regions.The module uses the representation of change context to enhance the representation of change pixels and promote the integrity of change targets.The combination of the dual-perspective fusion module and the change context module maintains the balance between precision and recall,achieving high-precision change interpretation in high-resolution remote sensing images.(2)To address the problem of limited change labels and high training costs,this article investigates semi-supervised approaches for change detection based on consistency learning.Specifically,two task scenarios are considered: typical semisupervised change detection and semi-supervised change detection with single temporal semantic information.For these two scenarios,the Temporal Consistency Network(TCNet)and the Multi-Task Consistency Network with single temporal supervision(MTCNet)are proposed,respectively,to effectively alleviate the dependence on change labels.In the semi-supervised change detection method based on the Temporal Consistency Network,a data-level perturbation that conforms to the characteristics of change detection is designed to learn unlabeled data by promoting the consistency of predicted results from different input sequences and fully exploiting change information in unlabeled data.As for the semi-supervised change detection method with the Multi-Task Consistency Network,it takes advantage of the existing single-temporal semantic labels.A task-level consistency learning strategy is proposed to indirectly optimize the change branch prediction results by minimizing the difference between the predicted results of the two T1 branches,thereby improving the generalization ability of the change detection model.This method can use singletemporal semantic labels as a low-cost substitute for change labels,thereby reducing model training costs.