CGA-Net: A CNN-GAT Aggregation Network Based on Metric for Change Detection in Remote Sensing

Change detection aims to reveal the changes of specific regions or objects in a time series. Object-level change detection methods are more suitable for existing needs because they can locate and identify changed objects more accurately. Attempting to solve the problems in existing object-level change detection methods, such as ignoring the relationship between dual-branch features, insufficient utilization of feature point information, and unreasonable fusion weight allocation mechanism, this article proposes an object-level change detection network, CGA-Net, based on metric, which combines similarity measurement with the feature extraction and fusion. Specifically, the feature exchange module (FEM) exchanges the feature point information of the dual branches under the guidance of cosine similarity; the feature aggregation module driven by graph attention network (GAT) aggregates features locally and globally using cosine similarity and GAT; the dual-time feature fusion module assigns weights to different parts for fusion based on feature similarity and correlation. The experimental results show that CGA-Net exhibits excellent performance on the LEVIR-CD and WHU-CD datasets. Our method achieves 94.20% and 86.85% in mAP@0.5 on the two datasets, respectively, and 75.12% and 77.39% in mAP@0.5:0.95, respectively, which has a significant improvement compared to both the algorithms based on bounding boxes and the algorithms based on masks. It is fully proven that CGA-Net can effectively improve the accuracy of change detection in different scenarios.