AGFormer: An anchor-guided transformer for class imbalance in remote sensing change detection

Abstract

Remote Sensing Change Detection (RSCD) aims to assess changes by comparing two or more images recorded for the same area but taken at different time stamps. Mainstream research improves the representation of models through the optimization of model architecture design, ignoring the importance of correcting classifiers. However, the issue of class imbalance in the RSCD field inevitably introduces biases into the classifier, damaging the model performance. In this paper, we propose an Anchor-Guided transFormer-based model, named AGFormer, to address this problem. Specifically, the HAR (Hypersphere Anchor Regularization) calibrates the classification layer from an anchor view, which ensures both inter-class separability and intra-class balance between compactness and diversity by initializing class anchors on the hypersphere and applying similarity-based contrastive learning in different phases. In addition, a disentanglement anchor optimization strategy is designed to avoid the influence of class imbalance in the RSCD field. By supervising the main features and calibrating classifiers with mapped class anchors, more discriminative representations and robust classifiers are obtained. In addition, we design the CEM (Change Enhancement Module) based on flow to highlight the changed features. The proposed HAR and CEM are plug-and-play and can be integrated into existing architectures. Extensive experiments are conducted on four benchmark datasets, and state-of-the-art performance is achieved by the proposed AGFormer. All the codes are available at https://github.com/jiaenchen2024/AGFormer.