HFA-Net: hierarchical fusion attention network for remote sensing image change detection

Abstract

Remote sensing image change detection plays a crucial role in fields such as urban planning and land use. However, existing methods often face challenges such as class imbalance, pseudo-change interference, and the difficulty of balancing computational efficiency with detection accuracy. To address these issues, this article proposes a hierarchical fusion attention network (HFA-Net). This network adopts a symmetric encoder–decoder architecture and integrates four key modules to enhance performance. First, we design the grouped multi-scale attention module to efficiently capture multi-scale context information within a single computational path and adaptively enhance key features. Second, we construct a multilevel feature fusion mechanism, including the efficient residual feature interaction module for early feature enhancement, the adjacent-scale feature aggregation module to fuse deep semantic and shallow details within a single temporal image, and the shared channel–spatial feature attention module to dynamically fuse bitemporal features using shared weights during the decoding stage. These modules work together to precisely suppress pseudo-changes, enhance real change features, and effectively balance computational efficiency and representation capability of the model. Experimental results on three public datasets show that HFA-Net achieves F1-scores of 92.13%, 83.41%, and 86.21% on the LEVIR-CD, SYSU-CD, and GZ-CD datasets, respectively. Compared with the best-performing competing method on each dataset, HFA-Net improves the F1-score by 0.89, 0.44, and 0.06 percentage points and the Intersection over Union by 1.52, 0.62, and 0.07 percentage points, respectively. These results demonstrate the effectiveness and competitive performance of HFA-Net for accurate and robust change detection in complex scenes.