HWA-Net: Hierarchical window aggregate network for cross-resolution remote sensing change detection

Abstract

Cross-resolution remote sensing change detection (CD) is a critical task in various applications, including urban monitoring, environmental changes, and disaster management, where images captured at different times often possess varying spatial resolutions. Current methods typically address this by resampling low-resolution (LR) images to high-resolution (HR) formats, but such image-level strategies lead to significant artifacts and misalignment in the change map. These imperfections not only reduce detection accuracy but also lead to misleading or false change identifications, resulting in incorrect or incomplete conclusions in time-sensitive applications, such as land-use change detection or disaster monitoring. To address these challenges, we propose the Hierarchical Window Aggregate Network(HWA-Net), a novel framework that directly operates on cross-resolution image pairs without preprocessing, aiming to accurately aggregate cross-resolution representations for robust CD. HWA-Net initially employed window-based feature extraction to produce scale-independent representations, subsequently transferring these features to layered decoding. This process effectively enhances detection accuracy across diverse resolutions. Our approach establishes new state-of-the-art results on three synthesized datasets and one real-world cross-resolution change detection dataset.