GCCANet: A Small-Object Sensitive and Environment-Robust Network for Remote Sensing Change Detection

Abstract

Remote sensing change detection (RSCD), widely used in industrial and military applications, aims to accurately identify surface changes by comparing two temporally separated images of the same region. In recent years, deep learning has greatly advanced the development of RSCD. However, in complex environments, subtle and background-induced changes often compromise the accuracy of change detection (CD). To address these challenges, we propose the global-context and change-aware network (GCCANet). We adopt a dual-encoder architecture to extract rich semantic features and design an adaptive layer attention (ALA) module that condenses multilevel Transformer semantics into compact global guidance. We further design a background suppression module (BSM) to inject this global guidance into each convolutional neural network (CNN) stage, effectively suppressing background noise and stabilizing the subsequent differencing process. Moreover, we design a difference learning module (DLM) that couples local differencing with global attention matching to extract robust and detail-aware change cues. Then, we introduce a lightweight decoder equipped with a multiscale small-object enhancement (MS-SOE) strategy to enhance subtle changes and preserve thin structures. Extensive experiments on five public benchmark datasets demonstrate that GCCANet consistently achieves state-of-the-art (SOTA) performance. The code is available at https://github.com/E1ison/GCCANet