EDG-Net: A Lightweight Frequency-Aware CNN-Transformer Hybrid Network for Efficient Remote Sensing Change Detection

Abstract

Accurate change detection (CD) in high-resolution remote sensing imagery is often affected by “pseudochange” interference (e.g., seasonal and illumination variations) and by the computational constraints of edge devices. To address these challenges, we propose the efficient difference-gated network (EDG-Net), a lightweight frequency-aware convolutional neural network (CNN)-Transformer hybrid architecture that strategically couples convolutional feature extraction with efficient attention-based aggregation. The hybrid design employs CNNs for local frequency-aware encoding and transformer-style linear attention for global context modeling, forming a coupled pipeline that combines frequency-domain feature recalibration, temporal evidence interaction, and multiscale aggregation. First, a Siamese frequency-aware encoder (CNN-based) equipped with a discrete cosine transform mechanism is used to emphasize structure-sensitive spectral responses while weakening features dominated by environmental variation. Second, a difference-correlation interaction module with linear attention jointly models temporal differences and structural consistency to reduce pseudochange responses. Finally, a gated multiscale attention aggregation module adaptively fuses global context and fine-grained boundary cues with limited computational overhead. Extensive experiments across five high-resolution benchmarks (LEVIR-CD, WHU-CD, SYSU-CD, LEVIR-CD+, and CDD-CD) validate the effectiveness and robustness of the proposed design. With a computational cost of 7.35 GFLOPs, EDG-Net attains an F1-score of 83.50$\pm$0.27% (mean$\pm$std over 5 random seeds) and an intersection over union of 71.67$\pm$0.40% on the challenging SYSU-CD dataset, while maintaining a favorable accuracy-efficiency tradeoff against recent transformer- and Mamba-based models. These results indicate that EDG-Net is a practical lightweight solution for robust remote sensing CD.