Spectral-Attention Cooperative Encoding with Dynamic Activation for Remote Sensing Change Detection

Change detection (CD) in high-resolution remote sensing imagery is vital for resource monitoring and disaster assessment but faces challenges such as spatiotemporal heterogeneity, spectral variability, and computational inefficiency. This paper proposes an efficient CD method that hybridizes Convolutional Neural Networks (CNNs) and Transformers. A CNN backbone first extracts multi-level features from bi-temporal images. A Semantic Token Generator then compresses these features into compact, low-dimensional semantic tokens, reducing computational load. The core of our model is a novel cooperative encoder integrating a Spectral layer and an Attention layer. The Spectral layer enhances sensitivity to high-frequency components like edges and textures in the Fourier domain, while the Attention layer captures long-range semantic dependencies via self-attention. Furthermore, we introduce a Dynamic Tanh (DyT) module to replace conventional normalization layers, using learnable parameters to adaptively adjust activation thresholds, thereby improving training stability and computational efficiency. Comprehensive evaluations on the LEVIR-CD, WHU-CD, and DSIFN-CD benchmarks demonstrate that our method maintains high accuracy while reducing complexity, offering a practical solution for real-time CD in resource-limited environments.