Language-Guided Adaptive Spatial–Frequency-Aware Network for Remote Sensing Change Detection

Remote sensing change detection (RSCD) plays a crucial role in urban planning, disaster assessment, and ecological monitoring. However, the inherent complexity of remotely sensed images and the interference of non-targeted changes, such as illumination changes and ground renovation, often lead to the problem of misdetection of change areas by models. Moreover, many real-world changes occur at the semantic level rather than through explicit geometric or structural differences, making them difficult to distinguish based solely on visual appearance. Although deep learning-based methods have achieved significant progress in this field, most existing approaches struggle to effectively balance and address the complexities mentioned above due to insufficient semantic guidance. To this end, we propose a Language-Guided Adaptive Spatial-Frequency-Aware Network (LGSFNet) to enhance the model’s semantic guidance and discrimination ability for changing regions by introducing textual high-level semantic information at low cost. In this approach, textual prompts guide the model to focus more on change-related regions, thereby learning more discriminative features. Through disentangled fusion, this method provides a new perspective for traditional unimodal RSCD to utilize the high-level semantic information of text. Furthermore, we design a lightweight Dual-Domain Collaborative Sparse Attention mechanism in both spatial and frequency domains, which mimics human vision by adaptively emphasizing salient information. Combined with wavelet transforms, this mechanism further enhances the model’s sensitivity to complex change patterns, improving both discriminative power and robustness. Extensive experiments conducted on three widely used benchmarks—LEVIR, BCDD, and SYSU—demonstrate that LGSFNet achieves competitive performance compared with state-of-the-art methods.