Mamba-LCD: Robust Urban Change Detection in Low-Light Remote Sensing Images

Remote-sensing-based urban change detection (CD) has seen substantial progress with the development of deep learning. Nevertheless, its performance under low-light imaging conditions, such as nighttime, twilight, and cloudy or rainy weather, remains limited due to image degradation, reduced contract, and weak spatial features. To address these challenges, we propose Mamba-LCD, a novel CD method that integrates a Siamese visual state-space encoder, a bitemporal feature fusion module, and a mask decoder. The encoder captures spatial dependencies via multidirectional state-space scanning, while the fusion module enhances semantic consistency through illumination-aware recalibration. The decoder integrates multiscale features using attention-inspired pooling for precise pixel-level change masks. Comprehensive experiments on three benchmark datasets including S2Looking, WHU-CD, and LEVIR-CD demonstrate the effectiveness of Mamba-LCD. Specifically, on the low-light-dominated S2Looking dataset, Mamba-LCD outperforms ten state-of-the-art convolutional neural network-, Transformer-, and Mamba-based methods, surpassing the previous best results by 0.73 and 1.58, respectively. Moreover, it maintains competitive performance on LEVIR-CD and WHU-CD under normal lighting. We further conducted ablation studies to evaluate the impact of different activation functions and fusion strategies. The results suggest the robustness and adaptability of Mamba-LCD, indicating its potential for urban monitoring across diverse illumination scenarios.