DLT-Mamba: Unsupervised Change Detection by Fusing Latent Token Graph Alignment and Graph Interaction Mamba Segmentation

Unsupervised change detection (CD) remains an active yet highly challenging research direction in remote sensing. Domain shifts caused by environmental factors, such as illumination, seasonality, and atmospheric variations often severely degrade detection performance. Currently, most unsupervised methods still treat domain alignment and CD as separate paradigms, resulting in insufficient synergy between the two processes and underutilization of structural information in images. To address these limitations, we propose DLT-Mamba, an end-to-end unsupervised CD framework that fuses domain-alignment-centric and detection-centric paradigms within a single architecture. The framework integrates two tightly coupled components. The dynamic latent token graph convolutional network (DLTNet) uses a small set of learnable latent tokens as prototype nodes to enable efficient domain alignment through adaptive graph-based message passing and graph interaction Mamba (GIMamba) incorporates structural information into bidirectional state-space modeling, thereby achieving precise CD with enhanced boundary and structural awareness. Furthermore, we design a hybrid structure-chroma reconstruction loss function that supports joint optimization, ensuring the alignment process remains sensitive to structural changes while simultaneously promoting sparse and reliable predictions during fully unsupervised training. Extensive experiments on multiple widely used datasets demonstrate that the proposed method consistently achieves state-of-the-art performance and improves structural precision and robustness under domain-shift scenarios.