TALS-CD: Causal Continuity-Based Task-Aware Mamba for Remote Sensing Binary Change Detection

Scan strategies are critical for Mamba-based visual tasks, yet predefined hand-crafted scan paths often fail when applied to task-specific data. In change detection, irregular discontinuities between changed and unchanged regions disrupt spatial continuity, causing existing scan methods to increase effective dependency lengths and reduce the causal effectiveness of Mamba-based change feature representations. To address these challenges, a task-aware learnable scan (TALS) strategy is proposed for Mamba-based change detection (TALS-CD). TALS learns causal correlations between local windows and establishes an inter-window priority sequence for bi-temporal difference feature representation learning. With TALS, Mamba first models the global dependencies of difference features and then progressively refines them in a top-down manner, leveraging the local properties of CNNs. Additionally, a Mixture of Difference Experts (MoDE) is introduced to utilize complementary difference representations, dynamically selecting optimal routing paths to enhance difference feature expression. Experimental results on three public datasets demonstrate that TALS-CD achieves superior change detection performance with a small number of parameters and computational cost. The code is available at: https://github.com/fangcunxinghe/TALS-CD.