TA-VSSM: Topology-Aware Visual State Space Model for Road Change Detection

Abstract

Road change detection (RCD) in remote sensing imagery is inherently challenging due to the elongated geometry, complex topology, and strong directionality of road networks, which often result in fragmented predictions and inconsistent structures. To address these issues, we propose a topology-aware visual state space model (TA-VSSM) that reformulates RCD as a structured reconstruction problem rather than conventional pixelwise classification. Specifically, the proposed framework integrates geometry-aligned feature extraction and omnidirectional sequence modeling to better capture the intrinsic characteristics of road structures. By introducing directional inductive bias and extending state-space propagation beyond conventional axial scanning, the model is able to effectively model long-range dependencies along arbitrary orientations. Furthermore, a topology-preserving decoding strategy, together with a hybrid supervision scheme, is employed to ensure structural continuity and accurate boundary delineation in the final predictions. Extensive experiments on the WRCD and CRCD datasets demonstrate that TA-VSSM achieves superior performances against the state-of-the-art methods.