Abstract
© 2025 The AuthorsMulti-temporal Semantic change detection (MT-SCD) provides crucial information for a wide variety of applications, including land use monitoring, urban planning, and sustainable development. However, previous deep learning-based SCD approaches exhibit limitations in time-series semantic change analysis, particularly in understanding Earth surface change dynamics. Specifically, literature methods typically employ Siamese networks to exploit the multi-temporal information. This hinders temporal interactions, failing to comprehensively model spatio-temporal dependencies, causing substantial classification and detection errors in complex scenes. Another key issue is the neglect of temporal transitivity consistency, resulting in predictions that contradict the multi-temporal change chain rules inherent to MT-SCD. Furthermore, literature approaches do not consider dynamic adaptation to the number of observation dates, failing to process time-series remote sensing images (RSIs) with arbitrary time steps. To address these challenges, we propose a graph-enhanced spatio-temporal Mamba (GSTM-SCD) for MT-SCD (including both bi-temporal SCD and time-series SCD). It employs vision state space models to capture the spatio-temporal dependencies in multi-temporal RSIs, and leverages graph modeling to enhance inter-temporal dependencies. First, we employ a single-branch Mamba encoder to efficiently exploit multi-temporal semantics and construct a spatio-temporal graph optimization mechanism to facilitate interactions between multi-temporal RSIs, while maintaining spatial continuity of feature representations. Second, we introduce a bidirectional three-dimensional change scanning strategy to learn underlying semantic change patterns. Finally, a novel loss function tailored for time-series SCD is proposed, which regularizes the multi-temporal topological relationships within data. The resulting approach, GSTM-SCD, demonstrates significant accuracy improvements compared to the state-of-the-art (SOTA) methods. Experiments conducted on four open benchmark datasets (SECOND, Landsat-SCD, WUSU and DynamicEarthNet) demonstrate that our method surpasses the SOTA by 0.53%, 1.66%, 9.32% and 0.78% in SeK, respectively. Moreover, it significantly reduces computational costs in comparison with recent SOTA methods. The associated codes is made available at: https://github.com/liuxuanguang/GSTM-SCD.
NETL
NSTL
Elsevier