<p>Remote sensing image change detection (CD) is a crucial technique for monitoring dynamic changes on the Earth’s surface, with broad applications in disaster assessment, land-use analysis, and urban planning. However, state-of-the-art (SOTA) CD methods still face two main challenges: (1) limited cross-temporal feature interaction at the encoder stage, which weakens the model’s discriminative capability; and (2) hard feature fusion at the decoder stage, which impairs fine-grained CD performance in dense prediction tasks. To address these issues, we propose STAR-VMamba, a spatiotemporal adaptive refinement network built upon the VMamba architecture to enhance cross-temporal feature interactions and enable adaptive feature fusion. Specifically, we design a spatiotemporal interleaved reinforcement module (ST-IRM) that explicitly models global spatiotemporal dependencies to promote cross-temporal semantic alignment. Furthermore, we introduce an adaptive attention re-weighting mechanism that dynamically emphasizes multi-level bi-temporal features. This mechanism consists of a single-level fusion module (SFM) for preserving local details and a multi-level aggregation module (MAM) for maintaining cross-level semantic consistency. Extensive experiments on three benchmark datasets demonstrate that STAR-VMamba outperforms existing SOTA methods by effectively suppressing background interference, enhancing fine-grained change representation, and maintaining the structural integrity of change maps.</p>