GLCSNet: A Building Change Detector With Global-Local Dual-Domain Awareness and Cross-Spatiotemporal Joint Modeling

Building change detection (CD) in high-resolution remote sensing imagery is crucial for urban sustainable development, rapid disaster response, and environmental monitoring. However, existing deep learning methods still struggle with the inherent complexity, dynamic change patterns, and high-noise characteristics of such imagery, particularly because conventional feature fusion approaches fail to effectively capture and decouple complex spatiotemporal evolution semantics between bitemporal images. To address these issues, this study proposes the global-local dual-domain co-aware and cross-spatiotemporal joint modeling network (GLCSNet), an efficient CD framework based on a hierarchical encoder–decoder mixed module. GLCSNet consists of two core modules. The global-local dual-domain co-aware block reduces computational cost through progressive spatial reduction while enhancing the discriminability and signal-to-noise ratio of local details by integrating a spatio-frequency dual-domain feature refinement network. The cross-spatiotemporal joint modeling block constructs heterogeneous relational views via spatiotemporal semantic reorganization, employs an omnidirectional scanning Mamba mechanism for efficient sequence modeling, and further applies temporal decomposition and difference fusion to model spatiotemporal evolution. Extensive experiments on urban datasets (LEVIR-CD and WHU-CD) and disaster datasets (CEBD and CWBD) demonstrate that GLCSNet consistently outperforms state-of-the-art methods across all metrics, for example, achieving 0.90%–5.36% IoU improvement on the LEVIR-CD dataset. Meanwhile, GLCSNet maintains a lightweight parameter scale and high computational efficiency, providing an effective solution for high-resolution remote sensing building CD.