NDVI Change-Guided Transformer–Mamba Hybrid Network for Spatiotemporal Fusion

High spatiotemporal resolution imagery is essential for dynamic earth observation. Efficiently fusing multisource imagery to obtain high spatial and temporal resolution surface information has become a core research problem in the field of remote sensing. However, existing spatiotemporal fusion methods have limited ability to model phenological dynamics and land cover changes across both short and long temporal intervals, resulting in suboptimal fusion performance under complex surface change conditions. To address this problem, this paper proposes an innovative dual-input normalized difference vegetation index (NDVI) change-guided spatiotemporal fusion model, termed NDVI change-guided Transformer-Mamba hybrid network for spatiotemporal fusion (NCTM-STF). An NDVI change-guided feature enhancement module (NFEM) was constructed based on the Swin Transformer architecture, explicitly incorporating NDVI change information into the attention mechanism. Then, combined with a differential feature extraction module (DFEM) also constructed based on the Swin Transformer architecture, the proposed model could automatically focus on change regions while extracting differential features, thus enhancing the characterization of both short and long temporal intervals surface dynamic changes. Moreover, a channel-aware Mamba fusion module (CA-MFM) is designed to achieve feature fusion across both spatial and channel dimensions. Experimental results on the CIA and LGC datasets demonstrate that the fusion images of NCTM-STF consistently outperform existing methods under both two- and three-input settings in both qualitative and quantitative evaluations. Ablation studies further verify the effectiveness of each core module. Its advantages are especially evident in complex scenarios with significant phenological changes and drastic land cover variations, even under long-interval conditions, highlighting its superior spatiotemporal fusion performance.