STFMamba: Spatiotemporal satellite image fusion network based on visual state space model

Remote sensing images provide extensive information about Earth's surface, supporting a wide range of applications. Individual sensors often encounter a trade-off between spatial and temporal resolutions, spatiotemporal fusion (STF) aims to overcome this shortcoming by combining multisource data. Existing deep learning-based STF methods struggle with capturing long-range dependencies (CNN-based) or incur high computational cost (Transformer-based). To overcome these limitations, we propose STFMamba, a twostep state space model that effectively captures global information while maintaining linear complexity. Specifically, a super-resolution (SR) network is firstly utilized to mitigate sensor heterogeneity of multisource data, then a dual U-Net is designed to fully leverage spatio-temporal correlations and capture temporal variations. Our STFMamba contains the following three key components: 1) the multidimensional scanning mechanism for global relationship modeling to eliminate information loss, 2) a spatio-spectral-temporal fusion scanning strategy to integrate multiscale contextual features, and 3) a multi-head cross-attention module for adaptive selection and fusion. Additionally, we develop a lightweight version of STFMamba for deployment on resource-constrained devices, incorporating a knowledge distillation strategy to align its features with the base model and enhance performance. Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed method. Specifically, our method outperforms compared methods, including FSDAF, FVSDF, EDCSTFN, GANSTFM, SwinSTFM, and DDPMSTF, with average RMSE reductions of 24.25%, 25.94%, 18.15%, 14.36%, 9.63%, and 12.82%, respectively. Our code is available at: https://github.com/zhaomin0101/ STFMamba.