Mamba for Remote Sensing: Architectures, Hybrid Paradigms, and Future Directions

Modern Earth observation combines high spatial resolution, wide swath, and dense temporal sampling, producing image grids and sequences far beyond the regime of standard vision benchmarks. Convolutional networks remain strong baselines but struggle to aggregate kilometre-scale context and long temporal dependencies without heavy tiling and downsampling, while Transformers incur quadratic costs in token count and often rely on aggressive patching or windowing. Recently proposed visual state-space models, typified by Mamba, offer linear-time sequence processing with selective recurrence and have therefore attracted rapid interest in remote sensing. This survey analyses how far that promise is realised in practice. We first review the theoretical substrates of state-space models and the role of scanning and serialization when mapping two- and three-dimensional EO data onto one-dimensional sequences. A taxonomy of scan paths and architectural hybrids is then developed, covering centre-focused and geometry-aware trajectories, CNN– and Transformer–Mamba backbones, and multimodal designs for hyperspectral, multisource fusion, segmentation, detection, restoration, and domain-specific scientific applications. Building on this evidence, we delineate the task regimes in which Mamba is empirically warranted—very long sequences, large tiles, or complex degradations—and those in which simpler operators or conventional attention remain competitive. Finally, we discuss green computing, numerical stability, and reproducibility, and outline directions for physics-informed state-space models and remote-sensing-specific foundation architectures. Overall, the survey argues that Mamba should be used as a targeted, scan-aware component in EO pipelines rather than a drop-in replacement for existing backbones, and aims to provide concrete design principles for future remote sensing research and operational practice.