Li et al. (2026) Mamba for Remote Sensing: Architectures, Hybrid Paradigms, and Future Directions

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-01-12
• Authors: Zefeng Li, Long Zhao, Yihang Lu, Yue Ma, Guoqing Li
• DOI: 10.3390/rs18020243

Research Groups

Not specified (survey paper).

Short Summary

This survey evaluates the practical application of visual state-space models (e.g., Mamba) in Earth observation, analyzing their performance across various tasks and scales, and proposes design principles for their targeted integration into remote sensing pipelines.

Objective

• To provide a comprehensive review and practical guidance on the application of visual state-space models (like Mamba) in Earth observation, evaluating their effectiveness, developing a taxonomy of architectures, and outlining future research directions.

Study Configuration

• Spatial Scale: Discusses Earth observation data ranging from high spatial resolution to kilometre-scale context and large tiles.
• Temporal Scale: Discusses Earth observation data with dense temporal sampling and long temporal dependencies, including very long sequences.

Methodology and Data

• Models used: Visual state-space models (e.g., Mamba), Convolutional Networks (CNNs), Transformers. The survey analyzes these models.
• Data sources: Earth observation (EO) data, including high spatial resolution, wide swath, dense temporal sampling, hyperspectral, and multisource fusion data. The survey discusses applications to these data types.

Main Results

• Visual state-space models like Mamba offer linear-time sequence processing with selective recurrence, making them promising for remote sensing.
• The survey reviews the theoretical foundations of state-space models and the role of scanning and serialization for mapping two- and three-dimensional EO data onto one-dimensional sequences.
• A taxonomy of scan paths (centre-focused, geometry-aware) and architectural hybrids (CNN-Mamba, Transformer-Mamba, multimodal designs) is developed for various EO applications (segmentation, detection, restoration, scientific applications).
• Mamba is empirically warranted for specific task regimes involving very long sequences, large tiles, or complex degradations.
• Simpler operators or conventional attention mechanisms remain competitive in other task regimes.
• The survey advocates for Mamba's use as a targeted, scan-aware component in EO pipelines rather than a universal drop-in replacement.
• Discussions include green computing, numerical stability, reproducibility, and future directions for physics-informed state-space models and remote-sensing-specific foundation architectures.

Contributions

• First comprehensive survey analyzing the practical realization of visual state-space models (Mamba) in Earth observation.
• Development of a taxonomy for scan paths and architectural hybrids specifically tailored for 2D and 3D Earth observation data.
• Delineation of specific task regimes where Mamba demonstrates empirical advantages compared to existing models.
• Provision of concrete design principles for the targeted integration of Mamba into Earth observation pipelines.
• Discussion of critical operational aspects such as green computing, numerical stability, reproducibility, and outlining future research directions including physics-informed models.

Funding

Not specified in the provided text.

Citation

@article{Li2026Mamba,
  author = {Li, Zefeng and Zhao, Long and Lu, Yihang and Ma, Yue and Li, Guoqing},
  title = {Mamba for Remote Sensing: Architectures, Hybrid Paradigms, and Future Directions},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18020243},
  url = {https://doi.org/10.3390/rs18020243}
}