Mamba-RSI: a state-space deep learning framework for efficient land-use and land-cover classification in remote sensing imagery

Accurate and efficient land-use and land-cover (LULC) classification from remote sensing imagery remains challenging. This is because it requires capturing long-range spatial dependencies while maintaining computational scalability. Recent transformer-based models improve global context modeling. However, they suffer from quadratic complexity and are limited in applicability to high-resolution imagery. We introduce Mamba-RSI: a linear-time, state-space deep learning framework using selective recursion, hierarchical multi-scale feature extraction, and lightweight global representations. Mamba-RSI captures both fine-grained spectral/texture information and coarse structural patterns with significantly less computational overhead than existing quadratic self-attention transformers. Extensive experimentation on EuroSAT and NWPU-RESISC45 demonstrated that Mamba-RSI achieves state-of-the-art performance. It achieved 99.72% accuracy on EuroSAT and 96.84% on RESISC45. This represents a +0.40% improvement over the strongest transformer baseline, ATMformer, on EuroSAT, a +0.29% improvement on RESISC45, and more than +0.53% over ViT-B on EuroSAT. Robustness tests under severe Gaussian noise (σ = 0.10) showed that Mamba-RSI maintains 97.43% accuracy. MaxViT, by comparison, maintains 94.01% in the same setting. Mamba-RSI also preserves 91.15% accuracy under 30% patch occlusion, outperforming ViT-B by +7.41%. Mamba-RSI provides an attractive blend of accuracy, robustness, and efficiency. It serves as a scalable foundation for new insights into remote sensing analytics and LULC mapping systems.