MSMF: a multiscale superpixel framework based on Mamba for hyperspectral image classification

Hyperspectral image (HSI) classification is one of the key tasks in remote sensing applications, which help to analyze the land cover details using high amounts of spectral data. Conventional techniques often struggle with high as well as irregular spatial distributions, which result in reduced accuracy and calculations. The Multiscale Superpixel Mamba Framework (MSMF) is proposed based on the Mamba model, which adds multiscale superpixel segmentation to capture the spatial long-range dependency and print the multiscale spectral feature. This method overcomes the shortcomings of standalone Mamba models, which improves the handling of HSI non-uniformity, improves the classification accuracy, and guarantees a compact model with parameters limited to 1.2 times the baseline. Evaluations on the Indian Pines dataset gives overall accuracy (OA) as 92.00% and average accuracy (AA) as 91.99% and Kappa coefficients as 91.46% outperforming the baselines such as SVM, CNN, HyT and SMN. Robustness Test in Noise Environments and Low Complexity Demonstrate its Applicability for Practical Deployments Analysis of many examples demonstrate the effectiveness of multiscale fusion and superpixel components. This framework advances HSI classification with a novel sequence modeling integrated with dynamic spatial segmentation that can provide avenues for improved and efficient remote sensing solutions.