SCS-Mamba: Spectral-curriculum Sharpness-aware Optimization With a Gated State-space Network for Few-epoch Hyperspectral Image Classification

Hyperspectral image classification has advanced quickly through convolutional, transformer and state-space architectures, yet progress is reported at convergence and says little about what a classifier achieves when the training schedule is short.Architectures are designed in one community and optimizers in another, so spectral resolution, losslandscape geometry and the training budget are not scheduled together.This paper presents SCS-Mamba, a framework in which the epoch budget is a declared constraint rather than an outcome.A spectral geometry analyzer scores every band group by entropy, redundancy and class separability; a curriculum scheduler admits groups from low to high complexity across stages advancing on a validation plateau; and SC-SAM, a sharpness-aware optimizer whose perturbation radius is tied to the complexity of the active stage, seeks flat minima in the region the curriculum has just opened.The backbone mixes, through a learned gate, a depthwise-separable three-dimensional convolution, a bidirectional selective state-space scan and a spectral-token residual path per band group.Under a leakage-controlled protocol with spatially disjoint blocks and guard bands, and with up to 30 labeled samples per class (three small Indian Pines classes receive half their pixels, giving 437 rather than 480), SCS-Mamba reaches 89.67%, 95.45%, 91.27% and 98.95% overall accuracy on Indian Pines, University of Pavia, Houston 2013 and WHU-Hi-LongKou, improving on the strongest baseline of each scene, among twelve retrained baselines, by 0.34 to 1.38 points while reaching early stopping in 47.1% fewer epochs on average, 35.1% less wall-clock time and 49.6% fewer training FLOPs.Ablations show that the curriculum and the adaptive radius overlap by less than 0.15 points, so neither can be dropped, and the training scheme carries part of its benefit to five backbones it was not designed for.Accuracy is also reported at matched epoch, FLOP and wall-clock budgets, under per-baseline near-optimal configurations and over repeated spatial partitions.Code: https://github.com/Poojadahiya01/SCS_MAMBA.

Authors

Institutions

Publication Details

Journal
International journal of intelligent engineering and systems
Published
2026-11-17
DOI
https://doi.org/10.22266/ijies2026.1231.66
Primary Topic
Remote-Sensing Image Classification
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

SCS-Mamba: Spectral-curriculum Sharpness-aware Optimization With a Gated State-space Network for Few-epoch Hyperspectral Image Classification

Kavita Rathi, Pooja Dahiya
1 links
International journal of intelligent engineering and systems
Remote-Sensing Image Classification
article

SCS-Mamba: Spectral-curriculum Sharpness-aware Optimization With a Gated State-space Network for Few-epoch Hyperspectral Image Classification

Kavita Rathi, Pooja Dahiya
article en

Abstract

Hyperspectral image classification has advanced quickly through convolutional, transformer and state-space architectures, yet progress is reported at convergence and says little about what a classifier achieves when the training schedule is short.Architectures are designed in one community and optimizers in another, so spectral resolution, losslandscape geometry and the training budget are not scheduled together.This paper presents SCS-Mamba, a framework in which the epoch budget is a declared constraint rather than an outcome.A spectral geometry analyzer scores every band group by entropy, redundancy and class separability; a curriculum scheduler admits groups from low to high complexity across stages advancing on a validation plateau; and SC-SAM, a sharpness-aware optimizer whose perturbation radius is tied to the complexity of the active stage, seeks flat minima in the region the curriculum has just opened.The backbone mixes, through a learned gate, a depthwise-separable three-dimensional convolution, a bidirectional selective state-space scan and a spectral-token residual path per band group.Under a leakage-controlled protocol with spatially disjoint blocks and guard bands, and with up to 30 labeled samples per class (three small Indian Pines classes receive half their pixels, giving 437 rather than 480), SCS-Mamba reaches 89.67%, 95.45%, 91.27% and 98.95% overall accuracy on Indian Pines, University of Pavia, Houston 2013 and WHU-Hi-LongKou, improving on the strongest baseline of each scene, among twelve retrained baselines, by 0.34 to 1.38 points while reaching early stopping in 47.1% fewer epochs on average, 35.1% less wall-clock time and 49.6% fewer training FLOPs.Ablations show that the curriculum and the adaptive radius overlap by less than 0.15 points, so neither can be dropped, and the training scheme carries part of its benefit to five backbones it was not designed for.Accuracy is also reported at matched epoch, FLOP and wall-clock budgets, under per-baseline near-optimal configurations and over repeated spatial partitions.Code: https://github.com/Poojadahiya01/SCS_MAMBA.

International journal of intelligent engineering and systemsVol. 19(12)
Deenbandhu Chhotu Ram University of Science and Technology (IN)
Remote-Sensing Image Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon