Answers are grounded in this paper and cite their sources.
Convolutional neural networks (CNNs) are widely used in hyperspectral image (HSI) classification but are often hindered by excessive parameters and computational cost. To enable efficient deployment in resource-constrained scenarios, we propose ELS ² C-Net, a lightweight spectral-spatial network built upon two core modules: a convolutional kernel decomposition-based multiscale feature extractor (CMSF) that decomposes standard convolutions into efficient multiscale operations, and a dual-aware efficient channel attention (DAE) mechanism that jointly models global channel dependencies and local spatial context for low-cost feature calibration. We also contribute a new hyperspectral dataset for geological survey. Extensive experiments on three public benchmarks (Pavia University, Salinas, and Dioni) and our self-built HyperGeo dataset demonstrate that ELS ² C-Net achieves 98.82%, 99.63%, and 97.07% overall accuracy on the three public benchmarks, outperforming eight state-of-the-art algorithms while requiring only 6.33 K parameters and 1.42 M MACs on the Dioni dataset.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: