Answers are grounded in this paper and cite their sources.
Human color vision may exhibit dichromatic, trichromatic, and potential tetrachromatic forms within the visible spectrum. However, hyperspectral remote sensing records surface reflectance over a much broader spectral range, where the red-edge and near-infrared (NIR) regions contain important discriminative information for land-cover classification. To bridge biological color perception and hyperspectral representation learning, this article proposes an NIR-aware tetrachromacy-inspired opponent representation for hyperspectral image classification. Hyperspectral pixels are first projected onto cone-like spectral sensitivity functions to generate perceptual color features. In addition to the conventional L, M, and S cone responses, an engineered O channel with a cone-like spectral weighting function is introduced to capture red-edge and NIR information for hyperspectral representation. The resulting L, M, S, and O responses are then transformed into opponent color features by extending the classical opponent processing mechanism, producing compact, and interpretable visible-NIR contrast representations. These features are finally processed by a spiking neural network classifier, forming a biologically inspired perception-opponent-classification pipeline. Experiments conducted on four hyperspectral datasets demonstrate that the proposed framework generally outperforms conventional spectral feature representations and achieves competitive classification performance against representative convolutional neural network-, transformer-, Mamba-, and spiking neural network-based approaches.
No takes yet. Share an insight, caveat, or question.
Xuan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: