Abstract
Low-rank tensor recovery-based feature extraction from authentic hyperspectral images (HSIs) has become widely employed to improve classification performance by removing sparse errors while preserving multi-dimensional structures. However, these techniques operate on either individual planar slices or unfolded matrix of each dimension, thereby overlooking spectral cross-channel correlation. Although quaternion-based low-rank recovery takes advantage of spectral cross-channel correlations, it often depends on spectral compression of global spectral bands, overlooking most spectral information, including the heterogeneity of various spectral regions and the homogeneity of regional spectrums. To address the challenges, this paper proposes a super-spectrum parallel low-rank quaternion recovery (SS-PLRQR) model composed of a hybrid super-spectrum grouping strategy (HSSGS) and a parallel low-rank quaternion recovery (PLRQR) algorithm. HSSGS streamlines the algorithmic flow and provides necessary conditions for subsequent adaptive analysis by visually partitioning global bands into heterogeneous sub-regions. Following this, PLRQR effectively and independently eliminates noisy errors from each super-spectrum, utilizing fitted parameter values tailored to the heterogeneity of the different super-spectrums. It maintains the complete planar spatial structure and cross-channel correlation of homogeneous spectrums within the same quaternionic super-spectrum. Comparative experiments on three real-world HSIs demonstrate the proposed model’s remarkable applicability, effectiveness, and robustness to various noises in classification.
Funding Information
  • Humanities and Social Science Fund of Ministry of Education (23YJAZH111)
  • Natural Science Foundation of Hunan Province (2023JJ10054)
  • Open Project of Key Laboratory of High Impact Weather (special), China Meteorological Administration
  • National Natural Science Foundation of China (42375016)

This publication has 59 references indexed in Scilit: