Non-Destructive Classification of Raw-Material Cashmere and Wool Based on GOOSE-LightGBM Using Near-Infrared Spectroscopy, a Preliminary Investigation

Cashmere and wool exhibit highly overlapping near-infrared (NIR) absorption features, making their discrimination difficult. In this study, 390 washed fiber samples free from visible impurities and grease, including 210 wool samples from seven sources and 180 cashmere samples from six sources, were collected and analyzed, covering diverse fiber types with different origins and physical characteristics. Conventional methods such as principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA), and manually tuned classifiers often show limited class separability and strong dependence on feature selection and hyperparameter tuning. To overcome these limitations, a novel identification method combining NIR spectroscopy with a GOOSE-optimized Light Gradient Boosting Machine (LightGBM) was proposed. Competitive Adaptive Reweighted Sampling (CARS) was used to select informative spectral variables, followed by Linear Discriminant Analysis (LDA) for dimensionality reduction and enhanced class separability. The GOOSE algorithm optimized LightGBM hyperparameters, providing efficient global search and stable convergence compared with commonly used optimization strategies. The proposed GOOSE-LightGBM model achieved 98.9% classification accuracy for cashmere–wool discrimination, providing an effective solution for identifying fibers with highly overlapping spectral features and a technical reference for rapid, non-destructive fiber identification.