Polarization standard filtering enables multi-attribute tobacco grading by near-infrared spectroscopy

Objective, non-destructive grading of tobacco remains elusive because near-infrared (NIR) spectra of leaves from different quality classes are almost indistinguishable. Here we combine a large multi-label spectral data set with a single-pass pre-processing operator to resolve these subtleties.

High-resolution NIR reflectance spectra were collected from 971 flue-cured samples and annotated by expert tasters for ten sensory attributes, each compressed into 7-10 calibrated grades. We introduce Polarization-Standard Filtering (PSF), an analytic fusion of modified max-min scaling, variance-aware centering, Centralization transformation and Savitzky-Golay smoothing. PSF increases inter-sample Euclidean distances sevenfold while preserving spectral envelopes, overcoming the 0.99 cosine similarity that hampers raw data. Support-vector machines trained on PSF spectra achieve 99.7 % accuracy for seven-level quality grading-75 % higher than models using unprocessed spectra and 23 % above the best conventional pipeline. Across nine additional sensory attributes, accuracy ranged from 67 % to 98 %, with a median accuracy of 85.2 %. Notably, Fragrance 98.3 %, Impurity 98.3 %, and Mellow 94.9 % also achieved high accuracies. Sliding-window occlusion localizes the most informative wavelengths (1700-1850 nm and 2080-2300 nm) to O-H, N-H and C-H combination bands, enabling chemically interpretable feature reduction to ≤ 8 % of the original channels with minimal accuracy loss.

The open PSF-NIR framework transforms tobacco flavor evaluation from subjective bench tests to an inline, multi-attribute, machine-readable process.