Convolutional Neural Networks in Vis-NIR Chemometrics: From Contradiction to Conditional Design

Source raw items (1)

• arXiv2026-07-29 07:17:09
  Convolutional Neural Networks in Vis-NIR Chemometrics: From Contradiction to Conditional Design

  Near-infrared (NIR and Vis-NIR) spectroscopy is widely used for rapid, non-destructive analysis in food, agriculture, pharmaceuticals, process analytical technology, and bioprocess monitoring. Nevertheless, deep-learning studies in NIR chemometrics often reach conflicting conclusions about convolutional neural network (CNN) design, including kernel size, depth, preprocessing, model complexity, and transfer robustness. This review argues that many apparent contradictions reflect incomplete experimental conditioning rather than incompatible findings. CNN performance depends on interactions among spectral physics, dataset regime, acquisition protocol, validation design, and deployment conditions. We organize the literature around three moderators. First, NIR signals are indirect, highly collinear, and shaped by broad overlapping bands, scattering, temperature, and matrix effects. Second, CNN design should be interpreted through receptive-field reasoning: kernel size, depth, dilation, and multi-scale branches determine the wavelength span available to the model, whereas the effective receptive field indicates which parts are actually used. Third, validation design can behave as a hidden hyperparameter because random splits may reward architectures that exploit shared batch, instrument, season, or process-run structure instead of transferable chemical information. We therefore propose a conditional design framework in which preprocessing, architecture, hyperparameter optimization, transfer evaluation, interpretability, and reproducibility are treated as coupled components. Rather than seeking a universally optimal CNN, the framework aims to support physics-aware, shift-aware, and reproducible model comparison in NIR chemometrics.