Leakage-safe benchmarking of conventional chemometric and deep learning models for food NIR spectroscopy regression

Near-infrared (NIR) spectroscopy is widely used in food systems for rapid, non-destructive estimation of compositional and quality attributes. Although machine-learning and deep-learning models are increasingly reported for spectral regression, practical model selection is often hindered by heterogeneous validation setups and the risk of information leakage during preprocessing and tuning. In this study, we provide a leakage-safe benchmark across nine food-related regression tasks spanning different sample sizes and spectral dimensionalities. Conventional chemometric and regression baselines (partial least squares regression, ridge regression, and support vector regression) are compared with lightweight deep-learning baselines (a shallow artificial neural network and a one-dimensional convolutional neural network) under a unified leakage-safe evaluation framework combining nested cross-validation with acomplementary fixed held-out test-set assessment, fold-wise preprocessing, and repeated training for stochastic models. Performance is summarized primarily using prediction error (RMSE in nested cross-validation and RMSEP on the held-out test set), while the coefficient of determination is retained as a secondary descriptive measure, complemented by cross-validation and fixed held-out-test learning-curve analyses together with model-complexity analysis. Across both validation analyses, ridge regression achieved the largest number of numerical task-level wins, while partial least squares regression and support vector regression remained competitive across multiple tasks. Paired randomization tests following van der Voet further showed that some small numerical differences between models were not statistically distinguishable. Artificial neural network and one-dimensional convolutional neural network baselines did not provide a reliable generalization advantage and showed greater variability in some data-limited settings. From a food-measurement perspective, these findings indicate that well-validated conventional models remain a strong default for routine NIR calibration workflows and non-destructive quality assessment. Deep learning appears most justified when calibration settings provide larger datasets, stronger experimental control, or spectroscopy-specific inductive bias that can translate added model complexity into practical measurement gains. Overall, this study serves as a controlled benchmark and reference for model selection under realistic food-spectroscopy data regimes.