Abstract
Abstract
Near-infrared (NIR) spectroscopy is widely used for the rapid, nondestructive determination of constituents such as protein and moisture in cereal grains, but model comparisons in this field are often confounded by differences in the inputs received by each model. We benchmark a compact one-dimensional CNN derived by one-factor-at-a-time ablation (CNN-Baseline) and a randomly searched CNN (CNN-RS) against PLSR, SVR, XGBoost on six cereal-grain datasets (n=500-5046) under three input conditions: raw spectra, optimal preprocessing, and preprocessing plus wavelength selection. A single-filter with a kernel size of 11 of convolution and batch normalization proved sufficient. Given a unified raw input, CNN-RS was most accurate on five of the six datasets, while CNN-Baseline came within 0.003-0.017 in test
R
2
on the four larger datasets. PLSR led only on the smallest protein dataset (XDS, n=500) and was the most stable model. Preprocessing had almost no effect on CNN accuracy on the four larger datasets, but improved accuracy by 0.034-0.047 (CNN-Baseline) and 0.017-0.055 (CNN-RS) on the two smallest, and brought the networks to their optimum in a median of 49% fewer epochs. Therefore, replacing explicit preprocessing with learned filters was only justified when sufficient training data were available. Once each model received its own preprocessing and wavelength subset, SVR became most accurate on four of six datasets and XGBoost rose from the weakest model to within 0.01-0.03 of the best model on four of the six datasets, suggesting that the principal advantage of CNNs lied in robustness to raw input data rather than in attainable accuracy. Overall, model selection for NIR calibration of cereal grains depends more on sample size and the input preparation than on network architecture.