Prediction of the ethanol yield of dry-grind maize grain using near infrared spectroscopy

The primary objective was to investigate the potential of using dry-grind maize spectral data for predicting ethanol yield, as well as to identity important wavelengths related to ethanol yield production. Spectral data from 222 ground maize grain samples were obtained over the spectral region of 400–2498 nm in 2 nm intervals. Thirty replicate runs were conducted, and for each run, 166 out of 222 samples were randomly selected for calibration, and the remaining 56 samples were used for validation. Partial least square regressions (PLSRs) were conducted on the complete spectra and on the wavelengths selected by bootstrapping based on the computed variable importance for projection values (Boot_VIP). Models with the wavelengths selected by the Boot_VIP and models with the complete spectra had similar prediction capabilities in the independent validation, with the average root mean square error of prediction (RMSEP) of 0.56%, which is comparable to the standard deviation of the dry fermentation reference method. The Boot_VIP procedure selected wavelengths consistently in 30 replicate runs, averaging 109 wavelengths selected at each run, and there were 101 wavelengths selected for all 30 runs. The selected wavelengths concentrated on 400–550 nm of the visible region and 2300–2360 nm of the NIR region. The first two factors from the Boot_VIP models were almost identical to the respective first two factors from models with the complete spectra.

► Gravimetric method is comparable to HPLC direct measurement.
► Boot_VIP selected wavelengths consistently and in narrow intervals.
► RMSEP of the prediction model is comparable to laboratory error.