Location: Healthy Processed Foods ResearchTitle: Nondestructive detection of zebra chip disease in potatoes using near-infrared spectroscopy
|Haff, Ronald - Ron|
|Hua, Sui Sheng|
|Munyaneza, Joseph - Joe|
Submitted to: Biosystems Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/28/2017
Publication Date: 12/19/2017
Citation: Liang, P., Haff, R.P., Hua, S.T., Munyaneza, J.E., Yilmaz, M.T., Sarreal, S.L. 2017. Nondestructive detection of zebra chip disease in potatoes using near-infrared spectroscopy. Biosystems Engineering. 166:161-169. https://doi.org/10.1016/j.biosystemseng.2017.11.019.
Interpretive Summary: NIR spectroscopy was evaluated as the basis for a rapid, non-destructive method for the detection of Zebra Chip disease in potatoes. NIR spectra of 363 potatoes were acquired and analyzed using advanced multivariate statistical methods. A classifier to discriminate infected tubers from the non-infected ones using the spectral data was built and achieved 98% classification accuracy. The high accuracy achieved indicates that reflection of light in the NIR region is a suitable basis for high-speed, non-destructive detection of Zebra Chip disease. A mathematical model was also built to predict concentrations of sugars (glucose, sucrose, and fructose) in tubers using the spectral data and the regression between actual concentrations and predicted concentrations showed good linearity (about 70%). The linearity provided us some insights into the scientific background of the NIR spectroscopy detection method.
Technical Abstract: Near-Infrared (NIR) spectroscopy in the wavelength region from 900 nm to 2600 nm was evaluated as the basis for a rapid, non-destructive method for the detection of Zebra Chip disease in potatoes and the measurement of sugar concentrations in affected tubers. Using stepwise regression in conjunction with canonical discriminant analysis applied to raw NIR spectra, a total classification error rate of 1.65% (i.e. 98.35% accuracy) with the type II error (2% false negative) rate slightly higher than the type I error (1% false positive) rate was achieved. The same analysis applied to 2nd derivative spectra, yielded a total classification error rate of 2.75% (i.e. 97.25% accuracy) with almost equal rates of type I and type II errors. Canonical discriminant analysis applied to sucrose, glucose, and fructose concentrations previously determined by high-performance liquid chromatography (HPLC) resulted in 4.3% false positive, 2.3% false negative, and 3.3% total classification error rates in detecting infected tubers. Classification was slightly improved when fructose was excluded from the variables, with 3.7% false positive, 2.3% false negative, and 3% total classification error rate. Partial least squares (PLS) regression models built to predict sugar concentrations from the NIR spectra resulted in regression R2 for actual vs. predicted concentrations of 0.68 and 0.71 respectively for sucrose and glucose, and 0.56 for fructose. However, the fructose model had the lowest standard error of cross-validation (SECV). The predictive ability and required number of factors for each sugar were significantly improved when the 2nd derivative spectra were used.