A comparison of different optical instruments and machine learning techniques to identify sprouting activity in potatoes during storage

• Published in: Journal of food measurement & characterization Vol. 14; no. 6; pp. 3565 - 3579
• Main Authors: Rady, Ahmed M., Guyer, Daniel E., Donis-González, Irwin R., Kirk, William, Watson, Nicholas James
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2020; Springer Nature B.V
• Subjects: Chemistry; Chemistry and Materials Science; Chemistry/Food Science; Classification; Cultivars; Data integration; Discriminant analysis; Engineering; Food Science; Hyperspectral imaging; Learning algorithms; Machine learning; Marketability; Near infrared radiation; Optical instruments; Optics; Original Paper; Potatoes; Shelf life; Tubers; Vegetables; Sprouting; Potatoes; Near-infrared; Classification; Machine learning; Sensor fusion; Primordial leaf count; Hyperspectral imaging
• ISSN: 2193-4126; 2193-4134
• DOI: 10.1007/s11694-020-00590-2

The quality of potato tubers is dependent on several attributes been maintained at appropriate levels during storage. One of these attributes is sprouting activity that is initiated from meristematic regions of the tubers (eyes). Sprouting activity is a major problem that contributes to reduced shelf life and elevated sugar content, which affects the marketability of seed tubers as well as fried products. This study compared the capabilities of three different optical systems (1: visible/near-infrared (Vis/NIR) interactance spectroscopy, 2: Vis/NIR hyperspectral imaging, 3: NIR transmittance) and machine learning methods to detect sprouting activity in potatoes based on the primordial leaf count (LC). The study was conducted on Frito Lay 1879 and Russet Norkotah cultivars stored at different temperatures and classification models were developed that considered both cultivars combined and classified the tubers as having either high or low sprouting activity. Measurements were performed on whole tubers and sliced samples to see the effect this would have on identifying sprouting activity. Sequential forward selection was applied for wavelength selection and the classification was carried out using K-nearest neighbor, partial least squares discriminant analysis, and soft independent modeling class analogy. The highest classification accuracy values obtained by the hyperspectral imaging system and was 87.5% and 90% for sliced and whole samples, respectively. Data fusion did not show classification improvement for whole tubers, whereas a 7.5% classification accuracy increase was illustrated for sliced samples. By investigating different optical techniques and machine learning methods, this study provides a first step toward developing a handheld optical device for early detection of sprouting activity, enabling advanced aid potato storage management.