Submitted to: Cereal Chemistry
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/1/2007
Publication Date: 3/1/2007
Citation: Lee, K., Herrman, T.J., Bean, S., Jackson, D.S., Lingenfelser, J. 2007. Classification of dry-milled maize grit yield groups using quadratic discriminant analysis and decision tree algorithm. Cereal Chemistry. 84:152-161. Interpretive Summary: Physical properties of cereal grains are important end-use quality characteristics. The ability to predict end-use properties based on these grain attributes is important to the cereal processing industry. This study utilized a diverse set of maize hybrids to build statistical models to predict milling yield of maize based on physical grain attributes. Pattern recognition techniques predicted milling yields correctly 90%+ with discriminant analysis and regression methods correctly predicting milling yield 70-80% of the time. Such methods could be utilized by the grain processing industry to help predict processing performance of cereals.
Technical Abstract: A genetically and environmentally diverse collection of maize (Zea maize L.) samples were evaluated for physical properties and grit yield to help develop a standard set of criteria to identify grain best suited for dry milling. Application of principal component analysis reduced a set of approximately 500 samples collected from 6 states to 154 maize hybrids. Regression analysis explained approximately 50% of the variability in dry milling grit yield. Prediction of dry-milled grit yield groups using two pattern recognition techniques, discriminant analysis and decision tree algorithm improved classification accuracy. Selected maize hybrids were assigned into seven groups according to their dry-milled grit yields. Patterns of differences in the physical properties for the seven grit yield groups were observed, allowing the seven yield groups to be set into two or three groups. The correct classification rates were in the range of 82.1-91.4% when the samples were divided into three yield groups and 93.8-96.1% when samples were divided into two yield groups. Quadratic discriminant analysis and classification and regression tree (CART) models developed using a training data set had a moderate prediction ability of approximately 70% when samples were in three groups and 81% when divided into two yield groups.