Unravelling the physicochemical features of US wheat flours over the past two decades by machine learning analysis

Abstract

The physicochemical features of US wheat flours harvested from 1999 to 2020 were subjected to machine learning analysis for classifying wheat flour varieties and predicting bread loaf volumes. Principal component analysis demonstrated that 79.9% of the total variability was explained by two principal components that were mainly related to protein content and water absorption. The multilayer perceptron neural network with tuned hyperparameters was the most effective in correctly classifying wheat flours by showing high accuracy (98.91%) and F1-score (0.97), followed by the support vector machine, k-nearest neighbor, and decision tree algorithms. In addition, when three machine learning hyperparameters were tuned for predicting bread loaf volume, the use of Adam optimizer at a learning rate of 0.01 highly contributed to increasing the prediction accuracy, which was further improved by increasing the number of hidden layers up to 3 (R2 = 0.95–0.97 and RMSE = 19.30–27.45).