Bioimpedance-based prediction of dry matter content and potato varieties through supervised machine learning methods

Abstract

This study explored bioimpedance measurements and their derived parameters as potential indicators of dry matter content and potato variety. We investigated how bioimpedance correlates with the dry matter composition of potatoes,through destructive tests on tissue slices. Using the machine learning approach, we identified patterns and associations that can aid in predicting dry matter content, using regression models, and discriminating among potato varieties, focusing on classification models. In particular, four feature selection methods (correlation matrix, minimum-redundancy–maximum-relevance, neighborhood component analysis, and t-test) were evaluated against a baseline with all features. It was found that the best regression model for dry matter was a neural network regression model with t-test features, which achieved $R^2$ of 0.62 and RMSE of 2.3% in testing, while the best classification model for potato variety was a neural network classification model with correlation matrix features, achieving an $F_1$ score of 0.92.