Machine learning to support geographical origin traceability of Coffea Arabica

Abstract

The species, variety and geographic origin of coffee directly influence the characteristics of the coffee beans and, consequently, the quality of the beverage. The added economic value that these features bring to the product has boosted the use of non-designative tools for authentication purposes. In this work, the feasibility of implementing a traceability system for Arabica coffee by country of origin was investigated using quality attributes and supervised machine learning algorithms: Multilayer Perceptron (MLP), Random Forest (RF), Random Tree (RT) and Sequential Minimal Optimization (SMO). We use an available database containing quality parameters for coffee beans produced in 15 countries, including the largest exporters and importers. Overall, Ethiopia, Kenya and Uganda had the highest coffee quality index (Total Cup Points). Differences between countries were found with 99% confidence using Robusta Multivariate Data Science with original data and 98% accuracy using Bootstrapping resampling method and Supervised Machine Learning algorithms. The model obtained by RF provided the best classification accuracy. The most important attributes to discriminate Arabica coffee by country of origin, in descending order, were body, moisture, total cup points, cupper points, acidity, aftertaste, flavor, aroma, balance, sweetness and uniformity. The coffee variety proved to be a promising variable to increase accuracy and can be incorporated among the quality attributes for classification and grading of coffee beans.