Comparison of Peruvian coffee varieties applying near infrared spectroscopy (NIRS) and machine learning

In the coffee industry, the traditional cupping process, based on subjective judgments, has been superseded by implementing more objective techniques. Near-infrared spectroscopy (NIR) and Machine Learning (ML) algorithms were used to compare Peruvian coffee varieties in this context. Data preparation was critical, addressing outlier detection and management and partitioning into training and test sets. The evaluation focused on four models: PLS, SVM, and Random Forests. The results revealed that Random Forests stood out with exceptional performance, achieving an average accuracy of 96%. This key metric was complemented by RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and overall accuracy. The comparison of these metrics allowed us not only to evaluate the predictive ability of the models but also to understand the magnitude and nature of the prediction errors. Implementing these models provided a rapid and non-destructive evaluation of the quality of coffee varieties, overcoming the limitations of the traditional cupping approach. This study supports the feasibility of integrating NIR with ML as an effective tool for comparing Peruvian coffee varieties, as well as providing insight into the objective evaluation of coffee quality, opening new perspectives in the coffee industry.