Predicting Pineapple Quality from Hyperspectral Data of Plant Parts Applied to Machine Learning

Abstract

Food quality detection by machine learning (ML) is more practical and sustainable as it does not require sample preparation and reagents. However, the prediction of pineapple quality by hyperspectral data applied to ML is not known. The aim of this study was to verify accurate ML models for predicting pineapple fruit quality and the best inputs for algorithms: Artificial Neural Networks (ANNs), M5P (model tree), REPTree decision trees, Random Forest (RF), Support Vector Machine (SMV) and Zero R. Three inputs were used for each model: leaf reflectance, peel reflectance, and fruit reflectance. The machine learning model SVM, stood out for its best results, demonstrating good generalization capacity and effectiveness in predicting these attributes, reaching accuracy values above 0.7 for Brix and ratio, using fruit reflectance. In terms of the overall efficiency of the input variables, peel and fruit were the most informative, with peel standing out for the estimation of secondary metabolism compounds, while the fruit showed excellent performance in predicting flavor-related attributes, such as acidity, °Brix and ratio, as mentioned previously, above 0.7. These results highlight the potential of using spectral data and machine learning in the non-destructive assessment of pineapple quality, enabling advances in monitoring and selecting fruits with better sensors.