Avaliação e predição das propriedades físicas e de qualidade de sementes de jatobá-do-cerrado processadas e armazenadas em diferentes condições, utilizando modelos de Machine Learning

Abstract

Post-harvest seed evaluation is subject to sampling and laboratory errors. Machine learning techniques can therefore help to optimize processes and obtain more accurate results for making decisions about the processing and conservation of stored seeds. The aim of this study was to evaluate and predict the physical properties and physicochemical quality of jatobá-do-cerrado seeds. The experimental design used was entirely randomized (5×3×2), with five different packages (polyethylene plastic, kraft paper, glass container, tetrapack, polyethylene container), three forms of processing (with pulp, without pulp - scarified, without pulp - fermented) and two storage conditions (10 and 23 °C). The variables studied were water content, seed mass, length, thickness, width, volume, apparent specific mass, projected area, sphericity, average diameter, circular area, circularity, drag coefficient, crude protein, ash and acidity index. The data was analyzed using Pearson correlation, Artificial Neural Networks, RepTree decision tree algorithms, M5P, Random Forest and Linear Regression. The results show that the processing of jatobá-do-cerrado seeds by fermentation and storage at 10 °C minimized the physical alterations and preserved the physical-chemical quality of the seeds, regardless of the type of packaging. The combination of the variables processing, temperature and packaging for the Artificial Neural Networks, RepTree and Random Forest and M5P algorithms outperformed linear regression, providing higher accuracy rates. The Artificial Neural Network and Random Forest models were the ones that best predicted the effects of treatments on changes in the physical properties and physicochemical quality of jatoba seeds.