Sementes de milho popped kernel e furadas: armazenabilidade e classificação de defeitos por meio de algoritmos de aprendizado de máquina

Abstract

Corn is one of the most widely produced crops in the world, with great economic and nutritional importance. The seed, as a raw material, is responsible for transferring the advances of genetic improvement, guaranteeing high yields under proper management. Its quality integrates physical, physiological, genetic, and sanitary attributes, forming the basis of the crop's productive potential. However, biotic and abiotic stresses can reduce their quality, favoring the emergence of defective seeds such as popped kernel (PK) and perforated seeds. The objective of this study was to evaluate the physiological quality of corn seeds with PK and perforated defects during storage, as well as to assess the ability to classify these defects using supervised machine learning algorithms. Samples containing five defect ratios (0%, 2,5%, 7,5%, 15% e 25%) were stored under ambient conditions. Physiological quality assessments were performed at baseline and every three months, up to nine months for PK seeds and 12 months for samples with a perforated defect. Images of defective and healthy seeds were captured, and after processing, four supervised learning models and the GroundEye system were used for classification. It was observed that proportions higher than 2.5% for perforated seeds and 7.5% for PK seeds compromise germination and vigor after 3 and 6 months of storage, respectively. The MLP (Multi-Layer Perceptron) and SVM (Support Vector Machine) models demonstrated accuracies above 82% for samples containing perforated seeds and 96% for the prediction of PK seeds, respectively. On the other hand, the GroundEye system, using a reduced number of variables selected by the Kruskal–Wallis method, demonstrated good classification performance, with an accuracy of 98%. Tolerance levels for the presence of defective seeds were determined for each defect without compromising the physiological potential of the seeds. Image-based methodologies can assist laboratory analyses, providing accurate results and reducing subjectivity.