Predição de danos mecânicos em sementes de soja armazenadas em diferentes ambientes utilizando Aprendizado de Máquina

Abstract

Artificial intelligence has been widely applied in data prediction for better decision making and process optimization. In the post-harvest, the control of biotic and abiotic factors is fundamental for the conservation of seed quality in the storage stage. The tetrazolium test has been used to evaluate seed quality, however, with several limitations that can lead to evaluation errors. Thus, the aim of this study was to identify the best machine learning model for predicting mechanical damage, vigor and viability of soybean seeds during storage, depending on different conditions (10, 15 and 25 ºC), packaging (with coating and uncoated) and storage times (0, 3, 6, 9 and 12 months). The Quinlan algorithm M5 (M5P) and random forest (FA) models showed the best performance for predicting seed vigor and viability, while the artificial neural network model best predicted the results of mechanical damage to seeds. The Pearson correlation coefficient (r) and mean absolute error (MAE) were of low magnitude in predicting mechanical damage and seed moisture. In predicting the vigor and viability of soybean seeds, higher results were found for r and lower results for MAE. The Quinlan algorithm M5 (M5P) and random forest (FA) models are those that best predict soybean seed quality results, with a more simplified and agile analysis for determining the vigor and viability of soybean seeds in storage.