Predição da qualidade de sementes de soja armazenadas em diferentes ambientes e embalagens usando aprendizado de máquina

Abstract

Storage as a post-harvest stage is responsible for postponing deterioration and maintaining seed quality, and during this time some variables have a direct impact on the material, such as storage time, temperature and packaging. Traditional seed quality analyzes are processes that sometimes test several in laboratories to obtain results. Therefore, using learning techniques can maximize time, and thus collaborate with better decision making in storage processes. This work analyzed the feasibility of applying machine learning algorithms to predict the physiological quality of soybean seeds. For this, different inputs were tested (storage temperature (C), packaging (E), time (T), C + E, C + T, E + T and C + E + T) and machine learning models ( linear regression, artificial neural networks, decision tree, M5P and random forest). For each combination, the Person coefficient (r) values and the mean absolute error between the observed and predicted values with 10 repetitions (folds) were captured. These values were collected in an analysis of variation and grouping of means by the Scott-Knott test considering a 5x7 factorial (machine processing models x inputs) in a completely randomized design. The results of this research indicate the possibility of predicting variables related to seed quality using information about the storage environment as input to the random forest model. The proposed approach stands out in terms of cost and speed when compared to routinely used analysis methods.