Algoritmo floresta aleatória e estatística multivariada utilizados para classificação do estádio de maturação de cagaitas (Eugenia Dysenterica)

Abstract

The Cerrado biome, due to its vast territorial extent and high biological diversity, plays a strategic role in both food security and biodiversity conservation in Brazil. Covering approximately 24% of the national territory, this ecosystem contributes significantly to Brazilian agricultural and livestock production. However, increasing human interventions have threatened the region’s ecological balance, making the monitoring and preservation of native species essential. Eugenia dysenterica, commonly known as cagaita, is an endemic fruit of the Cerrado with commercial potential, yet it poses postharvest conservation challenges due to the delicate texture of its pulp. In this context, the present study aimed to classify the different maturation stages of cagaita based on physical, physicochemical, and chemical attributes, in order to determine the optimal harvest point and support large-scale commercialization. The variables analyzed included weight, diameter (transverse and longitudinal), surface area, color (lightness, hue angle, and chroma), firmness (peel and pulp), soluble solids, pH, titratable acidity, and maturation index. Data were collected using a completely randomized design (CRD) with one factor — days after anthesis (DAA) — comprising five treatments (28, 31, 34, 37, and 40 DAA), with 40 replications and three fruits per experimental unit. Statistical analyses involved multivariate methods (cluster analysis and principal component analysis) and the random forest algorithm, known for its ability to handle complex classification problems. Principal component analysis identified the most relevant variables for differentiating maturation stages, while cluster analysis revealed the formation of three distinct groups based on DAA. The random forest model showed satisfactory performance, with balanced metrics of precision, recall, and F1-score, as well as a Matthews correlation coefficient (MCC) of 0.62. These findings demonstrate the potential of integrating multivariate statistics and machine learning for the characterization of native fruits, providing valuable insights for improving harvest, handling, and processing practices in Cerrado fruit farming.