Predição de Flavonoides com o uso de dados Hiperespectrais em plantas de milho submetidas a doses de nitrogênio

Abstract

Due to the importance of corn in human and animal nutrition, research into agricultural practices and secondary metabolites, such as flavonoids (FLV), has increased. These pigments help protect against oxidation caused by reactive oxygen species. The aim of this study was to analyze the amount of FLV in maize plants subjected to nitrogen (N) doses using hyperspectral reflectance and a machine learning (ML) algorithm. An experiment was conducted in a randomized block design (RBL), with a 4x5 factorial scheme. Four doses of N applied in top dressing were evaluated, using urea as the source of N. The N doses were as follows: N1 (no nitrogen fertilization), N2 (30% of the recommended dose - 54 kg ha-1 of N), N3 (60% of the recommended dose - 108 kg ha-1 of N) and N4 (120% of the recommended dose - 216 kg ha-1 of N). Physiological assessments included the extraction of isoflavones from leaves at different stages of corn development (V6, V8, V10, V12 and V14). Spectroradiometry was used to measure the spectral reflectance of maize leaves in different wavelength ranges. ML algorithms such as artificial neural networks, decision trees and random forests were used to predict flavonoid levels in plants based on spectral data. The study evaluated the relationship between N application and FLV levels in maize plants using hyperspectral spectroscopy and ML, highlighting the importance of these factors in maize production and offering insights into how the technology can be used to optimize crop production. The results show that the different ML algorithms had varying performances in predicting these levels based on the spectral data. The SVM algorithm played an extremely important role in predicting the variables D1, D2, G1, G2 and total isoflavones, outperforming the other algorithms, especially when the input consisted of wavelength (WL). In general, with regard to the correlation coefficient (r), the SVM performed best when the input was WL for all variables, and with regard to the Mean Absolute Error (MAE), the SVM stood out when WL was used as the input for all variables.