Laser-Induced Breakdown Spectroscopy Associated with the Design of Experiments and Machine Learning for Discrimination of <i>Brachiaria brizantha</i> Seed Vigor
Guilherme Cioccia,
Carla Pereira de Morais,
Diego Victor Babos,
Débora Marcondes Bastos Pereira Milori,
Charline Z. Alves,
Cícero Cena,
Gustavo Nicolodelli,
Bruno S. Marangoni
Affiliations
Guilherme Cioccia
SISFOTON-UFMS—Laboratório de Óptica e Fotônica, UFMS—Universidade Federal de Mato Grosso do Sul, Campo Grande 79070-900, MS, Brazil
Carla Pereira de Morais
Embrapa Instrumentation, São Carlos 13560-970, SP, Brazil
Diego Victor Babos
Embrapa Instrumentation, São Carlos 13560-970, SP, Brazil
Débora Marcondes Bastos Pereira Milori
Embrapa Instrumentation, São Carlos 13560-970, SP, Brazil
Charline Z. Alves
Programa de Pós-Graduação em Agronomia, UFMS—Universidade Federal de Mato Grosso do Sul, Chapadao do Sul 79560-000, MS, Brazil
Cícero Cena
SISFOTON-UFMS—Laboratório de Óptica e Fotônica, UFMS—Universidade Federal de Mato Grosso do Sul, Campo Grande 79070-900, MS, Brazil
Gustavo Nicolodelli
Departamento de Física, Universidade Federal de Santa Catarina, Florianópolis 88020-302, SC, Brazil
Bruno S. Marangoni
SISFOTON-UFMS—Laboratório de Óptica e Fotônica, UFMS—Universidade Federal de Mato Grosso do Sul, Campo Grande 79070-900, MS, Brazil
Laser-induced breakdown spectroscopy (LIBS) associated with machine learning algorithms (ML) was used to evaluate the Brachiaria seed physiological quality by discriminating the high and low vigor seeds. A 23 factorial design was used to optimize the LIBS experimental parameters for spectral analysis. A total of 120 samples from two distinct cultivars of Brachiaria brizantha seeds exhibiting high vigor (HV) and low vigor (LV) in standard tests were studied. The raw LIBS spectra were normalized and submitted to outlier verification, previously to the reduction data dimensionality from principal component analysis. Supervised machine learning algorithm parameters were chosen by leave-one-out cross-validation in the test samples, and it was tested by external validation using a new set of data. The overall accuracy in external validation achieved 100% for HV and LV discrimination, regardless of the cultivar or the classification algorithm.
