Diagnóstico do potencial fisiológico de sementes de canola por descritores e aprendizado de máquina

Abstract

The physiological quality of seeds is essential for the initial establishment of crops, especially in expanding species such as canola (Brassica napus L.). The species is strategic for the global energy transition, as a raw material for biofuel production, in addition to industrial and food uses. Traditional tests to evaluate seed quality, although well established, have limitations related to their destructive nature, subjectivity, and the time required to obtain results. In this context, multispectral image analysis is an alternative, as it enables rapid and non-destructive evaluations. Therefore, the objective of this study was to investigate the efficiency of multispectral image analysis, associated with machine learning algorithms, in assessing the physiological quality of canola seeds. Six batches of seeds from two cultivars (Nuola and Diamond) were analyzed, characterized by traditional physiological tests. Multispectral images were captured using VideometerLab 4 equipment at 19 wavelengths of reflectance and autofluorescence. The main spectral descriptors were selected using the Gini decrease index. Supervised classification models were adjusted using Random Forest, Neural Networks, and PLS-DA. The results showed that spectral descriptors, especially reflectance at 780 nm and autofluorescence in the 470/500 nm combination, were efficient in discriminating vigor levels, showing an inversely proportional correlation with physiological tests. Therefore, multispectral image analysis associated with machine learning algorithms proved to be efficient for diagnosing the physiological potential of canola seeds, with high potential for application in automated quality control systems. Keywords: Brassica napus L. var. oleifera Moench; vigor; autofluorescence; machine learning.