Classical and machine learning tools for identifying yellow-seeded Brassica napus by fusion of hyperspectral features.

Abstract

Introduction: Due to its favorable traits-such as lower lignin content, higher oil concentration, and increased protein levels-the genetic improvement of yellow-seeded rapeseed has attracted more attention than other rapeseed color variations. Traditionally, yellow-seeded rapeseed has been identified visually, but the complex variability in the seed coat color of Brassica napus has made manual identification challenging and often inaccurate. Another method, using the RGB color system, is frequently employed but is sensitive to photographic conditions, including lighting and camera settings.

Methods: We present four data-driven models to identify yellow-seeded B. napus using hyperspectral features combined with simple yet intelligent techniques. One model employs partial least squares regression (PLSR) to predict the R, G, and B color channels, effectively distinguishing yellow-seeded varieties from others according to globally accepted yellow-seed classification protocols. Another model uses logistic regression (Logit-R) to produce a probability-based assessment of yellow-seeded status. Additionally, we implement two intelligent models, random forest and support vector classifier to evaluate features selected through lasso-penalized logistic regression.

Results and discussion: Our findings indicate significant recognition accuracies of 96.55% and 98% for the PLSR and Logit-R models, respectively, aligning closely with the accuracy of previous methods. This approach represents a meaningful advancement in identifying yellow-seeded rapeseed, with high recognition accuracy demonstrating the practical applicability of these models.