Rapid and non-destructive identification of resistance levels to Phomopsis seed decay in soybean via near-infrared spectroscopy and machine learning algorithms

Phomopsis longicolla -induced soybean Phomopsis seed decay (PSD) threatens yield and food security. The method combing near-infrared spectroscopy with chemometrics was applied to predict the resistance levels of soybeans to fungal diseases in this study. Chemical composition analysis of soybean germplasms with different resistance levels indicated resistance may relate to the content of saponins, which can enhance the plant defense system. The spectral data showed saponin content differences can be reflected in the 1660-1700 nm region via different quantities of molecules with functional groups such as C-H bonds. Near-infrared spectroscopy (NIRS) employed chemometric methods to establish predictive models. These approaches integrated spectral preprocessing with machine learning algorithms. Preprocessed NIR spectral was used to build seven classification models for resistance identification. Especially, the Particle Swarm Optimization (PSO) algorithm was employed for data feature selection to optimize the support vector machine (SVM) and Bayesian-optimization multilayer perceptron (BO-MLP). Particle Swarm Optimization-Bayesian-optimization multilayer perceptron (PSO-BO-MLP) model demonstrated the best effect for classification, with the prediction accuracy to 89.18% and loss to 0.49. This study established a "spectral-biochemical-resistance" correlation, providing an accurate, non-destructive, and efficient method for evaluating the resistance of PSD in soybean, thereby offering a new tool for food-oriented soybean seed raw materials screening.