Rapid and accurate PLS-RBF neural networks for multi-constituent prediction in intact fresh maize kernels using NIR spectroscopy

Developing robust and accurate near-infrared (NIR) neural network calibration models for intact fresh maize kernels remains challenging in fresh maize breeding. In this study, four rapid and reliable PLS-RBF neural networks were established by integrating PLS feature extraction with ridge-regression-optimized RBF neural networks, to predict the contents of amylopectin, protein, crude fiber, and total sugar in intact fresh maize kernels, respectively. Compared with PCA-based models, PLS-based models consistently exhibited superior predictive performance, with RPD improvements of at least 15.89%,72.85%,54.77% and 19.01% for amylopectin, protein, crude fiber, and total sugar, respectively. The optimized PLS-RBF neural networks presented acceptable predictive performance, with RP2 value ranging from 0.871 to 0.970 and RPD values between 2.799 and 5.836. Overall, the proposed PLS-RBF neural networks provide efficient, nondestructive approaches for multi-constituent assessment of intact fresh maize kernels and demonstrate considerable potential for practical quality control applications within the collected sample population.