Comparison of partial least squares-discriminant analysis, support vector machines and deep neural networks for spectrometric classification of seed vigour in a broad range of tree species

Seed vigour significantly influences the seed production and plant regeneration performance. The capability of NIR spectroscopy to identify seed vigour across multiple tree species rapidly and cost-effectively has been examined. The NIR spectra of seeds from five different tree species have been taken. Standard germination testing has also been used to verify seed vigour. Three classification models were trained, i.e., partial least squares-discriminant analysis (PLSDA), support vector machine (SVM) and Multilayer Deep neural network (DNN). Three types of spectral pre-processing methods and their combination were used to fit for the best classification model. The DNN model has shown good performance on all pre-processing methods and yielded higher accuracy than other models in this study, with accuracy, sensitivity, precision and specificity all equal to 1. Compared with other pre-processing methods, the second derivative spectra have shown a robust and consistent classification result in both PLSDA and DNN models. Five important regions including 1270, 1650, 1720, 2100, 2300 nm were found highly related to the seed vigour. This study has found a rapid and efficient methodology for seed vigour classification, which could serve for industrial use in a rapid and non-destructive way.