Abstract
Early and accurate identification of phytopathogenic conidia, which cause substantial agricultural and economic losses, is critical for preventing disease and prescribing management. However, traditional methods based on morphology and molecular biology are time-consuming, labor-intensive, and ineffective at the species level. Here, we aimed to develop a new classification approach through Raman spectroscopy and data-driven modeling. Among the seven selected fungal species, three characteristic Raman wavenumbers at 1003-1005 cm(-1), 1153-1157 cm(-1), and 1515-1522 cm(-1) shared a consistent pattern across species and were attributed to carotenoids. Principal component analysis (PCA) results of their spectra showed substantial overlap that is insufficient for clustering. Consequently, three data-driven models - support vector machines (SVMs), decision trees (DTs), and eXtreme Gradient Boosting Forest (XGBoost) - were trained with three categories of features (number of peaks, maximum peak, and curve roughness) identified within eight characteristic wavenumber ranges. The optimal SVM, DT, and XGBoost determined by hyperparameter tuning achieved prediction precision of 0.88, 0.88, and 0.96, respectively. PCA-XGBoost trained by feeding principal components of PCA to XGBoost achieved prediction precision of 0.94, suggesting that features extracted from the raw datasets outperformed those extracted with PCA in terms of data-driven classification. In summary, a high classification precision has been achieved for conidial Raman spectra through XGBoost based on raw spectral feature extraction. This study lays the ground for the precise classification of phytopathogenic conidia by data-driven classification of Raman spectra thereby proving a great potential for preventing and controlling of plant fungal disease.