Raman spectroscopy coupled with interpretable machine learning for early detection of citrus Huanglongbing

Citrus Huanglongbing (HLB), caused by Candidatus Liberibacter asiaticus, threatens global citrus production, while early diagnosis is hindered by a prolonged asymptomatic period and laboratory-intensive methods. This study establishes a non-destructive framework integrating portable Raman spectroscopy, reference compound validation, optimized machine learning, and model interpretability for early HLB detection. A total of 1025 Thai pomelo leaves yielding 3075 spectra were classified into four categories: healthy, early HLB, symptomatic HLB, and nutrient deficiency. Class-specific spectral variations occurred in carotenoid, chlorophyll a, carbohydrate, and polyphenol regions. Measurements of pure reference standards supported key molecular assignments beyond literature-based interpretation. UMAP followed by K-means clustering achieved a purity of 91.72% ± 0.96%, indicating spectral separability among the four classes. For supervised classification, 72 preprocessing workflows were evaluated across six classifiers, followed by hyperparameter optimization using four search strategies. The optimized support vector machine (SVM) achieved the highest accuracy of 95.93% ± 1.42% and an F1-score of 0.9597 ± 0.0146. SHAP analysis linked classification decisions to these molecular assignments, identifying pigment-associated features as major contributors to overall classification and a cellulose-associated feature at 1096.8 cm^-1 as the leading contributor to early HLB identification. By connecting molecular assignments, classification performance, and model interpretation, this framework provides a biologically grounded approach to early HLB detection with potential for field deployment.