Early Detection and Dynamic Grading of Sweet Potato Scab Based on Hyperspectral Imaging

This study investigates the early detection of sweet potato scab by using hyperspectral imaging and machine learning techniques. The research focuses on developing an accurate, economical, and non-destructive approach for disease detection and grading. Hyperspectral imaging experiments were conducted on two sweet potato varieties: Guangshu 87 (resistant) and Guicaishu 2 (susceptible). Data preprocessing included denoising, region of interest (ROI) selection, and average spectrum extraction, followed by dimensionality reduction using principal component analysis (PCA) and random forest (RF) feature selection. A novel dynamic grading method based on spectral-time data was introduced to classify the early stages of the disease, including the early latent and early mild periods. This method identified significant temporal spectral changes, enabling a refined disease staging framework. Key wavebands associated with sweet potato scab were identified in the near-infrared range, including 801.8 nm, 769.8 nm, 898.5 nm, 796.4 nm, and 780.5 nm. Classification models, including K-nearest neighbor (KNN), support vector machine (SVM), and linear discriminant analysis (LDA), were constructed to evaluate the effectiveness of spectral features. Among these classification models, the MSC-PCA-SVM model demonstrated the best performance. Specifically, the Susceptible Variety Disease Classification Model achieved an overall accuracy (OA) of 98.65%, while the Combined Variety Disease Classification Model reached an OA of 95.38%. The results highlight the potential of hyperspectral imaging for early disease detection, particularly for non-destructive monitoring of resistant and susceptible sweet potato varieties. This study provides a practical method for early disease classification of sweet potato scab, and future research could focus on real-time disease monitoring to enhance sweet potato crop management.