Table 1_Early detection of sesame leaf disease using hyperspectral imaging and machine learning: a patch-based spatial-spectral integration approach.docx

IntroductionSesame (Sesamum indicum L.) is highly susceptible to leaf blight caused by Corynespora cassiicola and bacterial leaf spot caused by Pseudomonas syringae, both of which can cause substantial yield and quality losses. Conventional detection methods are limited to post-symptom diagnosis, limiting opportunities for timely intervention. Early asymptomatic detection is essential for stable production and economic sustainability.
MethodsWe developed a hyperspectral imaging framework integrated with machine learning algorithms for early sesame leaf diseases detection by implementing a 3 × 3 patch-based approach to capture spatial–spectral disease patterns. Hyperspectral images (450–950 nm) were acquired using dedicated hyperspectral imaging equipment. The 3 × 3 patch extraction method captured spatial–spectral disease progression patterns, including infected areas and surrounding healthy tissues, to enable effective early disease detection. Three preprocessing methods [Raw reflectance spectra, Savitzky–Golay (SG) smoothing, and 1st derivative] were evaluated. Wavelength selection was performed using the Sequential Projection Algorithm (SPA), Genetic Algorithm (GA), and Competitive Adaptive Resampling (CARS). Feature extraction included three spectral features (mean, standard deviation, and range) and five vegetation indices [Normalized Difference Vegetation Index (NDVI), Red Chlorophyll Index (RCI), Photochemical Reflectance Index (PRI), Anthocyanin Reflectance Index (ARI), and Plant Senescence Reflectance Index (PSRI)]. Each preprocessing–effective wavelength combination was applied to Support Vector Machine (SVM), Partial Least Squares Discriminant Analysis (PLS-DA), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) models.
ResultsComparative analysis revealed that the SG-SPA-SVM and Raw-GA-SVM models achieved the highest classification performance, each with 97.8% accuracy on the test set. The proposed framework enabled early asymptomatic diagnosis as soon as 1 day post-inoculation, preceding visual symptom development.
DiscussionIntegrating hyperspectral imaging with patch-based spatial–spectral analysis and machine learning enables accurate, early, and pathogen-specific detection of sesame leaf diseases. The reduced wavelength requirements of the GA-based approach improve feasibility for lightweight, cost-effective multispectral sensor development, thereby supporting sustainable sesame production and precision agriculture.