Early Detection of Pepper Wilt Disease Using Unmanned Aerial Vehicle-Based Multispectral Time-Series Imagery

This study evaluates early detection of pepper wilt disease using vegetation indices obtained from unmanned aerial vehicle (UAV)-based multispectral time-series imagery and machine learning classification. Multispectral images were acquired from June-August 2023, and disease status was determined by field diagnosis and visual interpretation of the imagery using diagnosed plants as references. Nine vegetation indices were used to develop k-nearest neighbors (KNN), support vector machine, and logistic regression (LR) models. For the June and July datasets, the best model was LR, achieving validation accuracies of 0.860 and 0.925, recall of 0.850 and 1.000, and F1-scores of 0.850 and 0.930, respectively. The best combined-dataset model was KNN; however, it did not outperform the monthly dataset models. The difference vegetation index showed the highest variable importance in the two monthly dataset models, and the variable importance of the normalized difference vegetation index was markedly higher in the July model. Similar trends were observed in the histogram analysis of vegetation indices and spectral features. Thus, date-specific models reflecting crop growth and disease progression were more effective than pooled observations across dates. Overall, UAV-based multispectral time-series imagery provides a practical basis for the field-scale early detection and prompt management of pepper wilt disease.