Convolutional Neural Network-Based Image Classification of Bacterial Wilt Symptoms in Pepper (Capsicum annuum L.)

초록

Bacterial wilt, caused by Ralstonia solanacearum, is a destructive disease that affects pepper and other solanaceous crops. Reliable phenotypic evaluations are essential for resistance screening, but conventional visual scoring can be affected by observer-dependent variations. In this study, we developed an image classification framework based on a convolutional neural network (CNN) for objective assessments of bacterial wilt symptoms in pepper. Three pepper genotypes representing resistant (MC4), moderately resistant (ECW30R), and susceptible (Subicho) responses were inoculated with R. solanacearum SL1931, after which symptom development was evaluated using a five-level disease index. Under standardized imaging conditions, 1,500 multi-view symptom images were acquired and assigned to three symptom-associated resistance classes. To reduce information leakage, images obtained from the same plant individual were assigned to the same dataset partition. The model was trained using the training set, monitored with the validation set, and evaluated using an independent test set. The final model correctly classified 266 of 300 test images, corresponding to an accuracy rate of 88.67%. Misclassifications occurred mainly between adjacent symptom classes, suggesting partial visual overlap between intermediate and severe wilt responses. These results indicate that the CNN-based image analysis method developed here can support objective and reproducible bacterial wilt phenotyping in pepper. Further validation across broader germplasm, pathogen isolates, and imaging environments will be required for large-scale breeding applications.

키워드