Comparative Analysis of YOLOv8 and MobileNetV2 for Digital Image-Based Bird's Eye Chili Leaf Disease Classification

Background: Bird’s eye chili is an important horticultural commodity in Indonesia but is highly vulnerable to leaf diseases such as mosaic virus, leaf spot, and chlorosis. Visual disease identification is often inaccurate and can delay treatment, while previous deep learning studies have mainly used laboratory-acquired datasets with limited applicability to field conditions. Objective: This study compares the performance of YOLOv8 classification (YOLOv8-cls) and MobileNetV2 in classifying bird’s eye chili leaf diseases using digital images acquired under field conditions. Methods: A dataset of 800 leaf images representing four classes (healthy, mosaic virus, leaf spot, and chlorosis) was collected under natural field conditions. The images were preprocessed and used to train both models through transfer learning using ImageNet-pretrained weights. Performance was evaluated using stratified 5-fold cross-validation and measured using accuracy, precision, recall, F1-score, and the Matthews correlation coefficient (MCC). Results: Both models achieved average accuracies above 98%. YOLOv8-cls achieved the best overall performance, with 98.38% accuracy, 98.55% precision, 98.35% recall, and 98.41% F1-score, while requiring only 2.84 MB of storage, making it suitable for deployment on edge devices. MobileNetV2 achieved 98.25% accuracy, 98.36% precision, 98.16% recall, and 98.24% F1-score, while reducing training time by approximately 40%. Conclusion: YOLOv8-cls is the more suitable architecture for mobile and edge-based plant disease detection because of its superior classification performance and compact model size, whereas MobileNetV2 is advantageous when rapid model retraining is required. These findings provide a practical basis for developing accessible digital agriculture applications for early disease detection.