Abstract:
This study investigates the potential of computer vision approach for classifying banana
leaf diseases - Cordana, Pestalotiopsis, Sigatoka in the context of regional variations. We
collected a dataset of banana leaf images from Habiganj, Sylhet, Bangladesh, to account
for these variations and employed data augmentation techniques to enrich its size and
diversity. Five pre-trained CNN architectures (VGG19, VGG16, ResNet50,
MobileNetV1, MobileNetV2) were evaluated for their disease classification performance.
The evaluation compared the models' performance with and without image
preprocessing. Our findings highlight the exceptional performance of MobileNetV2,
achieving an impressive 94.41% accuracy on raw, unprocessed data. This accuracy
further improved to 96.41% after image preprocessing, demonstrating the model's
robustness and resilience to variations that might be encountered in real-world
applications. These results emphasize the potential of CNNs, particularly lightweight
architectures like MobileNetV2, for accurate banana leaf disease classification using
computer vision. This study provides valuable insights for developing future plant disease
identification systems, ultimately contributing to improved disease management and
promoting sustainable agricultural practices.
