A Banana Disease Detection Using MobileNetV2 Model Based on Adam Optimizer

The main objective of this study is to develop a deep learning-based disease detection system for banana plants using the MobileNetV2 architecture through a comprehensive comparison with VGG16. This study utilizes a dataset of 3,653 images categorized into 12 classes, including Aphids, Bacterial Soft Rot, Bract Mosaic Virus, Cordana, Insect Pest, Moko, Panama, Fusarium Wilt, Black Sigatoka, Yellow Sigatoka, Pestalotiopsis, and healthy specimens. The methodological framework includes architecture comparison, data balancing, preprocessing techniques, and performance evaluation. The dataset was divided with a distribution ratio of 75% for training, 15% for validation, and 10% for testing. Comparative analysis shows excellent performance of MobileNetV2 with an accuracy of 96.21% compared to 90.15% for VGG16, while maintaining a significantly smaller model size of 10.0 MB compared to 57.8 MB for VGG16. Statistical validation through the McNemar test confirms significant superiority with a p-value of 0.008. The findings of this study contribute positively to the development of agricultural technology, particularly in the development of automated systems for disease detection in banana plants.