Cassava Diseases Classification using EfficientNet Model with Imbalance Data Handling

Abstract

This research highlights the urgent need for classifying cassava diseases into five classes, such as Cassava Bacterial Blight (CBB), Cassava Brown Streak Disease (CBSD), Cassava Green Mottle (CGM), and Cassava Mosaic Disease (CMD), and Healthy. The study proposes the utilization of the EfficientNet model, a lightweight deep learning architecture, for classifying cassava diseases based on leaf images. However, the datasets available for this classification task are all unbalanced, made it difficult for researchers to perform. To tackle this imbalance issue, the authors compared several imbalance data handling methods commonly used for image classification, including SMOTE (Synthetic Minority Oversampling Technique), basic augmentation, and neural style transfer, to be applied before fed into EfficientNet. Initially, EfficientNet model without addressing dataset imbalances, the F1-Score stands at 78%, with most images misclassified into the majority class. Integration with SMOTE notably boosts the F1-Score to 82%, showcasing the efficacy of oversampling methods in enhancing model performance. Conversely, employing data augmentation, both basic and deep learning-based, lowers the F1-Score to 74% and 65% respectively, yet it results in a more balanced distribution of true positives across disease classes. The findings suggest that SMOTE surpasses the other methods in handling imbalanced data.