Strawberry Plant Diseases Classification Using CNN Based on MobileNetV3-Large and EfficientNet-B0 Architecture

Abstract

Strawberry is a plant that has many benefits and a high risk of being attacked by pests and diseases. Diseases in strawberry plants can cause a decrease in the quality of fruit production and can even cause crop failure. Therefore, a method is needed to assist farmers in identifying the types of diseases in strawberry plants. Currently, there are many methods to assist farmers in identifying types of disease in plants, including strawberry plants. In this study, a system is proposed to be able to detect strawberry plant diseases by classifying the disease based on healthy and diseased strawberry leaf images. The proposed system is the Convolutional Neural Network (CNN) algorithm using MobileNetV3-Large and EfficientNet-B0 models to train pre-processed datasets. The results of this study obtained the best accuracy reaching 92.14% using the MobileNetV3-Large architecture with the hyperparameter optimizer RMSProp, epochs 70, and learning rate 0.0001. The percentage of the evaluation model using MobileNetV3-Large for precision, recall, and F1-Score achieved 92.81%, 92.14%, and 92.25%. Whereas in the EfficientNet-B0 architecture, the best accuracy results only reach 90.71% with the hyperparameter optimizer Adam, 70 epochs, and a learning rate of 0.003. Then, the precision, recall, and F1-scores for EfficientNet-B0 reached 92.65%, 90.00%, and 90.37%. Overall, it presents fairly good results in classifying strawberry leaf plant disease. Furthermore, in future work, it needs to obtain higher accuracy by generating more datasets, trying other augmentation techniques, and proposing a better model.

Plant phenotyping relevance

イチゴ葉画像から植物病害状態をCNNで分類する手法が研究の中心であり、植物の病害表現型を直接推定しているため含める。

Code and data availability

The supplied blocks describe a self-collected strawberry leaf image dataset (1336 images) and CNN training, but contain no public dataset deposit, no author code/model release, and no availability statement. All URLs in the blocks are citations to prior work, not paper-specific assets.

No evidence-backed public reproduction asset is currently recorded.