A Comparative Study of Deep Transfer Learning Architectures for Multi-Class Plant Leaf Disease Detection

Abstract

Plant Leaf Diseases represent a significant risk to global agricultural production. Crops ranging from Peppers and Tomatoes to Potatoes are affected by these diseases. Traditional methods of identifying leaf diseases based primarily on visual inspection have historically been slow and relatively inaccurate. A deep learning-based solution for the automated identification of plant leaf diseases utilizes Convolutional Neural Networks (CNNs) as the primary methodology. To identify leaf images into 15 disease categories, both pre-trained models such as VGG16, EfficientNetB3, Inception V3 and a custom CNN model were used. Using pre-trained models to allow for the use of Transfer Learning helps to mitigate some of the issues associated with computational resource limitations and data limitations in providing faster convergence rates and higher accuracy when compared to training a model from scratch. The Plant Village image collection which contains over 20,000 images of different plant leaf diseases was utilized for training and testing purposes. Each model's performance was evaluated based on its accuracy, loss and generalization capabilities. Additionally, each model was fine-tuned through hyperparameter optimization. As a result, the model that achieved the highest validation accuracy rate of 95% was the EfficientNetB3 model while the second highest accuracy rate was achieved by the Inception V3 model at 92%. This methodology provides an excellent answer to addressing early disease detection, enabling farmers to take the necessary actions quickly to reduce their losses and maximize their harvest.

Plant phenotyping relevance

植物葉画像から病害状態を推定する深層学習手法を比較・評価しており、病害フェノタイピングと手法検証が研究の中心です。

Code and data availability

The paper's plant-phenotyping measurements (CNN classification of 15 leaf disease classes) were performed on the public PlantVillage-derived Kaggle 'Plant Disease Dataset' by E. Marrex, which the authors cite as their training/testing image source. No author analysis code, trained model checkpoints, or paper-specific衍生