Transfer Learning-Based Multiclass Classification of Pea Diseases from Leaf Images

Abstract

Accurate detection of plant leaf diseases is an important need to support early decision-making in agriculture, especially in conditions where datasets are limited and have high visual variation. This study aims to evaluate the performance of the transfer-learning-based deep learning approach in the multiclass classification of pea leaf disease and determine the most optimal Convolutional Neural Network architecture. The research method used pea leaf imagery from open source with five ImageNet-based pre-trained CNN architectures, namely ResNet50, EfficientNet B0, DenseNet121, MobileNetV3 Large, and ConvNeXt Tiny, which were trained using the same preprocessing configuration, data augmentation, dataset sharing, and training strategies. Performance evaluation was carried out using accuracy metrics, macro precision, macro recall, macro F1 score, confusion matrix, and ROC one vs rest curve. The results showed that ConvNeXt Tiny provided the best performance with an accuracy value of 90.05 percent and a macro F1 score of 0.9137 as well as strong discriminating ability in all disease classes. The conclusions of this study show that the transfer-learning-based deep learning approach is effective and reliable for the multiclass classification of pea leaf disease, as well as that the selection of the right CNN architecture is highly influential in achieving optimal performance on small-scale agricultural datasets with high visual complexity.