Abstract:
Java Plum (Syzygium cumini) is a valuable fruit crop in terms of nutritional, medicinal and economic value. Nevertheless, the leaves are very prone to disease that can lower yield and quality which can be a burden on growers. There is thus a need to detect the disease early and accurately to facilitate effective disease management and to maintain sustainability of cultivation. Although recent research on plant disease detection has been based on deep learning models, including convolutional neural networks (CNNs) and Vision Transformers, most methods do not go beyond raw classification performance. There have been only few approaches that use explainability techniques, and, therefore, end-users like farmers and agricultural experts do not have a clear picture on how and why a prediction is delivered. We presented and tested two complementary approaches to Java Plum leaf disease detection in this work: transfer learning using fine-tuning of the previously trained CNNs and hybrid classifiers that combine CNN feature extraction with traditional machine learning classifiers. A publicly available Kaggle dataset of healthy and diseased Java Plum leaf images was used, with rigorous preprocessing and augmentation to enhance generalization. Pretrained models like ResNet50, VGG16, DenseNet121, MobileNetV2, variants of EfficientNet models were tested in pre- and post-fine-tuning. After the results, it was found that fine-tuning offered a strong performance improvement of most models, and EfficientNetB5 reached its highest classification accuracy of 97.9%, making it the most successful architecture with Grad-Cam visualizations showing accurate results as well. Competitive but slightly worse results were achieved with hybrid CNN and Machine Learning Classifier pipelines as compared to the fine-tuned deep learning models. Generally, the findings suggest that hybrid feature extraction models tend to be stronger, yet fine-tuning CNNs models using transfer learning prove to be more precise and addition of explainability methods like Grad-CAM builds confidence and reliability in real-world agriculture. This framework, besides achieving state of art performance, also takes a step toward some practical interpretable solutions to disease monitoring in Java Plum leaf.
