An effective framework based on deep learning (DL) is developed in this study for reliable and accurate performance. The multi-class detection of crops such as corn, tomato and potato is accurate and reliable. The aim is to improve early disease detection which guarantees good classification accuracy, strong generalization across datasets, enhanced interpretability through XAI methods and enabling realistic agricultural applications. : The study uses two large publicly available datasets of plant leaf disease images of corn, tomato and potato. The first dataset designated as D-I consists of 39,203 images belonging to 18 classes of disease and the second dataset designated as D-II consists of 65,565 images also belonging to 18 classes of disease. The two datasets contain images showing different visual scenarios and variations of disease which will help form a multi-class classifier. In this study, five DL architectures were used; InceptionNetV3, ResNet152V2, ViT, BERT, and the proposed hybrid model (ResViT-152) that combined convolutional feature extraction with transformer-based global attention. Every model was trained, validated, and tested under the same experimental setup. Cross validation and multi-phase testing assessed their performances in their capacity to learn discriminative parameters in corn, tomato and potato disease classes. The hybrid model exhibited a better performance in all test conditions. In IntraTest1, the accuracies were 99.12%, 98.94% and 99.06% for corn, tomato and potato respectively. In IntraTest2, the model achieves accuracy of 99.23% for corn, 98.97% for tomato, and 98.98% for potato on D-II. The precise percentages for the cross-tests were 96.27% (corn), 95.14% (tomato), 95.06% (potato) for CrossTest1 and 95.77% (corn), 96.22% (tomato), 96.15% (potato) for CrossTest2. The performance across datasets for all three crops is good and generalization is robust. A study is conducted on an efficient, high-performing, and interpretable DL framework for plant leaf disease detection. The experimental results verify that the proposed work provides superior performance. Generalization outperforms standard architectures in effectiveness. In addition, there is also stable performance with varying datasets. With Explainability included, the model becomes more transparent and a strong candidate for further validation toward deployment in precision agriculture, pending evaluation on real-world field datasets.
Srinivasan et al. (Sun,) studied this question.