Abstract:
Field peas are a significant legume crop cultivated worldwide for their nutritional value and
sustainable agriculture. They are vulnerable to diseases such as ascochyta blight, fusarium wilt,
powdery mildew, etc. which significantly impact yield quality and quantity. Lack of knowledge
about disease, pesticides, and insecticides to monitor disease in communities hinders the economy,
brings yield loss for farmers, and food loss for consumers. The conventional mechanism of disease
identification predominantly depended on the necked eye, manual observation, and subjective
judgment, which is expensive, laborious, and prone to inaccuracies. This study aimed to develop
an explainable AI-based deep learning approach for field pea leaf disease classification. We
followed an experimental research design through a mixed research approach, notably utilizing
CNN and transfer learning combined with Random Search hyperparameter optimization. In the
transfer learning approach, we utilized InceptionV3, VGG16, VGG19, ResNet50, DenseNet201,
and DenseNet121, then we applied explainable AI techniques to CNN and DenseNet121 for model
explainability. We employed SHAP, and LIME Explainer for more understanding of the model’s
decision. Additionally, we investigated the DCGAN generative model to enhance training data by
generating synthetic images from existing data for better performance. Utilizing these approaches,
most models achieved accuracy beyond 97%. Notably, the proposed CNN model gained a
commendable accuracy of 95.8% and improved to 97.5% accuracy by utilizing synthetic images.
DenseNet121 showcased an exceptional accuracy of 99.6%. The model has great significance for
producers, and agricultural experts providing an automated and rapid classification of diseases
and making informed decisions regarding crop cultivation, and marketing. Depending on
experimental results exploring the XAI-aided DL approach to identify field pea leaf disease is
worthwhile for future endeavors.
