<p>Leaf disease identification and classification are important in agricultural crop production and food protection. Traditional models struggle to capture spatial hierarchies and lack interpretability. The proposed CapsNet-XAI model is based on a capsule network and explainable AI techniques for leaf disease identification and classification. The model uses a dynamic routing-by-agreement to maintain part–whole relationships in leaf disease structures, which enable the disease identification across varying spatial orientations. The model also follows the squash function and the votes computation. The normalization of capsule outputs encodes feature presence probabilities effectively. We also ensure dimensional consistency during routing, strengthening the model’s ability to generalize. The model is trained and tested the efficiency on four heterogeneous datasets: PlantVillage1, PlantDoc, Mango, and PlantVillage. The model decision-making is strengthened by XAI techniques (LRP, Grad-CAM, LIME, Integrated Gradients, SmoothGrad, and Guided Backpropagation). The model shows accuracy of 98.95%, 98.01%, 99.46%, and 98.76%, on respective datasets. The cross-dataset generalization also helps to test model interpretability. The proposed framework to achieve state-of-the-art performance also delivers interpretable insights, making it a valuable tool for real-world agricultural disease management.</p>