A Robust and Interpretable Ensemble CNN Framework for Plant Leaf Disease Classification

The detection of leaf disease in plants is a crucial area on the way to increasing agricultural productivity. Although deep convolutional neural networks (CNNs) are highly accurate, the predictions of visually similar diseases may vary in single models. In this paper, I have compared several pretrained CNN models, such as VGG16, VGG19, ResNet101V2, InceptionV3, and DenseNet121 on the PlantVillage dataset. The highest standalone accuracy of 99% percent is attained by denseNet121. A softvoting ensemble is also used to enhance the prediction robustness and stability by minimizing the misclassification of the classes. Grad-CAM is applied to visualize the area of the disease and improve the interpretability and trust towards the user. The framework suggested has a valid and interpretable diagnosis of plant diseases decision-support system.