Image-Based Plant Leaf Disease Detection using Deep Learning

Abstract-Deep learning (DL) has recently gained a wide attention in detection of plant leaf diseases. However, the effectiveness of DL depends on the vast amount of data to derive the meaningful patterns which is crucial for accurate detection. This led to the appreciation of pretrained models, which are trained on large data for improved detection despite limited data in the required domain. This paper presents an ensemble-based DL model for detecting diseases in cucumber plant leaves. The proposed ensemble learning framework includes Resnet V50, MobileNetV2 and EfficientNet-B0 as the pretrained models which aggregate the output of these models by weighted ensemble averaging. Data augmentation and Hyperparameter tunings were performed to increase the model generalization and capability. The proposed ensemble approach was validated on Cucumber Disease Recognition Dataset and achieved an accuracy of 99.351%, precision of 99.363% and F1 Score of 99.352% that outperforms the underlying learning framework. Additionally, to provide more insights into the extent of a model to predict each class, Gradient-weighted Class Activation Mapping (Grad-CAM) is employed that produces heat maps highlighting the regions which contributed most to the prediction.