Automatic Detection of Pumpkin Leaf Diseases Using Transfer Learning and a Custom Dataset from Bangladesh

Pumpkin plays a significant role in Bangladeshi agriculture, serving as both a vital food source and a component of the local economy. In this work, we have created a new dataset of pumpkin leaves and trained Convolutional Neural Network (CNN) models for pumpkin leaf disease detection. The dataset contains 4000 images, illustrating various pumpkin diseases including Downy Mildew, Powdery Mildew, Mosaic Disease, Bacterial Leaf Spot, and images of healthy leaves. We trained four CNN models on this dataset—EfficientNet, MobileNet, DenseNet, and ResNet50—to categorize the pumpkin leaf diseases. Data augmentations were implemented to improve the strength of the models. The accuracy of ResNet50, EfficientNet, MobileNet, and DenseNet for classifying pumpkin leaf diseases was found to be 74.68%, 79.74%, 95.44%, and 97.82%, respectively, with DenseNet demonstrating superior accuracy. By identifying the most efficacious model, this research contributes to the advancement of efficient and adaptable solutions for plant disease diagnosis, ultimately aiming to enhance crop management and productivity in Bangladesh.