Scalable Detection of Wheat Diseases using DenseNet-201: A Transfer Learning Approach

Wheat diseases, namely, Fusarium Head Blight, Powdery Mildew, Leaf Rust (Brown Rust), Stripe Rust (Yellow Rust), pose major harm to the security of the food all over the world, resulting in serious yield losses per annum. Traditional disease detection methods are labor-intensive and lack the ability in scalability for huge-scale agricultural applications. This work adapts the DenseNet-201, which is a Deep Learning model that is initially built for general classification tasks, to detect numerous wheat diseases utilizing an arranged dataset of RGB images composed of healthy wheat leaves and diseased wheat leaves. The DenseNet-201 model accomplished an accuracy of 99.20% in training and an accuracy of 83.00% in validation through transfer learning and fine-tuning. Examination on the validation set had generated a macro-average F1 score of 16% and a weighted precision of 20%, with the Stripe Rust (Yellow Rust) class exhibiting the highest recall at 40%. These outcomes emphasize the model’s capability for scalable and accurate disease identification in precision agriculture. The future work will concentrate on approaching issues of the dataset, on improving the performance of the class level, and enabling real-time deployment in the circumstances of the field.