Transformative Approaches to Wheat Disease Detection: An Efficient Deep Learning Framework for Enhanced Crop Resilience

Abstract

Wheat is the third most consumed grain globally and a critical staple food for billions of people; yet its production is severely impacted by fungal diseases that cause annual yield losses of 15% to 20%, posing a direct threat to global food security and agricultural economics. Traditional methods of disease detection, reliant on manual field inspections, are inefficient, labour-intensive, require expertise that is not always available, and are prone to subjectivity—motivating the need for automated, reliable, and scalable solutions for early and accurate wheat disease identification. This study presents a controlled comparative evaluation of four CNN-based pre-trained architectures—DenseNet121, ResNet50, VGG19, and InceptionV3—for wheat disease classification using the Large Wheat Disease Classification Dataset (LWDCD2020), employing targeted data augmentation strategies to improve model robustness and generalizability without overfitting. The results reveal that InceptionV3 achieved the highest performance with a Macro F1-score of 96.75%, outperforming all other proposed models and existing approaches in the literature, while requiring only ten training epochs compared to 50–200 epochs in prior studies. The introduction of the Revolutions Per Inference (RPI) metric further quantifies the efficiency–accuracy trade-off, demonstrating InceptionV3’s practical suitability for deployment in resource-constrained agricultural environments.