A novel YOLOv11 framework for enhanced tomato disease detection

Abstract

Plant diseases severely threaten global agriculture, causing significant crop losses and jeopardizing food security. Traditional manual diagnostic methods are inefficient, time-consuming, and prone to human error, underscoring an urgent need for accurate, efficient, and scalable automated detection systems. While deep learning offers transformative potential, existing models often contend with high computational demands, limited scalability, and insufficient robustness for real-world agricultural deployment. This article presents a novel and highly efficient framework leveraging the cutting-edge You Only Look Once (YOLO)v11 architecture, enhanced with a sophisticated Attention-Guided Multi-Scale Feature Fusion (AGMS-FF) Enhancer, for the precise classification of 10 distinct diseases affecting tomato plants, alongside healthy specimens. Our proposed AGMS-FF module meticulously refines feature representations by integrating multi-scale convolutional paths with both channel and spatial attention mechanisms, all supported by residual connections to improve feature learning and model stability. The framework was rigorously evaluated on the extensive Zekeriya Tomato Disease Model dataset, comprising 42,606 annotated images (4,260 in the test set). Our enhanced YOLOv11 model achieved an outstanding overall accuracy of 99.93%, demonstrating exceptional performance across all disease classes, with many reaching perfect 100.00% precision, recall, and F1-scores. A comprehensive ablation study confirmed the efficacy of the AGMS-FF components, showing that while the baseline YOLOv11 already achieved near-perfect accuracy, the enhanced variants maintained this high level of performance with slightly varied metrics ( e.g ., 99.84% accuracy for full AGMS-FF), underscoring the robust and stable nature of our additions even at performance saturation points. Furthermore, the model exhibited excellent computational efficiency, with a training duration of 126 min, inference time of 31.4 ms, memory usage of 3.2 GB, and a throughput of 38.2 FPS. These results collectively establish a new state-of-the-art in tomato disease classification, providing a powerful, accurate, and computationally practical solution. The developed framework significantly bridges the gap between advanced deep-learning research and practical agricultural deployment, offering real-time diagnostic capabilities essential for enhancing crop health, optimizing yields, and bolstering global food security.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.