Paper Details
Abstract
Accurate and robust classification of plant leaf diseases is vital for timely intervention in agricultural management systems. While existing convolutional neural network (CNN) models, such as YOLOv8, have shown competitive performance in plant disease classification, their vulnerability to distributional shifts and image noise limits deployment in real-world farming environments. In this study, we propose an enhanced YOLOv8n classification model integrated with Efficient Channel Attention (ECA) to strengthen feature recalibration and improve model generalisation under varying visual conditions. The ECA module is embedded directly after the Spatial Pyramid Pooling-Fast (SPPF) layer to amplify informative channels prior to global pooling and final classification. We evaluate the proposed model on a nine-class plant leaf disease dataset using three test scenarios: clean (original), moderately augmented (Mid Enhanced), and heavily perturbed (Heavy Real-World) conditions. On the original dataset, the ECA-enhanced YOLOv8n achieves a Top-1 accuracy of 99.28% and Top-5 accuracy of 99.94%, comparable to the baseline. Under moderate noise, our model outperforms the baseline with a Top-1 of 61.94% versus 60.74%. In the most challenging test condition, characterised by severe weather simulation, occlusions, and lighting variations, the proposed model demonstrates superior robustness with a Top-1 accuracy of 37.68%, compared to 35.27% from the original YOLOv8n. These results underscore the benefit of lightweight channel attention in enhancing model resilience to domain shifts, highlighting the potential of our approach for deployment in practical, resource-constrained agricultural settings.