LeafDet: A lightweight and interpretable deep learning framework for tomato leaf disease detection

Abstract

Ensuring global food security depends on timely and reliable plant disease identification. Traditional disease detection methods often prove inefficient because of the lack of necessary precision. Furthermore, public datasets typically suffer from the class imbalance issue, which can obstruct reliable model testing and lead to biased performance evaluations. This paper introduces LeafDet, an object detection model based on the YOLOv8 architecture, specifically designed for the effective detection of tomato leaf diseases. Moreover, a revised, balanced dataset, named PlantTom, is developed by combining images from various public sources to reduce the existing dataset limitations. PlantTom has 7836 images with 8 distinct classes, each representing a tomato leaf disease. The proposed LeafDet model includes CBM, C2f, SPPF, and ECA attention modules in the backbone section; BiFPN, GSConv, VoVGSCSP, and Shuffle Attention in the neck section. Efficient attention methods like ECA and Shuffle Attention are used to improve both accuracy and speed. LeafDet model achieves 91.6% mAP@0.5 on the PlantTom dataset, which is a 2.2% improvement over the original YOLOv8n with 2.69M parameters and an inference time of 2.4ms. The proposed model also outperforms several other state-of-the-art object detection models, including the latest YOLOv11n and YOLOv12n. Ablation studies show that each part of the model helps to improve its performance, and the PIoUv2 loss function is found to be the optimal choice for this use. The model predictions are then validated using Eigen-CAM, which provides a visualization of the decision-making process. These results demonstrate that LeafDet provides a deployable and interpretable framework for plant disease detection in smart agriculture.