A fast and lightweight detection model for wheat fusarium head blight spikes in natural environments

Abstract

Fusarium head blight (FHB) is one of the most common and destructive infections in wheat, posing a serious threat to food security and human health. Real-time detection and severity assessment of wheat FHB are crucial for effective management and loss assessment. In this study, we proposed a novel and advanced YOLOv5s model for fast and accurate detection of wheat FHB. Firstly, we replaced the original backbone with MobileNetV3 and integrated spatial pyramid pooling-fast (SPPF). Subsequently, we replaced the C3 modules with the C3Ghost to reduce the parameters and complexity of this model without compromising detection performance. The proposed model was trained and evaluated on the wheat FHB dataset, achieving a high mean average precision (mAP) of 97.15 % using only 3.64 M parameters and 4.77 G floating-point operations (FLOPs). Compared with the original YOLOv5, these values have decreased by 49.72 % and 71.32 %, respectively. To assess the severity of FHB damage, the diseased spike rate was calculated and the coefficient of determination (R2) was 0.9802, achieving a satisfactory statistical effect. These results validate the effectiveness of the improved YOLOv5s model for real-time detection of wheat FHB spikes. This study contributes to the development of accurate and efficient plant disease detection systems and provides a valuable reference for accurate quantitative assessment of crop losses, thus ensuring food security.