APD-YOLOv7: Enhancing Sustainable Farming through Precise Identification of Agricultural Pests and Diseases Using a Novel Diagonal Difference Ratio IOU Loss
Jianwen Li,Shutian Liu,Dong Chen,Shengbang Zhou,Chuanqi Li
DOI: https://doi.org/10.3390/su16208855
IF: 3.9
2024-10-14
Sustainability
Abstract:The diversity and complexity of the agricultural environment pose significant challenges for the collection of pest and disease data. Additionally, pest and disease datasets often suffer from uneven distribution in quantity and inconsistent annotation standards. Enhancing the accuracy of pest and disease recognition remains a challenge for existing models. We constructed a representative agricultural pest and disease dataset, FIP6Set, through a combination of field photography and web scraping. This dataset encapsulates key issues encountered in existing agricultural pest and disease datasets. Referencing existing bounding box regression (BBR) loss functions, we reconsidered their geometric features and proposed a novel bounding box similarity comparison metric, DDRIoU, suited to the characteristics of agricultural pest and disease datasets. By integrating the focal loss concept with the DDRIoU loss, we derived a new loss function, namely Focal-DDRIoU loss. Furthermore, we modified the network structure of YOLOV7 by embedding the MobileViTv3 module. Consequently, we introduced a model specifically designed for agricultural pest and disease detection in precision agriculture. We conducted performance evaluations on the FIP6Set dataset using mAP75 as the evaluation metric. Experimental results demonstrate that the Focal-DDRIoU loss achieves improvements of 1.12%, 1.24%, 1.04%, and 1.50% compared to the GIoU, DIoU, CIoU, and EIoU losses, respectively. When employing the GIoU, DIoU, CIoU, EIoU, and Focal-DDRIoU loss functions, the adjusted network structure showed enhancements of 0.68%, 0.68%, 0.78%, 0.60%, and 0.56%, respectively, compared to the original YOLOv7. Furthermore, the proposed model outperformed the mainstream YOLOv7 and YOLOv5 models by 1.86% and 1.60%, respectively. The superior performance of the proposed model in detecting agricultural pests and diseases directly contributes to reducing pesticide misuse, preventing large-scale pest and disease outbreaks, and ultimately enhancing crop yields. These outcomes strongly support the promotion of sustainable agricultural development.
environmental sciences,environmental studies,green & sustainable science & technology