Advancing Precision Agriculture: Enhanced Weed Detection Using the Optimized YOLOv8T Model

Abstract

Precision agriculture relies on effective weed management for high yields and crop quality. Deep learning (DL)-based techniques show potential for providing effective solutions. However, their practicality is sometimes limited by insufficient datasets. Our research has utilized a comprehensive instance-level annotated weed dataset derived from existing agricultural imagery to address this critical gap. This dataset encompasses various weed and crop species, with images featuring detailed bounding box annotations to mark individual instances. This refinement facilitates the application of advanced DL models by providing more granular, real-world training data. Utilizing this dataset, we extensively evaluated the latest object detection models, focusing on the YOLO series, including YOLOv7, YOLOv8 variants, and our newly proposed YOLOv8T model. Our findings reveal that the YOLOv8T model surpasses its predecessors, achieving a mean average precision (mAP) of 82.5%. This notable improvement underscores the model’s enhanced capability to accurately distinguish between crop and weed species. Moreover, our study delves into the impact of data augmentation techniques to mitigate class imbalance within the dataset, further elevating the YOLOv8T’s performance metrics. These techniques improved the mAP results and showed how DL models, especially the YOLOv8T, can improve weed detection systems in the field. Through rigorous testing and analysis, our research confirms the viability of the YOLOv8T model as a cornerstone for developing automatic, efficient, and scalable weed detection systems.