YOSBG: UAV image data-driven high-throughput field tobacco leaf counting method

Abstract

Abstract Background: Estimating tobacco leaf yield is a crucial task. The number of leaves is directly related to yield. Therefore, it is important to achieve intelligent and rapid high-throughput statistical counting of field tobacco leaves. Unfortunately, the current method of counting the number of tobacco leaves is expensive, imprecise, and inefficient. It heavily relies on manual labor and also faces challenges of mutual shading among the field tobacco plants during their growth and maturity stage, as well as complex environmental background information. This study proposes an efficient method for counting the number of tobacco leaves in a large field based on unmanned aerial vehicle (UAV) image data. First, a UAV is used to obtain high-throughput vertical orthoimages of field tobacco plants to count the leaves of the tobacco plants. The tobacco plant recognition model is then used for plant detection and segmentation to create a dataset of images of individual tobacco plants. Finally, the improved algorithm YOLOv8 with Squeeze-and-Excitation (SE) and bidirectional feature pyramid network (BiFPN) and GhostNet (YOSBG) algorithm is used to detect and count tobacco leaves on individual tobacco plants. Results: Experimental results show YOSBG achieved an average precision (AP) value of 93.6% for the individual tobacco plant dataset with a model parameter (Param) size of only 2.5 million (M). Compared to the YOLOv8n algorithm, the F1 (F1-score) of the improved algorithm increased by 1.7% and the AP value increased by 2%, while the model Param size was reduced by 16.7%. In practical application discovery, the occurrence of false detections and missed detections is almost minimal. In addition, the effectiveness and superiority of this method compared to other popular object detection algorithms have been confirmed. Conclusions: This article presents a novel method for high-throughput counting of tobacco leaves based on UAV image data for the first time, which has a significant reference value. It solves the problem of missing data in individual tobacco datasets, significantly reduces labor costs, and has a great impact on the advancement of modern smart tobacco agriculture.