Abstract:

Soybean plant physiological height, as one of the significant phenotypic traits, holds paramount importance in soybean breeding for selection and improvement purposes. Accurate measurement of soybean plant physiological height stands as a crucial task in soybean cultivation due to its pivotal role in variety evaluation. Currently, manual measurement methods suffer from inefficiency and error-proneness, rendering them inadequate for large-scale soybean plant measurement tasks. With the advancement of computer vision technology, an increasing number of studies have been devoted to utilizing computer vision algorithms for automated measurement and analysis of soybean plant physiological height. For instance, in the field of object detection, approaches often rely on measuring the straight-line length of plant endpoints to determine plant height. However, the natural curvature of soybean plants may introduce significant measurement errors using this method. Furthermore, instance segmentation techniques are employed to segment plant pixels for height measurement. Nevertheless, annotating segmentation data for entire plants entails high costs and is challenging to achieve. This study proposes a novel approach utilizing an improved YOLOv8n model to reconstruct the curved main stem skeleton of soybean plants, enabling high-precision measurement of soybean plant physiological height. Based on the original YOLOv8n model, the Backbone section has been augmented with CA (Coordinate Attention), while a new small object detection layer has been added to the Neck and Head sections. This enhancement improves the ability to detect main stem nodes of soybean plants, enabling accurate localization of the main stem nodes. Subsequently, the YOLOv8n-seg model is employed for root segmentation, obtaining the positions of root-stem junction points, thereby eliminating the influence of root length on plant height measurement. Based on the growth direction of the plants and combining the position information of the main stem nodes and the root-stem junction points, the main stem skeleton of soybean plants is constructed. An edge contour detection algorithm is used to detect the straight ruler reference object in the image, obtaining the actual length corresponding to the pixels. Finally, by utilizing the main stem skeleton information that accurately fits the curved morphology of soybean stems, the plant height measurement is completed, thereby improving the accuracy of plant height. The experimental findings demonstrate that the improved YOLOv8n model yields a mean average precision of 91.52%, marking a notable improvement of 2.09 percentage points over the original network. Meanwhile, the YOLOv8n-seg model achieves an impressive mean average precision of 95.54%, enabling the precise reconstruction of soybean plant main stem skeletons. Using both methods to detect soybean plant images, the results indicate that reconstructing the soybean main stem skeleton achieves a mean absolute error of 1.67 cm, a root mean square error of 1.91 cm, and an mean absolute percentage error of 3.25% in plant height measurement. Compared with the length measurement of the detection box for the entire plant, there is a significant decrease of 8.46 percentage points in the mean absolute percentage error. This signifies a notable enhancement in measurement accuracy. The experimental results were analyzed and compared with common methods for measuring plant height, leading to the conclusion that this approach surpasses the commonly used methods in the field of computer vision for physiological plant height measurement. The research findings demonstrate that this method yields accurate measurements of soybean plant physiological height, providing methodological and technological support for intelligent soybean breeding.

Key words: curvature, soybean, main stem skeleton, object detection, plant height measurement