A study on the non-contact measurement of sunflower disk inclination and its application to accurate phenotypic analysis

Qiang WangKaixuan LiZihao GaoXinyuan WeiYaoyu LiYangcheng LvWuping Zhang^*

• Shanxi Agricultural University, Jinzhong, China

The tilt angle of sunflower flower heads is an important phenotypic characteristic that influences their growth and development, as well as the efficiency of mechanical harvesting. Traditional manual measurement methods suffer from low accuracy, high costs, and the risk of damaging plants. This study proposes a non-contact measurement method based on deep learning and geometric analysis. By optimizing the lightweight YOLO11n-seg model, it achieves efficient instance segmentation of flower heads and stems (recall rate improved by 3.7%, mAP50 improved by 1.8%, parameters reduced by 0.29M, and computational load reduced by 0.5 GFLOPs). The mask map output by the model is used to extract the contours of the flower head and stem. Ellipse and curve fitting are then used to determine the main axis direction of the flower head and the tangent direction of the stem, respectively, and the angle between the two is calculated to quantify the tilt angle. This method establishes a collaborative workflow from instance segmentation to geometric modeling, providing precise perception information for path planning, grasping posture adjustment, and end-effector positioning in harvesting robots. It serves as a critical automated measurement module in sunflower precision agriculture. Due to the lack of publicly available standard datasets, this study was validated on 220 images, comparing the algorithm results with the actual values measured using a manual protractor. The results were RMSE = 2.93°, MAE = 2.43°, and R² = 0.94. This method significantly improves measurement efficiency while maintaining measurement accuracy, without requiring contact with the plant, and demonstrates good adaptability and application value.