Abstract:

[Objective] Accurate extraction of maize plant skeletons and phenotypic parameters is fundamental for acquisition of plant growth data, morphological analysis, and agricultural management. However, leaf occlusion and complex backgrounds in dense planting environments pose significant challenges to skeleton and parameters extraction. A maize plant skeleton and phenotypic parameters extraction method suitable for dense field environments was proposed in this research to enhance the extraction precision and efficiency, and provide technical support for maize growth data acquisition. [Methods] An improved YOLOv11-Pose multi-object keypoint detection network was introduced, a top-down detection framework was adopted to detect maize plant keypoints and reconstruct skeletons. A uniform sampling algorithm was used to design a keypoint representation method tailored for maize skeletons and optimize task adaptability. Additionally, a single-head self-attention mechanism and a convolutional block attention module were incorporated to guide the model's focus on occluded regions and connected parts, thereby improve its adaptability to complex scenarios. [Results and Discussion] In dense field maize environments, experimental results showed that when the number of uniformly sampled keypoints was set to 10, the Fréchet distance reached its minimum value of 79.008, effectively preserving the original skeleton's morphological features while avoiding the negative impact of redundant points. Under this configuration, the improved YOLOv11-Pose model achieved a bounding box detection precision of 0.717. The keypoint detection mAP50 and mAP50-95 improved by 10.9% and 23.8%, respectively, compared to the original model, with an inference time of 52.7 ms per image. The results demonstrated the model's superior performance and low computational cost in complex field environments, particularly in keypoint detection tasks with enhanced accuracy and robustness. The study further combined the results of skeleton extraction and spatial geometric information to achieve a plant height measurement mean average error (MAE) of 2.435 cm, the detection error of leaf age was less than one growth period, and the measurement error of leaf length was 3.482%, verifying the effectiveness and practicability of the proposed method in the application of phenotypic parameter measurement. [Conclusion] The proposed improved YOLOv11-Pose model can efficiently and accurately extract maize plant skeletons, meeting the demands of ground-based maize growth data acquisition. The research could provide technical support for phenotypic data acquisition in grain production and precision agricultural management.

Key words: crop growth, keypoint detection, attention mechanism, phenotypic parameter, maize plant skeleton, YOLOv11