基于表型机器人的小麦关键生育期表型检测方法

Phenotyping Identification Method for Key Wheat Growth Stage Based on Phenotyping Robot

摘要:

为解决传统田间小麦表型数据采集与解析自动化水平低和精准性差的问题，研制了小麦表型机器人底盘，并提出一种基于表型机器人的小麦关键生育期表型检测方法。首先，提出了TD-YOLO v11出苗检测模型，实现了田间小麦出苗精准识别。该模型在特征提取网络中引入可变性卷积模块（Deformable convolutional v4, DCNv4），增强模型捕捉上下文信息的能力，降低计算复杂度和参数量。此外，引入任务动态对齐检测头（Task dynamic align detection head, TDADH），通过动态选择特征，提高模型的分类和定位性能。然后，构建了融合多传感器与边缘计算的小麦表型解析系统，该系统集成了出苗检测方法与前期研究提出的抽穗期监测及开花期判定等表型解析方法，实现了田间表型数据的高效自动化采集与解析。结果表明，提出的方法具有较高的小麦出苗识别精度（R2为0.908，RMSE为11.73，rRMSE为23.04%），同时实现小麦抽穗期及开花期表型的动态监测。该方法可用于田间小麦表型数据的高通量采集和高效解析，为小麦育种田间表型获取工作提供了高效、可靠的技术支持。

Abstract:

Aiming to address the challenges of low automation and poor accuracy in traditional field wheat phenotyping data collection and analysis, a wheat phenotyping identification robot chassis was developed and a phenotypic detection method for key wheat growth stages was proposed based on a phenotyping robot. Initially, a TD-YOLO v11 seedling detection model was proposed to achieve automated and precise recognition of wheat seedling emergence in the field. The incorporation of the DCNv4 module into the feature extraction network enhanced its ability to capture contextual information, allowing for the extraction of feature representations with fewer network parameters, thereby reducing computational complexity and the number of parameters. Moreover, the introduction of a task dynamic alignment detection head further utilized information from intermediate layers, promoting consistency between classification and localization tasks, and improving the model’s classification and localization performance during seedling detection. Subsequently, a phenotyping identification system for wheat was constructed, integrating multi-sensor fusion and edge computing. This system integrated the seedling detection method with previously proposed phenotyping identification techniques for heading stage monitoring and flowering stage determination, enabling the efficient and automated collection and analysis of field phenotypic data. The results indicated that the proposed method had a high accuracy in wheat seedling emergence identification (R2=0.908, RMSE=11.73, rRMSE=23.04%). It also enabled dynamic monitoring of wheat heading and flowering stages, exhibiting excellent temporal feature capture capabilities. The system facilitated precise determination of wheat growth stages and accurate analysis of key phenotypic traits, including spike number, spikelet number, flower number, and seedling emergence. This method can be applied for high throughput collection and efficient analysis of field wheat phenotypic data, providing effective and reliable technical support for field phenotype acquisition in wheat breeding.