基于无人机RGB影像估测田间小麦穗数

单位面积穗数是小麦产量构成的重要因素，利用图像信息处理技术快速、准确地估测田间小麦穗数，可以为小麦长势监测和产量估测提供直接依据。利用无人机路径规划和控制系统（fragmentation monitoring and analysis with aerial photography， FragMAP）获取标准统一、高分辨率的田间小麦RGB航拍影像，通过高效的目标检测手段（YOLOv3）获得训练模型并自动识别麦穗，通过分析该方法（FY方法）与传统方法测定麦穗数量的关系来构建单位面积麦穗估测模型。结果表明，FY方法的样本获取效率和观测面积显著高于传统方法（P < 0.001）；YOLOv3训练模型识别麦穗的准确率随着训练样本数量和迭代次数增加而增加，500个训练样本迭代6 250次，获得模型识别麦穗的准确率超过90%；FY方法和与传统方法测定的田间小麦穗数量呈显著的线性相关关系，据此构建估测田间小麦穗数的模型为：y=0.816x-14.863（R² = 0.790，P<0.001）。上述结果表明，结合标准统一、高分辨率的无人机航拍影像和深度学习方法估测田间小麦穗数精度高、实时性强，可为小麦长势监测和产量估测提供重要的数据和技术支撑。

Estimating of Wheat Ears Number in Field Based on RGB Images Using Unmanned Aerial Vehicle

Ears number per unit area is the key element of wheat yield， so estimating ears number in field quickly and accurately based on aerial photographs and information extraction could provide the direct support to wheat growth monitoring and production estimates. This study acquired the unified standard and high resolution RGB aerial photographs using unmanned aerial vehicle （UAV） controlled by fragmentation monitoring and analysis with aerial photography （FragMAP）， then obtained the training model by quick and efficient target detection method （YOLOv3） and realized identifying wheat ears automatically， and finally established the estimation model of wheat ears number based on the relationships of the wheat ears number measured by proposed method （FY） and traditional methods. The results showed that the sampling efficiency and monitor area of FY were significantly higher than that of traditional method； the mean accuracy rate of identification based on YOLOv3 model was above 90%； the wheat ears numbers measured by FY and traditional method were significant linear correlation， and the estimate model was established as y=0.816x-14.863 （R²=0.790， P<0.001）. The results showed that the wheat ears estimation based on the unified standard and higher resolution images collected by UAV and deep learning was accurate， and could effectively monitor wheat growth and predicting the wheat yield.