It is important to obtain real‐time leaf density of fruit‐tree canopies for the precision spray control of plant‐protection robots. However, conventional detection techniques for the characteristics of fruit‐tree canopies cannot acquire the canopy internal information, which may provide an unsatisfactory accuracy of detection of leaf densities. This paper proposes a method for estimating canopy leaf density of fruit trees based on wind‐excited audio. A wind‐exciting implement was used to force fruit‐tree canopy leaves vibrating to produce audio. Then, some correlation analysis methods were used to extract key characteristic parameters of wind‐excited audio that were significantly correlated with leaf density. Finally, based on the data set of wind‐excited audio, a few machine‐learning methods were used to develop leaf‐density estimation models. Test results showed that: (1) there were five key feature parameters of wind‐excited audio that were significantly correlated with leaf density: the short‐time energy, spectral centroid, the frequency average energy, the peak frequency, and the standard deviation of frequency. (2) the estimation model of leaf density developed based on backpropagation neural network for fruit‐tree canopy showed the optimal estimation results, which can achieve the estimation of leaf density of fruit‐tree canopies accurately. The overall correlation coefficient (R) of the estimation model was more than 0.84, the root‐mean‐square error was less than 0.73 m2 m−3, and the mean absolute error was less than 0.53 m2 m−3. This study is expected to provide a technical solution for the leaf‐density detection of fruit‐tree canopies of plant‐protection robots.

中文翻译：

---

基于风激励音频的果树冠层叶密度估计

实时获取果树冠层叶片密度对于植保机器人的精准喷洒控制具有重要意义。然而，传统的果树冠层特征检测技术无法获取冠层内部信息，导致叶片密度检测的准确性不理想。本文提出了一种基于风激励音频的果树冠层叶片密度估计方法。使用一种激风装置来迫使果树树冠叶子振动以产生声音。然后，采用相关性分析方法提取与叶片密度显着相关的风激音频关键特征参数。最后，基于风激发音频数据集，使用一些机器学习方法来开发叶子密度估计模型。测试结果表明：（1）风激音频有5个关键特征参数与叶片密度显着相关：短时能量、谱质心、频率平均能量、峰值频率和标准差。频率。 (2)基于反向传播神经网络开发的果树冠层叶片密度估计模型显示出最优估计结果，可以实现果树冠层叶片密度的准确估计。总体相关系数（右估计模型的误差大于0.84，均方根误差小于0.73 m2米−3，平均绝对误差小于0.53 m2米−3。该研究有望为植保机器人果树冠层叶片密度检测提供技术解决方案。