张先洁 孙国祥 汪小旵 杨海慧 魏天翔

摘要： 針对无人机精确植保过程中，果树冠层区域颜色特征和杂草相似度较高、难以分割等问题，采用基于超像素特征向量的果树冠层分割方法，以消除不同杂草特征对树冠分离的干扰，减小农药喷雾区域，节省农药使用量。通过分析无人机采集合成的样本图像在HSV彩色空间上色调与饱和度的分布情况，选取合适的阈值范围，提取样本图像中包含果树冠层与杂草的绿色区域，将提取的绿色区域RGB图像转换生成Lab和HSV彩色空间模型下的图像，然后运用简单的线性迭代聚类（Simple linear iterative clustering，SLIC）超像素分割算法将RGB图像预设分割成250个超像素单元，结合超像素的分割信息与RGB图像、Lab图像、HSV图像以及灰度图，提取超像素单元的特征向量，随机选取25%的超像素样本的特征向量作为SVM分类器的训练集，利用SVM分类器对所有样本进行预测分类，实现果树冠层与杂草分割。将基于超像素特征向量的方法和基于光谱阈值、K-means聚类的2种方法进行对比分析，结果显示，基于超像素特征向量的方法在识别果树冠层位置方面生产者精度为90.83%，在提取果树冠层轮廓上F测度值为87.62%，总体分割性能优于后两种方法。说明，基于超像素特征向量的方法能够较为准确地分割果树冠层与杂草，为实现无人机在果园中精确植保提供重要支撑。

关键词： 无人机;植保;果树冠层;杂草;超像素;分割

中图分类号： TP242.6+2;S252+.3 文献标识码： A 文章编号： 1000-4440（2021）03-0724-07

Segmentation method of fruit tree canopy based on super pixel feature vector

ZHANG Xian-jie1， SUN Guo-xiang1， WANG Xiao-chan1，2， YANG Hai-hui1， WEI Tian-xiang1

（1.College of Engineering， Nanjing Agricultural University， Nanjing 210031， China;2.Laboratory of Modern Facility Agricultural Technology and Equipment Engineering in Jiangsu Province， Nanjing 210031， China）

Abstract： Aiming at the problem such as high similarity of color features between canopy area of fruit trees and weeds which are difficult to be segmented in the process of unmanned aerial vehicle （UAV） in precise plant protection， a canopy segmentation method based on super pixel feature vector was adopted to eliminate interference of different weed characteristics to crown separation， reduce the pesticide spray area and save pesticide usage. By analyzing the distribution of hue and saturation of sample images collected and synthesized by UAV in HSV color space， the appropriate threshold range was selected， the green area of the sample image including the canopy of fruit trees and weeds was extracted. The RGB image of the green area extracted was transformed to generate the image under Lab and HSV color space model， and simple linear iterative clustering （SLIC） super pixel segmentation algorithm was then used to divide the RGB image into 250 super pixel units by default. The feature vectors of super pixel unit was extracted by combining the segmentation information of super pixel with RGB image， Lab image， HSV image and gray scale image. Feature vectors of 25% super pixel samples were selected randomly as the training set of SVM classifier， SVM classifier was used in predicting and classifying all samples to realize the segmentation of fruit tree canopy and weeds. The method based on hyper pixel feature vector and methods based on spectral threshold and K-means clustering were compared. The results showed that， the producer accuracy of the method based on super pixel feature vector in recognizing position of fruit tree canopy was 90.83%， the F value for extracting contour of fruit tree canopy was 87.62%. The overall segmentation performance by the method based on hyper pixel feature vector was better than the latter two methods. It can be seen that， the method based on super pixel feature vector can segment fruit tree canopy and weeds accurately， which provides an important support for the realization of UAV in precise plant protection in the orchard.

Key words： unmanned aerial vehicle （UAV）;plant protection;canopy of fruit trees;weed;super pixel;segmentation

无人机喷药相对于传统的地面移动式喷药可以忽视复杂的地形阻碍，同时无人机旋翼可以吹动树冠，使农药能更均匀垂直地覆盖果树冠层。……