基于无人机多光谱影像的柑橘树冠分割方法研究

Research on citrus canopy segmentation method based on UAV multispectral image

树冠是树木的重要组成部分,利用无人机分辨率高、成本低、操作简单等优势,研究无人机影像上分割单木树冠的方法,实现操作流程一体化。以四川省成都市蒲江县柑橘种植基地为研究区域,飞行无人机获取多光谱影像,提出一种利用卷积神经网络对柑橘树冠分割的有效方法,该方法基于VGG16和U-Net,前者检测图片中是否包含柑橘树,选取柑橘最佳的光谱特征,后者实现对柑橘树冠的分割。首先,对原始数据进行灰度化、畸变差校正、正射拼接等一系列预处理,将正射影像裁剪至适中的尺寸后输入到VGG16卷积神经网络中对图片进行分类,过滤出主体为房屋、道路、湖泊的图片;在VGG16模型分类的基础上,选定分类准确率最高的波段,用该波段影像对柑橘树冠进行分割,使用U-Net卷积神经网络模型对分类为柑橘地的图片进行树冠分割与株树统计,提取样地内树冠区域和单木位置。在波段选择中,近红外波段对柑橘地和非柑橘地的分类准确率最高,为95.34%,优于RGB彩色图片和其他波段的数据,而其他波段的准确率分别为:RGB彩色图片88.37%,红色波段62.79%,绿色波段69.76%,蓝色波段65.11%,红边波段74.42%;在树冠区域提取中,使用U-Net卷积神经网络对近红外波段影像进行树冠分割,在测试集上的总体精度可以达到93.63%,错判误差为7.99%,漏判误差为4.52%,准确率为91.99%,Kappa系数为0.88。研究结果表明,结合深度学习可以快速准确地获取单木树冠范围,在保证调查精度的基础上提高调查效率。

Tree crown is an important part of trees.Based on the advantages of unmanned aerial vehicle(UAV),such as high resolution,low cost and simple operation,the method of obtaining single tree crown range from UAV image is studied to realize the integration of operation process.In this study,an effective method for citrus crown segmentation based on UAV multispectral image was proposed based on VGG16 and U-Net.The former detected whether there were citrus trees in the image,selected the best spectral features of citrus,and the latter realized citrus crown segmentation.Taking the citrus planting base in Pujiang County,Chengdu City,Sichuan Province as the research area,an effective method of citrus crown segmentation based on UAV multispectral image was proposed.Firstly,the original data was preprocessed by graying,distortion correction,ortho mosaicing,etc.The ortho image was cut to a moderate size and input into VGG16 convolutional neural network to classify the images,and the images of houses,roads and lakes were filtered out.On the basis of VGG16 model classification,the band with the highest classification accuracy was selected,and the orange image was classified with this band image crown segmentation using the U-Net convolution neural network model to segment the crown and tree statistics of the images classified as citrus land,extracted the crown area and individual tree position in the sample plot.In the band selection,near infrared(NIR)band had the highest classification accuracy of 95.34%for citrus and non citrus land,which was better than visible and other bands,while the accuracies of other five bands were 88.37%for RGB color picture,62.79%for the red band,69.76%for the green band,65.11%for the blue band,and 74.42%for the red edge band.In the crown region extraction,the U-Net convolution neural network was used to segment the NIR band image,and the overall accuracy on the test set could reach 93.63%,the commission error was 7.99%,the omissions error was 4.52%,the accuracy rate was 91.99%,and the Kappa coefficient was 0.88.The research results showed that the combination of deep learning could quickly and accurately obtain the crown range of single tree,and improved the investigation efficiency on the basis of ensuring the investigation accuracy.