基于高光谱图像的石斛兰氮素检测

摘要：石斛兰作为一种重要的中药材和极高的观赏价值植物，越来越受到人们的关注。为了满足对石斛兰生长状况的跟 踪检测需求，提出利用可见-近红外高光谱成像技术对石斛兰不同生长阶段的氮素营养进行活体检测。首先，采集 400~1000nm光谱范围内的石斛兰高光谱图像。其次，提取感兴趣区域（（ DOI）。采用2G-R-B算法分割背景和植株，再用 RGB阈值法分离叶鞘和茎。通过两次掩模或运算去除噪声，最后选取面积最大的区域提取DOI区域。然后提取DOI区 域的反射光谱，研究两种特征提取方法和两种优化波段选择方法对高光谱图像进行降维，最后建立支持向量机（SVM） 模型对金钗石斛氮素水平进行分类。结果表明，LDA 结合 SVM 算法的分类精度最高，3 个生长阶段训练集分类精度分 别为97.47%、95.03%、95.97%，测试集分类精度达到97.00%、88.8%、92.67%。可见光-近红外高光谱成像技术结合LDA- SVM分类模型，可以有效区分梯度氮栽培的金钗石斛在各个生长阶段的营养状况，是一种很有潜力的精准营养决策技术。

关键词：设施园艺，高光谱成像技术，金钗石斛，氮素检测，SVM

期刊：Int J Precis Agric Aviat

DOI：10.33440/j.ijpaa.20200302.87

设备：GaiaField

Nitrogen detection of Dendrobium nobile based on hyperspectral images

Abstract：Dendrobium nobile Lindl. (D. nobile), as an important traditional Chinese medicine and highly ornamental value plant, has attracted more and more people’s attention. In order to meet the needs of tracking and testing the growth status of D. nobile, the visible-near infrared hyperspectral imaging technology was proposed for nitrogen nutrients detection in vivo in its different growth stages. Firstly, collecting the hyperspectral images of D. nobile in spectral range 400-1000 nm. Secondly, extracting the region of interesting (DOI). The 2G-R-B algorithm was used to segment the background and plants, and then the RGB threshold method was used to separate the leaf sheath and stems. Removing noise by two masks’ or-operations, and then the DOI area was finally extracted by selecting the largest area. After that, the reflectance spectrum of the DOI area was extracted, and then two kinds of feature extraction methods and two kinds of optimizing band selection methods were researched for dimension reduction of hyperspectral images. Finally, Support vector machine (SVM) model was established to classify the nitrogen level of D. nobile. The results showed that the LDA combined with the SVM algorithm had the highest classification accuracy. The classification accuracy of training sets in the three growth stages were 97.47%, 95.03%, and 95.97%, respectively, and the classification accuracy of test set reached 97.00%, 88.8%, 92.67%. The visible-near infrared hyperspectral imaging technology combining LDA-SVM classification model could effectively distinguish D. nobile cultivated by gradient nitrogen in each growth stage. It is a potential technology applied in decision-making of precise nutrition supply.

Keywords：facility gardening, hyperspectral imaging technology, Dendrobium nobile Lindl., nitrogen detection, SVM