基于高光谱分析的SVM马铃薯植株营养元素亏缺识别

SVM nutrient element deficit identification of potato plants based on hyperspectral analysis

针对目前大多数农作物缺素研究多为传统的实验室化学方法、形态法和施肥法等,检测成本较高且会破坏到植物本身,不宜在实际农业生产中推广使用的问题。研究提出一种高光谱分析结合图像处理的方法,尝试使用新型手持式高光谱相机直接到户外马铃薯示范基地进行实地采拍,通过高光谱图像不同地物的类别光谱曲线差异搭建一个1D-CNN网络结构提取感兴趣区域并进行标签化处理,建立自然环境下的正常、缺氮、缺磷、缺钾四类马铃薯植株高光谱图像数据集。在SVM下使用PCA的特征提取与特征选择方法对缺素马铃薯植株分类识别进行对比实验,结论表明经PCA特征提取降维到20后,缺素植株总识别率由91.7%提高到93.1%,在保证准确识别率的情况下,降维处理可极大提高运行速度,与特征波段选择相比特征提取更适合本次的无损定性研究分析,为高光谱技术监测农作物生长状况问题提供一个新思路。

In view of the problems that most of the current researches on crop nutrient deficiency mainly use traditional laboratory chemical methods,morphological methods and fertilization methods,etc.,which have high detection cost and damage to the plants themselves,and are not suitable for promotion and use in actual agricultural production.This paper proposes a hyperspectral analysis and image processing method,attempts to use a new handheld hyperspectral camera directly to the outdoor potato demonstration base for field shooting,builds a 1D-CNN network structure based on the spectral curve differences of different hyperspectral images,extracts the region of interest and carries out label processing.The hyperspectral image data sets of normal,nitrogen,phosphorus and potassium deficiency potato plants in natural environment were established.A comparison experiment was carried out on the classification and recognition of vegetative deficient potato plants based on PCA feature extraction and feature selection method based on SVM.The conclusion showed that the total recognition rate of vegetative deficient plants increased from 91.7%to 93.1%after dimension reduction by PCA feature extraction to 20.Under the condition of ensuring accurate recognition rate,dimension reduction treatment could greatly improve the running speed.Compared with feature band selection,feature extraction is more suitable for this lossless qualitative research,It provides a new idea for monitoring crop growth by hyperspectral technology.