基于高光谱技术的草地牧草种类的识别及叶绿素含量预测的研究

发布时间：2018-03-05 10:52

本文选题：高光谱图像　切入点：数据降维　出处：《内蒙古农业大学》2017年硕士论文　论文类型：学位论文

【摘要】：利用高光谱仪采集草地牧草的高光谱图像可同时获取图像信息和光谱信息,目前已成为草地动态监测的一种新手段。本研究利用高光谱仪获取草地牧草在光谱范围为400nm～1000nm的可见-近红外光谱下的高光谱图片,首先用ENVI提取图片感兴趣区域(Region of Interest,ROI)的光谱数据;其次对提取的数据进行降维,再次对其进行分类,最后又做了牧草叶绿素含量的分析。论文的主要研究内容如下:(1)采用多元散射校正(Multiplicative Scatter Correction,MSC)、归一化(Normalize)、标准化处理(Standard Normal Variate,SNV)对提取的牧草平均光谱数据进行预处理,从而消除散射影响,增加与成份含量相关的光谱吸收信息。(2)采取连续投影(Successive Projections Algorithm,SPA)、局部线性嵌入(Locally Linear Embedding,LLE)、主成分分析(Principal Component Analysis,PCA)算法,选取30特征波段,消除数据冗余。(3)采取支持向量机(Support Vector Machine,SVM)、最近邻法(1-Nearst Neighbors,1NN)、BP神经网络算法对所选取的特征波段进行分类。实验采用了 10类500个样本的高光谱图像数据,其中35个样本作为训练集,15个样本作为预测集进行分类。(4)采用BP神经网络、偏最小二乘法、SPA模型实现牧草的叶绿素的含量预测。实验采用3类150个样本的高光谱图像数据,其中每类样本35个作为该样本的训练、集,15个作为该样本的预测集,实现对牧草的叶绿素含量的预测。实验结果表明:采取高光谱成像技术对草地牧草种类的识别是可行的,且在对全波段数据进行降维的时候,SPA用于提取光谱特征波段在30个特征波段的情况下对草地牧草分类可以达到100%,因此SPA算法降维效果较好,可以灵活简便的选择特征波段。对比分类综合时间和效果,SVM在判别草地牧草种类方面取得较好的效果;叶绿素含量预测中,运用偏最小二乘法回归模型可以取得较精确的预测结果。
[Abstract]:The hyperspectral image of grassland forage can be obtained by using the hyperspectral instrument, and the image information and the spectral information can be obtained at the same time. At present, it has become a new method for dynamic monitoring of grassland. In this study, the hyperspectral images of grassland forage in the spectral range of 400nm ~ 1000nm were obtained by using a high spectrometer. First, we use ENVI to extract the spectral data of region of interest (region of interest); secondly, we reduce the dimension of the extracted data and classify them again. Finally, the chlorophyll content of forage was analyzed. The main contents of this paper were as follows: 1) the average spectral data of forage were preprocessed by multivariate scattering correction, Normalizeg and Standard Normal Variate SNV. In order to eliminate the scattering effect and increase the spectral absorption information related to the content of components, we adopt continuous projection Projections algorithm, locally linearly embedded Linear embedded LLEA, Principal component Analysis Component Analysis (PCAA) algorithm, and select 30 characteristic bands. In order to eliminate data redundancy, support Vector Machine (SVM) and nearest neighbor method (1-Nearst neighbor) BP neural network algorithm are used to classify the selected feature bands. The experiment uses 10 kinds of 500 samples of hyperspectral image data. Among them, 35 samples were used as training set and 15 samples were classified as prediction sets. BP neural network was used to predict chlorophyll content of forage grass by partial least square method and SPA model. Among them, 35 samples of each type were used as training and set of samples, and 15 samples were used as prediction sets to predict the chlorophyll content of forage grass. The experimental results showed that it was feasible to recognize the species of pasture grass by using hyperspectral imaging technique. And when reducing the dimension of the whole band data, the SPA algorithm can be used to extract spectral characteristic bands in the case of 30 feature bands, and the classification of grassland forage can reach 100%, so the SPA algorithm has better dimensionality reduction effect. It can be used to select characteristic bands flexibly and easily. Compared with classification synthesis time and effect, SVM has achieved better results in judging grassland forage species. More accurate prediction results can be obtained by using partial least square regression model.
【学位授予单位】：内蒙古农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S812

【参考文献】