基于因子分析的苜蓿叶片叶绿素高光谱反演研究

因子分析是一种能够将具有错综复杂关系的变量归结为少数几个综合因子的多变量统计分析方法，在降低数据维数的同时又可以保存足够的信息，这为处理信息量丰富但冗余较大的高光谱数据提供了一种有效方法。本文利用2010年9月23日采集的16个样点的苜蓿叶片反射率及叶绿素含量数据，采用因子分析方法，分别提取苜蓿叶片反射率光谱400～900nm，以及可见光400nm～760nm和近红外760～900nm光谱区的公共因子，分析因子载荷分布、载荷总量对公共因子与叶绿素含量相关性的影响。利用逐步回归法建立基于公共因子的叶片叶绿素反演模型，并将反演模型与光谱指数建立的模型进行对比。研究表明，1）公共因子与叶片叶绿素的相关性，在更大程度上是与该因子在各个波段上载荷分布有关，而不是总载荷量；2）对波谱进行分区建立的反演模型略优于全区因子分析建立的反演模型；3）与常用于叶片叶绿素含量反演的光谱指数CARI、MCARI、mND680、mND705、mSR705、TVI、DmSR、BGI、BRI相比，因子分析建立的叶绿素反演模型精度更高。

A study on the hyperspectral inversion for estimating leaf chlorophyll content of clover based on factor analysis

Factor analysis is a statistical method used to describe variability among observed variables in terms of a potentially lower number of unobserved variables called factors. For the purpose of reducing the number of variables while retaining the most useful information, the factor analysis is an effective method to process hyperspectral data of rich useful information but much redundancy. The main objective of this study is to test if the factors of reflectance spectra of leaf can be used to inverse the chlorophyll concentration. The leaf reflectance spectra of 16 clover samples were collected using ASD (Analytical Spectral Devices) with the range of 325-1075nm and the spectral resolution of 3.5nm in September 23, 2010. The leaves were brought to laboratory to detect the chlorophyll concentration with 95% ethonal and ultraviolet spectrophotometer using the heat insulation barrel. To reduce the disturbance of systematic error, only the reflectance spectral range from 400nm to 900nm was analyzed in this paper. The reflectance data was standardized before the reflectance spectra of leaves was divided into two segments, visible light segment from 400nm to 760nm, and near infrared segment from 760nm to 900nm. And after doing that, the reflectance spectral range of 400-900nm, 400-760nm and 760-900nm were analyzed using factor analysis separately in SPSS 13.0. We generated 15 factors from 400-900nm, 7 factors from 400-760nm and 14 factors from 760-900nm. And according to the factor scores calculated from SPSS software, we generated the factor values of three different spectra ranges. The correlation coefficient between factors and chlorophyll concentration were calculated, and the spectral range of 400-760nm was taken as an example to analyze the impact of loading distribution and total loading capacity of factors on the correlation coefficient. Finally the inversion models for leaf chlorophyll concentration were established by using different factors with a stepwise regression method. These two models were compared with well established several spectral indexes. Only BRI, mND680, mND705, mSR705 have the determination coefficient R² above 0.5 among all the used spectral indexes. The R² of chlorophyll concentration inversion model established by overall factors, segment factors, BRI, mND680, mND705, mSR705 are 0.857, 0.869, 0.787, 0.728, 0.662, 0.597, and the relative errors are 15.3%, 14.3%, 23.7%, 21.5%, 24.9%, 29.7%. The result shows that: 1) To a great extent, the correlation coefficient between factors and leaf chlorophyll concentration was controlled by the loading distribution of factors rather than the loading capacity of factors. 2) Two chlorophyll concentration inversion models were established by factors, one using the first factor, the second factor and the eleventh factor of 400-900nm, and the other using the forth factor of 400-760nm and the first factor and the fourteenth factor of 760-900nm. The inversion model using factors from segmented reflectance spectral regions 400-760nm and 760-900nm, was slightly more accurate than the models estimated by factors from the full spectral range of 400-900nm. 3) Compared with spectral indexes including CARI, MCARI, mND680, mND705, mSR705, TVI, DmSR, BGI and BRI, the inversion models established by factors are more accurate due to a higher coefficient of determination R² and a lower relative error. This study demonstrated that factor analysis can be an effective method to process hyperspectral data and inverse chlorophyll concentration. However, this paper need improve the wider suitability of the models further, since only the clover samples were used, and in the future the research need simulate leaf reflectance spectra using physical models, such as PROSPECT and LIBERTY.