基于光谱特征变量的湿地典型植物生态类型识别方法——以北京野鸭湖湿地为例

光谱特征变量的选择对于湿地植被识别的精度和效率有着直接的影响作用.以华北地区典型的淡水湿地——野鸭湖湿地为研究区,采用Field Spec 3野外高光谱辐射仪,获取了野鸭湖典型湿地植物的冠层光谱.以野外高光谱数据为基础,首先利用一阶导数与包络线去除的方法,分析和对比不同植物生态类型的光谱特征,选定了用于识别植物生态类型的光谱特征变量,选定的8个光谱特征变量为红边位置WP_r、红边幅值Dr、绿峰位置WP_g、绿峰幅值Rg、510 nm附近的吸收深度DEP-510和吸收面积AREA-510、675 nm附近的吸收深度DEP-675和吸收面积AREA-675.其中,7种植物生态类型的一阶导数光谱特征差异较小,吸收特征差异性相对较大.除WP_r和WP _g外,沉水植物Rg和Dr平均值最低,湿生植物的Rg平均值最高,达到0.164,栽培植物的Dr平均值最高,达到0.012.7种植物生态类型在675 nm附近的DEP-675和AREA-675均高于510 nm附近的DEP-510与AREA-510,除去栽培植物,随着水分梯度的变化,其他6种植物生态类型的吸收深度和吸收面积都表现出先升高后降低的趋势.然后利用单因素方差分析(One-way ANOVA)验证了所选光谱特征变量的区分度,在P≤0.01的置信水平下,选取的8个光谱特征变量都能够较好的区分7种植物生态类型,区分度的最小值为13,最大值为18,并且吸收特征参数的区分度优于一阶导数参数.最后应用非线性的反向传播人工神经网络(BP-ANN)与线性判别分析(FLDA)的类型识别方法,利用选定的8个光谱特征变量进行湿地植物生态类型识别,取得了较好的识别精度,两种方法的总分类精度分别达到85.5％和87.98％.单因素方差分析(One-way ANOVA)和不同分类器的分类精度表明,所选的8个光谱特征变量具有一定的普适性和可靠性.

Identifying typical plant ecological types based on spectral characteristic variables: a case study in Wild Duck Lake wetland, Beijing

Certain spectral characteristic variables have a direct impact to the accuracy and efficiency of identifying the wetland vegetation. The Wild Duck Lake, a typical freshwater wetland of North China, was selected as the study area of this research. The canopy hyperspectral reflectances of typical wetland vegetation were measured by Field-Spec 3 high-spectrum radiometer. In this research, a total of seven vegetation covers were measured. The reflectance spectrum of submerged vegetation species was special because of the influence of water body, suspended solid in water, etc. The reflectance spectra of the other six vegetation covers were similar. But different vegetation cover had different feature, such as plant morphology, water content and chlorophyll content so that they had their own spectral reflectance characteristics. Thus, based on the canopy spectral reflectance, first-derivative and continuum removal were applied to analyze and contrast spectral features of different vegetation types. The spectral characteristic variables were selected for identifying plant ecological. The research established eight distinct spectral characteristic variables (red edge position WP_r, red edge amplitude Dr, green peak position WP_g, green peak amplitude Rg, absorption depth around 510 nm and 675 nm, absorption area around 510 nm and 675 nm) that have importance for mapping vegetation cover types in the Wild Duck Lake wetland. In addition, the absorption features of seven vegetation types were larger differences than first-derivative spectral features. Excepting WP_r and WP_g, the average of Rg and Dr of submerged plant were minimum, the average of Rg of hygrophilous plant was maximum (0.164) and the average of Dr of cultivated plant was maximum (0.012). The DEP-675 and AREA-675 around 675 nm of seven vegetation types were higher than the DEP-510 and AREA-510 around 510 nm. Absorption depth and absorption area of other six vegetation types had shown fall after rise along with the water environment gradient change except cultivated plant. Then we made use of single factor analysis of variance (One-way ANOVA) to verify the discrimination of the spectral characteristic variables. When the confidence level was less than or equal to 0.01, the selected spectral characteristic variables could distinguish seven plant ecological types better with a minimum discrimination of 13 and a maximum of 18. Moreover, the discrimination of absorption characteristic parameters was better than first-derivative parameters. Finally the nonlinear back propagation artificial neural network (BP-ANN) and fisher linear discriminant analysis (FLDA) were applied to identify the wetland vegetation types and the selected spectral characteristic variables were also included in. According to accuracy test, the results of overall accuracy of two methods were 85.5% and 87.98%, respectively. The single factor analysis of variance (One-way ANOVA) and the classification accuracy of different classifiers indicated that the selected eight spectral characteristic variables had great applicability and reliability. The results of this paper would not only provide a scientific base for hyperspectral remote sensing image processing and wetland vegetation mapping in the Wild Duck Lake, but also supply reference for identifying and classification of freshwater wetland vegetation applying remote sensing technology. In future research, we should increase sample numbers and refine the research objectives. Applying the results of this research to hyperspectral image interpretations, we can fully explore the potential and advantage of hyperspectral remote sensing technology.