基于最佳波段判别的湿地植物叶片全氮反演研究

Estimating Wetland plant leaf total nitrogen content based on optimal bands of reflectance from wetland vegetation

利用高光谱遥感技术定量估测湿地植被叶片全氮含量,对于监测和诊断湿地植被的生理状况及生长趋势具有重要意义。但叶片氮素遥感诊断研究多存在反演模型过拟合、入选波段与生化参量间因果关系不明确和入选变量间"多重共线性"等局限。以芦苇(Phragmites australis)和香蒲(Typha angustifolia)叶片全氮含量作为研究对象,通过谱带分区,分区最佳波段选取和偏最小二乘回归相结合的方法构建芦苇和香蒲叶片全氮含量反演模型,并利用交叉验证决定系数(R2cv)和均方根误差(RMSEcv)对模型精度进行检验,尝试克服传统反演方法中的不足。结果表明,不同湿地植物类型相比,利用芦苇反射光谱建立的预测模型精度都高于香蒲。不同回归模型相比,一阶导数光谱-偏最小二乘回归模型(FDS-PLSR)精度远高于原始光谱-偏最小二乘回归模型(OS-PLSR)。芦苇最佳模型交叉验证决定系数(R2cv)达到了0.84,方根误差(RMSEcv)为0.10,香蒲最佳模型交叉验证的决定系数(R2cv)达到了0.66,方根误差(RMSEcv)为0.13,是构建湿地植物芦苇和香蒲光谱与叶片全氮含量关系的最佳模型。在不降低湿地植物叶片氮含量反演精度的基础上,有效地避免了传统地物高光谱模型反演中的局限性,是无损害遥感探测方面的有益尝试。

Utilization of hyperspectral remote sensing technology to estimate wetland plant leaf nitrogen content quantitatively in large area is important to monitor and diagnosis of physiological condition and growth trend of wetland vegetation. However, there are many limitations such as over-fitting of inversion model, indeterminate causal relationship between the selected bands and bioc remote sensing research. The total n angustifolia, was selected as our stud hemical parameters,＂ muhicollinearity＂ of selected bands in leaf nitrogen diagnosis itrogen content of leaves of typical wetland plants, Phragmites australis and Typha y objects. These plants grow in South Wetland purification system in the Olympic Park in Beijing, a typical wetland using reclaimed water. The leaf reflectance spectra of main wetland plants were acquired by means of an ASD FieldSpec 3 spectrometer （350--2500nm）. Leaf total nitrogen content was determined by Kjeldahl nitrogen measurement method after acquiring the leaf reflectance spectra. The method,＂ Give a cross correlation analysis＂, was used to build the correlations between leaf nitrogen content and the original spectrum, first derivative spectrum. Then, the selected bands were divided into some areas according to QiSeGuang spectral range and the interval of selected bands. Those bands which are high frequency ,frequency more ,ranked high frequency are chosen as representative bands of different areas and considered to be the optimal bands used to build the regression model. Finally, partial least squares was used to build inversion model. The accuracy of this model was tested with cross-validated coefficient of determination （R^v） and cross, I validated root mean square error （RMSEo,）.The results show that the first derivative transformation can effectively improve the sensitivity of the original spectrum leaf nitrogen content inversion, and fully reflect the sensitivity of near infrared wave band representing leaf total nitrogen content. The accuracy of regression model based on the first derivative and the partial least-squares was much higher than that of the original spectra. In the regression model of the reed, verification accuracy （R^v） reached 0.84, square root error （RMSEov） was 0.11, in the regression model of the cattail, verification accuracy （R^v） reached 0.66, square root error （RMSEov） was 0.13, which were the optimal models to estimate leaf total nitrogen content. The determination of parameters in ＂Give a cross correlation analysis＂ provided a scientific basis for building the model of eliminating＂ singularband＂ and reducing ＂multicollinearity＂ problem. Spectrum zone division provides a scientific basis for revealing the causal relationships between optimal band and biochemical parameters. And partial least squares regression method was used to avoid ＂multicollinearity＂ of selected bands. The result from this study can not only fill the gaps in the detection of leaf nitrogen using remote sensing, but also provide a strong scientific basis for the nitrogen content monitoring and management of urban wetlands using reclaimed water. At last, Partial least squares regression method was used to avoid ＂muhicollinearity＂ of selected bands.