Y-梯度广义最小二乘加权校正的土壤速效氮野外原位光谱预测

摘要

土壤速效氮是影响作物生长发育的重要养分指标.野外原位可见近红外光谱(V IS-N IR)分析技术具有快速无损等特点,对速效氮的定量预测具有较好的应用前景.野外条件下进行原位光谱采集更节省人力物力,且为土壤养分实时传感器的开发提供了数据基础.但由于野外原位光谱中通常存在大量的无关环境因子干扰信息,易导致回归模型预测精度降低,达不到实用要求.针对位于以色列中部和北部的两个试验点共76个样本开展研究,提出利用Y-梯度广义最小二乘加权算法(Y-GLSW)对样本的野外原位VIS-NIR反射率光谱(350~2500 nm)进行滤波校正,以提高回归模型的预测能力.首先使用SG平滑、一阶导数变换、标准正态变换等常规方法对原始光谱进行预处理和变换;在此基础上再使用Y-GLSW构建滤波模型对变换后的光谱进行滤波校正;最后使用偏最小二乘回归算法(PLS-R)分别结合原始光谱RW、预处理变换后的光谱PPT和滤波校正后的光谱Y-GLSW建立回归分析模型对速效氮进行定量预测.结果表明:利用RW光谱建立的回归预测模型是不可靠的;利用PPT光谱建立的回归模型在测试集的相对分析误差(RPD)为1.41,解释总方差占实际总方差之比(SSR/SST)为0.57,模型具有一定的可靠性;Y-GLSW光谱建立的回归模型在测试集的RPD和SSR/SST分别为2.07和0.69,相对于PPT模型分别提高了46.81%和21.05%.因此,利用Y-GLSW对野外原位VIS-NIR光谱进行滤波校正,能够有效去除光谱中的无关信息数据,提高模型的预测精度和解释能力.%Soil available nitrogen is supposed to be an important nutrient constituent for the growth and development of crops . Insitu field visible-near infrared (VIS-NIR ,350~2500 nm) spectroscopic analysis is a rapid and non-destructive method that has the potential to predict nitrogen .Further ,it is cost-effective method compared with traditional laboratory analysis and can be used to provide a database for the development of real-time soil nutrient sensors .However ,prediction accuracy was greatly reduced due to unexpected environmental factors under field condition .In the current research ,field work contained 76 samples from two sites located in the center and north parts of Israel .Y-gradient general least squares weighting (Y-GLSW) algorithm was investigated to filtering correct the field VIS-NIR spectra for improving the prediction ability of nitrogen .Firstly ,Savitzky-Golay (SG) smoothing algorithm ,first derivative transformation and standard normal variate were sequentially conducted to preprocess and transform the raw field spectra (RS) .Then ,a filtering model was established based on the Y-GLSW algorithm to correct the preprocessed and transformed spectra (PPT ) .After that ,partial least square- regression (PLS-R) algorithm was applied to build regression models with RS ,PPT ,and Y-GLSW corrected spectra ,respectively .As a result ,the regression model based on RS was proved to be unfeasible .The ratio of performance to deviation (RPD) and the ratio between interpretable sum squared deviation and real sum squared deviation (SSR/SST) of the test set of the PPT-based regression model were found to be 1.41 and 0.57 ,respectively .The results of Y-GLSW-based regression model were RPD = 2.07 and SSR/SST=0.69 that significantly increased by 46.81% and 21.05% compared with PPT-based regression model .The results indicated that Y-GLSW was suitable to remove some unexpected variations (like the effect of environmental factors) of field spectra and improved the prediction accuracy and explanation ability of PLS-R model for predicting nitrogen .