基于高光谱数据的土壤全氮含量估测模型对比研究

构建基于高光谱数据的土壤全氮含量估测模型,为快速、准确监测农田土壤全氮含量,判断作物生长发育情况和评价土地质量提供新的技术和方法。以新疆南疆地区主要类型土壤为研究对象,于室内测定土壤全氮含量和光谱反射率数据,利用偏最小二乘回归(PLSR)、支持向量机回归(SVM)、随机森林回归(RF)与光谱反射率(R)及其4种数学变换相结合,建立全区和分区全氮含量估测模型。结果表明,PLSR最优模型的预测集R2和RPD分别为0.73和1.82;SVM最优模型的预测集R2和RPD分别为0.75和1.97;RF最优模型的预测集R;和RPD分别为0.86和3.52,3种模型的预测能力依次为RF>SVM>PLSR。除一阶微分(FD)变换外,其它数据变换均对模型精度有不同程度的提高。R及其4种变换后数据均以RF建模精度较高,而以PLSR和SVM建模精度相对较低。全区模型稳定性要高于分区模型,分区模型差异性较明显,稳定性较差。总体来看,RF模型的预测能力稳定,适用性较好,精度较高,可较精确地估测土壤全氮含量;而PLSR和SVM模型只能对全氮含量进行粗略估测。因此,利用RF模型可实现研究区土壤全氮含量的快速准确估测。

Comparative study on estimation models of soil total nitrogen content based on hyperspectral data

The establishment of soil total nitrogen content estimation model based on hyperspectral data can provide a new technology and method for rapid and accurate monitoring of soil total nitrogen content in farmland,judging crop growth and development and evaluating land quality. Taking the main types of soil in southern Xinjiang as the research object,soil total nitrogen content and spectral reflectance data are determined by indoors,using partial least square regression(PLSR),support vector machine regression(SVM),random forest regression(RF),the spectral reflectance(R) and their mathematical transformation to establish an estimation model for predicting the total nitrogen content of the whole district and subregion.The results show that the prediction set R;and RPD of the PLSR optimal model are 0.73 and 1.82,respectively;the prediction set R2 and RPD of the SVM optimal model are 0.75 and 1.97,respectively;and the prediction set R;and RPD of the RF optimal model are 0.86 and 3.52,respectively;the prediction abilities of the three models are in the order of RF>SVM>PLSR.Except for the first-order differential(FD) transformation,other data transformations have different degrees to improve the model accuracy.The R and its four transformed data have high modeling accuracy with RF,while the modeling accuracy with PLSR and SVM is relatively low.The stability of the whole region model is higher than that of the partition model and the partition model has obvious difference and poor stability.On the whole,RF model has stable predictive ability,good applicability and high accuracy,and can accurately estimate the total nitrogen content of the soil.However,PLSR and SVM models can only make rough estimates of total nitrogen content.RF model can be used to estimate the total nitrogen content of soil in the study area quickly and accurately.