基于REML的普通克里格和回归克里格在土壤属性空间预测中的比较

Comparing Prediction Accuracies of Ordinary Kriging and Regression Kriging with REML in Soil Properties Mapping

比较普通克里格与回归克里格预测精度的研究还存在着不同观点和结论,且现有的对比研究多使用矩量法(Mo M)计算变异函数。有限最大似然(REML)方法相比Mo M方法具有明显的优点,因而有必要使用REML方法来比较普通克里格与回归克里格的预测精度,为土壤属性空间预测提供方法优选的参考依据。本文在广西南宁高峰林场采集土壤样品,测定有机碳、pH和粘粒含量,再基于REML建立普通克里格和回归克里格模型,同时比较普通克里格、回归克里格以及多元线性回归的预测精度,并分析影响预测精度的因素。结果表明:空间相关性较弱且线性回归模型的决定系数较大时(约大于20%),回归克里格优于普通克里格;相反地,空间相关性较强或较弱且线性回归模型的决定系数较小时(约小于10%),普通克里格预测精度优于回归克里格。同时,线性回归模型的决定系数还影响普通克里格与回归克里格的精度提高的幅度。此外,回归克里格的精度一般不低于多元线性回归,且线性回归模型的决定系数越小,则回归克里格越优于多元线性回归。因此,本研究认为,线性回归模型的决定系数和土壤属性的空间相关性是影响普通克里格与回归克里格精度差异的主要因素。

There are different results and opinions on accuracies of ordinary kriging（OK） and regression kriging（RK）in spatial prediction of soil propertiesin literatures. Studies that have discussed about this issue generally used the method of moments（Mo M） to implement OK and RK, although the method of residual maximum likelihood（REML） is much advantageous over Mo M. Thus, it seems necessary to compare prediction accuracies of OK and RK with REML,and helpful to investigate factors that affect prediction accuracies of the two techniques.This study collected data on soil organic carbon（SOC）, pH and clay content in an area of Gaofeng Forestin Nanning, China, and built OK and RK model with REML, and then used OK,RK as well as multivariate linear regression（MLR） to predict soil properties accuracies, and finally analyzed the factors that affected prediction accuracies. The results showed that RKpredicted soil information more accurately than OK, in case of that spatial autocorrelation of a soilproperty wasweak and meanwhile linear relationship between the soil property and auxiliary attributes in terms of coefficient of determination of linearregression modelwas strong（more than 20%）. Otherwise, OK predicted soil information more accurately than RK when spatial autocorrelation of a soil property was weak or strong and meanwhile the coefficient of determination of linear regression model was weak（less than 10%）. In addition, the coefficient of determination of linearregression model also affected prediction accuracy of RK and OK. Generally, RK predicted soil information more accurately than MLR. The smaller the coefficient of determination of linear regression model was, the better RK did than MLR. This study concluded that the spatial autocorrelation of soil property and the determination coefficient of linear regression modelwere the main factors that affectedprediction accuracy of RK and OK.