基于Boruta-支持向量回归的安徽省土壤pH值预测制图

Predictive Mapping of Soil pH in Anhui Province Based on Boruta-Support Vector Regression

以安徽省为研究区域,将Boruta算法用于特征筛选,选择最优变量组合输入支持向量回归(SVR)模型,经参数优化和核函数对比后,选择最优的SVR预测模型进行土壤pH值空间分布制图。结果表明:1)使用Boruta算法筛选后的特征变量建模优于全部变量建模;特征变量重要性分析表明,年均降水(MAP)是影响安徽省土壤pH值的最重要因素,多尺度山谷平坦指数(MrVBF)、多尺度山脊平坦指数(MrRTF)和年均温(MAT)等特征变量均对土壤pH值有较重要的影响。2)选择径向基函数(RBF)作为核函数建立SVR模型进行土壤pH值预测最为合理;参数C=1,γ=0.125时,SVR模型精度最高,可以解释土壤pH值变异的74%,验证集R^2为0.62。3)土壤pH值预测制图结果表明,安徽省土壤pH值空间分布呈由北至南逐渐降低的趋势,符合“南酸北碱”特征,且预测制图的统计结果与样本点的统计结果基本一致。将Boruta算法与SVR模型结合可以提高土壤pH值的预测制图精度,且模型的泛化能力较强。

The aim of this study was to explore the accuracy of the Boruta algorithm combined with the support vector regression (SVR) model for predicting soil pH.In this paper,Anhui Province in East China was selected as a case for the study,where 140 soil samples were collected.The terrain factors,vegetation index and climate data,were collected using GIS spatial analysis technique.The Boruta algorithm was used for feature selection.The selected optimal variables were entered into SVR model.After parameter optimization,the most robust SVR model was obtained.The optimal SVR prediction model was selected to map the spatial distribution of soil pH.Results showed that:1) The prediction accuracy using the optimal combination of variables selected by the Boruta algorithm was higher than that using all variables.The feature selection method was critical for the establishment of the model.The weighted distribution of variables indicated that the mean annual precipitation (MAP) was the most important predictor of soil pH.In addition,the multi-resolution valley bottom flatness index (MrVBF),multi-resolution ridge top flatness index (MrRTF),mean annual temperature were important factors as well.2) The radial basis function adopted as the kernel function to establish the SVR model for soil pH prediction was the most reasonable.The soil pH prediction model was the most robust when the parameter C was set to 1,and γ to 0.125.The multi-source environmental variable combination can explain 74% of soil pH variation.The determination coefficient ( R^2 ) of the validation set was 0.62.3) The predicted soil pH mapping showed that the spatial distribution of soil pH in Anhui Province gradually decreased from north to south,which was consistent with previous findings.The statistical results of predicted soil pH mapping were basically consistent with the statistical results of sample points.Combining the Boruta algorithm with the SVR model could improve the prediction accuracy of soil pH.This method can be generalized to other studies as well.