摘要
提升土壤属性空间预测精度对实现农田精准施肥和保护生态环境具有重要意义。利用河北省滦平县采集的1773个样点耕地表层(0~20 cm)土壤有机质(SOM)及其地理环境数据,通过逐步回归分析方法筛选出最优环境变量;基于其中1426个农田样点分别建立多元线性回归(Multiple Linear Regression,MLR)、随机森林(Random Forest,RF)、贝叶斯正则化神经网络(Bayesianregularization neural network,BRNNBP)以及与普通克里格(OK)整合模型(MLR-OK、RF-OK和BRNNBP-OK)预测SOM空间分布,以其余347个样点数据为测试集检验分析不同模型预测精度,并对模型残差进行半方差函数和空间自相关分析以评价模型拟合效果。结果表明,研究区耕地表层土壤SOM处在8.62~35.64 g·kg^(-1)变化区间,变异系数为20.26%,属中等程度空间变异;SOM高值区主要分布在东北及东南海拔较高地区,低值区多分布在西南及中部河谷地区;海拔、坡度和年均温度与SOM关系密切(P<0.001);整合模型BRNNBP-OK的平均绝对误差MAE和均方根误差RMSE最低分别为2.162g·kg^(–1)和2.801g·kg^(-1),相较于OK、MLR、RF、BRNNBP、MLR-OK和RF-OK预测模型,R2提升1.84%~43.72%,成为SOM空间预测优选模型。与单一模型相比,整合模型残差块金系数大于0.75,Moran’sI指数均小于0且数值更趋近于0,表明整合模型残差空间自相关性减弱且空间分布呈离散状态。同时,各模型精度与模型残差Moran’sI指数呈显著相关。因此,整合模型可以拟合更多的趋势项,模型残差空间聚集性降低甚至趋于离散时,模型总体精度提升,揭示了模型精度提升的内在原因。
【Objective】Improving the spatial prediction accuracy of soil attributes is of great significance for achieving accurate fertilization of farmland and protecting the ecological environment.【Method】Soil organic matter(SOM)data was collected from 1773 samples from soil surface layer(0–20 cm)of cultivated land in Luanping County,Hebei Province.The optimal environmental variables were screened through a stepwise regression analysis method.Multiple linear regression(MLR),ordinary kriging(OK),random forest(RF),Bayesian regularized neural network(BRNNBP),and the corresponding three integrated models combined with a geostatistical model(MLR-OK,RF-OK and BRNNBP-OK)were utilized to predict SOM content via the training set including 1426 sampling points.Also,the prediction accuracy of each method was compared with 347 sampling points of the testing set.Autocorrelation analysis was carried out based on the residual of the integrated model to evaluate the fitting effect of the model.【Result】Results showed that the range of SOM content in the study area was 8.62–35.64 g·kg^(–1),and the coefficient of variation was 20.26%,which showed a moderate spatial variation.High concentrations of SOM were mainly distributed in the northeast and southeast areas with higher altitudes,while relative low concentrations of SOM were mostly observed in the southwest and central valleys of the study area.Elevation,slope and temperature selected by stepwise regression were closely related to SOM content(P<0.001).The lowest average absolute error and the root mean square error of the BRNNBP-OK model were 2.162 g·kg^(–1) and 2.801 g·kg^(–1),respectively.Compared with the OK,MLR,RF,BRNNBP,MLR-OK and RF-OK models,the goodness of fit of the BRNNBP-OK model increased by 1.84%–43.72%,making it an optimal model for SOM spatial prediction.Compared with the single model,the nugget coefficient of the integrated model residual was greater than 0.75,and the Moran's I was less than 0 and numerically closer to 0,indicating that the spat
作者
宋洁
王思维
赵艳贺
于东升
陈洋
王鑫
冯凯月
马利霞
SONG Jie;WANG Siwei;ZHAO Yanhe;YU Dongsheng;CHEN Yang;WANG Xin;FENG Kaiyue;MA Lixia(State Key Laboratory of Soil and Sustainable Agriculture,Institute of Soil Science,Chinese Academy of Sciences,Nanjing 210008,China;Agricultural and Rural Bureau of Luanping County,Luanping,Hebei 068250,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《土壤学报》
CAS
CSCD
北大核心
2023年第6期1569-1581,共13页
Acta Pedologica Sinica
基金
国家重点研发计划专项(2018YFC1800104)
国家自然科学基金项目(42001302,41571206)资助。
关键词
土壤有机质
机器学习
普通克里格
残差
数字化土壤制图
Soil organic matter
Machine Learning
Ordinary Kriging
Residual
Digital soil mapping
