样点数量与空间分布对县域尺度土壤属性空间预测效果的影响

Influences of Sample Size and Spatial Distribution on Accuracy of Predictive Soil Mapping on A County Scale

明确样点数量和空间分布对土壤属性空间预测的影响,有助于科学制定土壤采样策略、有效提高土壤空间预测精度。从5403个土壤样点中随机抽取验证数据集以及包含不同样点数量的训练数据子集(每个子集包括五种样点空间分布实例),在研究区表层土壤有机质含量普通克里格(Ordinary Kriging,OK)和反距离加权(Inverse Distance Weighted,IDW)插值结果的基础上,分析和探讨样点数量与空间分布对土壤空间预测效果的影响。结果显示,当样点数量从5000降至39个时,OK和IDW插值图的局部变异信息逐渐减少,基于20和10个样点的插值图存在失真畸变。当样点数量从5000降至1250个时,OK插值精度相近(r变幅为0.55~0.59、RMSE变幅为3.03~3.15);从样点数量减少至625个开始,OK插值精度明显下降,同一训练子集不同样点空间分布的插值精度分异明显。IDW插值精度随样点数量与空间分布的变化与OK插值相似,不同的是从1875个样点开始出现插值精度的明显下降和不同空间分布插值精度的明显分异。在插值图发生失真畸变之前,OK平均插值精度大于IDW。研究结果表明,样点数量及空间分布均可在不同程度上影响土壤属性空间预测结果,当样点数量足够多时,样点数量和空间分布对预测结果的影响非常有限;当样点数量减少至一定程度时,随着样点数量的减少,空间预测图的局部变异信息逐渐减少,预测精度逐渐下降,同时样点空间分布对预测结果的影响开始凸显;在空间预测结果发生失真畸变之前,与OK相比,IDW插值精度较低且更早响应样点数量和空间分布的变化。

【Objective】This study was conducted to investigate influences of sample size and spatial distribution on prediction of soil mapping,which is contributive to formulating soil sampling strategies scientifically and improving soil prediction accuracy effectively.【Method】Out of 5403 soil samples,a validation dataset and training sub-datasets different in number of soil samples were derived randomly,and each subset encompassed five examples different in sampling site spatial distribution pattern.Influences of sample size and spatial distribution on predictive soil mapping,embodied by spatial distribution characteristic and prediction accuracy,were explored on the basis of the prediction of organic matter content(OMC)in topsoil layer with the Ordinary Kriging(OK)or Inverse Distance Weighted(IDW)interpolation method.【Result】Results show that when the number of soil samples decreased from 5000 to 39 the OMC predictive maps based on OK or IDW interpolation was gradually losing details of local variation,and when the number dropped down to 20 or 10,the predictive maps became distorted.When the number varied in the range of 5000~1250,the predictive maps based on OK interpolation were quite similar in accuracy with r varying in the range of 0.55~0.59,and RMSE in the range of 3.03~3.15,but when the number dropped down to 625,the predictive maps based on OK interpolation declined significantly in accuracy,and varied sharply between the five groups different in spatial distribution pattern of sampling sites even in the same training subdset.The predictive map based on IDW interpolation varied in accuracy with the number and the distribution pattern of soil sampling sites on a trend similar to that based on OK interpolation,except that the predictive map based on IDW interpolation declined significantly in accuracy with the number of samples starting to drop from 1875 and varied sharply with spatial distribution pattern of the sampling sites.On average,the predictive maps based on OK interpolation were obviously higher than those based on IDW interpolation in accuracy before the map began to turn distorted.【Conclusion】All the findings in this study indicate that both sample size and spatial distribution pattern have certain impacts on predictive soil map,and the impacts are quite limited when the sample size is large enough.However,when the sample size drops below a certain level,the predictive maps will lose details of some local variations and prediction accuracy as well,while spatial distribution of sampling sites turns up to be the main affecting factor.Compared with OK,IDW is lower in accuracy and responds to the changes in sample size and spatial distribution earlier before the predictive map is distorted.