中国农业科学 ›› 2020, Vol. 53 ›› Issue (18): 3716-3728.doi: 10.3864/j.issn.0578-1752.2020.18.008

• 土壤肥料·节水灌溉·农业生态环境 • 上一篇    下一篇

基于3种空间预测方法的黄土区土壤颗粒组成空间分布研究—以宁夏海原县为例

申哲1(),张认连1,龙怀玉1,王转1,朱国龙1,石乾雄2,喻科凡1,徐爱国1()   

  1. 1中国农业科学院农业资源与农业区划研究所,北京 100081
    2北京大学地球与空间科学学院,北京100871
  • 收稿日期:2019-12-10 接受日期:2020-03-17 出版日期:2020-09-16 发布日期:2020-09-25
  • 通讯作者: 徐爱国
  • 作者简介:申哲,Tel:16619922154;E-mail: 18211097094@163.com
  • 基金资助:
    国家重点研发计划(2017YFD0200607);科技基础性工作专项(2012FY112100);中国农业科学院基本科研业务费专项(Y2020PT37);中国农业科学院基本科研业务费专项(1610132019043)

Research on Spatial Distribution of Soil Particle Size Distribution in Loess Region Based on Three Spatial Prediction Methods—Taking Haiyuan County in Ningxia as an Example

SHEN Zhe1(),ZHANG RenLian1,LONG HuaiYu1,WANG Zhuan1,ZHU GuoLong1,SHI QianXiong2,YU KeFan1,XU AiGuo1()   

  1. 1Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081
    2School of Earth and Space Sciences, Peking University, Beijing 100871
  • Received:2019-12-10 Accepted:2020-03-17 Online:2020-09-16 Published:2020-09-25
  • Contact: AiGuo XU

摘要:

【目的】探索适合地形复杂的黄土母质地区土壤颗粒组成的空间预测方法。【方法】以宁夏自治区海原县为研究区域,结合地形因子、土壤类型、归一化植被指数变量,采用基于对称对数比转换的经验贝叶斯克里格法(SLR-EBK)、回归克里格法(SLR-RK)、随机森林(SLR-RF)3种方法对训练集100个样点表层土壤颗粒组成的空间分布进行预测,并通过验证集24个样点比较了3种方法的预测精度。【结果】(1)最终进入土壤颗粒组成线性回归预测模型的辅助变量包括高程(Ele)和土壤类型;进入RF模型的辅助变量包括高程(Ele)、土壤类型、坡度(Slo)和风力作用指数(WEI),其中,高程(Ele)是最重要的辅助变量,其次是土壤类型,坡度(Slo)和风力作用指数(WEI)重要性相对较低。(2)3种方法预测的海原县土壤各粒级含量空间分布的趋势基本一致,表现为砂粒含量西南部低,东北部高,粉粒、黏粒则相反。与SLR-EBK相比,SLR-RK和SLR-RF能够更好地反映局部变异并减小平滑效应。(3)SLR-RF法对验证集3个粒级含量预测的平均绝对误差(MAE)和均方根误差(RMSE)均低于其他两种方法,且从平均Aitchison距离(MAD)来看,SLR-RF(0.208)

关键词: 土壤颗粒组成, 空间预测, 经验贝叶斯克里格法, 回归克里格法, 随机森林, 对称对数比转换

Abstract:

【Objective】 This study was oriented to explore for a model that was capable of predicting spatial distribution of soil particle size distribution in the region with complicated terrain and single parent material, like Haiyuan County, Ningxia. 【Method】 Empirical Bayesian Kriging with the symmetry logratio transform (SLR-EBK), regression Kriging with the symmetry logratio transform (SLR-RK) and random forest with the symmetry logratio transform (SLR-RF) were used to predict spatial distribution of soil particle size distribution in Haiyuan County based on the training set of 100 samples, combined with topographic factors, normalized difference vegetation index and soil types, and then the predictions were validated with a validating set of 24 validation points in the study area for comparison of prediction accuracy. 【Result】 (1) The auxiliary variables that came into the linear regression equator for prediction included elevation (Ele) and soil types. The auxiliary variables that came into the RF model included Ele, soil type, slope (Slo) and wind exposition index (WEI), of which Ele was the most important auxiliary variable, followed by soil types. Slo and wind exposition index (WEI) were less important. (2) Results predicted by three methods showed a spatial distribution trend that the sand content was lower in the southwest and higher in the northeast of the county, while the silt content and clay content were higher in the southwest and lower in the northeast. SLR-RK and SLR-RF could better describe local variation of different soil particle size contents. (3) The mean absolute error (MAE) and root-mean-square error (RMSE) of SLR-RF were lower than those of the other two methods. As for mean Aitchison distance (MAD), the sequences of MAD were obtained as following: SLR-RF (0.208)

Key words: soil particle size distribution, spatial prediction, empirical Bayesian Kriging, regression Kriging, random forest, symmetric log-ratio transform