基于土壤变异解释力的几种土壤制图方法的对比研究——以南阳市1m土体土壤有机碳密度制图为例
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

基金项目:

国家自然科学基金项目(40801080,41601210,40971128)、科技部科技支撑计划项目(2012BAD05B02-7)


Comparison between Soil Mapping Approaches Based on Their Ability Explaining Soil Variability-A Case of Mapping Soil Organic Carbon Density of Soil (0~1m)in Nanyang District
Author:
Affiliation:

Fund Project:

the National Nature Science Fund of China (No.40801080,41601210,40971128) and Science and Technology Support Program of Chinese Science and Technology Ministry(No.2012BAD05B02-7)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为克服方法的复杂性和数据的详细性解释土壤制图效果的不足,基于土壤变异解释力对多种方法进行对比研究。收集南阳1:5万土壤类型图、30 m分辨率数字高程模型和TM影像,计算出高程、坡度、坡向、归一化植被指数(NDVI)、穗帽变换的湿度(TCW)参数等,以439个土壤剖面为训练数据,分别按土壤类型连接法(SCLM)、加权最小二乘法(WLS)回归、地理权重(GWR)回归、随机森林(RF)、普通克里格(OK)、回归克里格(RK)进行1m土体土壤有机碳密度(SOCD)制图, 其余49个土壤剖面作为验证集。结果表明:(1)对SOCD变异的解释力是影响制图效果的本质因素。土壤类型、土壤表层有机质(OM)是主要预测变量,SCLM、WLS和GWR均只能利用其中一种主要变量,土壤图的详细化和回归模型的复杂化均不能明显改善SOCD制图效果。基于土属和OM变量,RF对SOCD变异的解释力最强,预测效果最优;地统计学空间变异函数对SOCD变异的解释力大于回归模型,小于RF,而与土壤类型相当,其相对制图效果亦如此。(2)预测变量建模和空间相关是两类不同的土壤变异解释机制,RK未必能使它们产生最佳组合:只有WLS回归、GWR回归和缺乏土壤类型信息的RF(OM+TCW)适合RK算法,在原始模型中它们对训练数据的拟合效果依次升高,但其RK结果的优劣排序则相反;所有RK的结果均未达到土属和OM参与下RF制图的精度。

    Abstract:

    【Objective】Before the digital soil mapping technology emerged, the soil category linkage method (SCLM),linking means or median values of properties of the soils of the same soil category with their corresponding polygons in the soil map, or linking soil properties with polygons based on pedological expertise (including type of the soil and its location), was the major method used in mapping of soil organic carbon density (SOCD). Even nowadays, it is still of quite high practical value, because it is quite hard to build up a DSM model for relationships of external environmental covariates with SOCD in deep soil layers and/or on large scale, e.g. Provincial, continental and global. To understand in-depth relative efficiency of the two linking methods, it is necessary to perform some comparative studies. In terms of the DSM technology, most of the comparative studies have come to the conclusions that sophisticated machine learning models are superior to simple ones and that mixed models (regression Kriging) are of high superiority in most cases. However, there are a few papers reported some contradictory results. All the conclusions suggest that SOCD mapping quality could not be explained merely by method and also affected by the effectiveness and accuracy of the parameters used in the method. To elaborate in-depth the contradictory conclusions and to analyze the essence of the problems, in this paper a comparison was performed of SCLM with weighted least squares regression(WLS), geographically weighted regression(GWR), random forest (RF),ordinary kriging (OK) and regression kriging (RK) in SOCD mapping, and establishment of relationships between abilities of the methods to explain SOCD variability and effects of their mapping was discussed. 【Method】A tract of land, 26600 km2 in area, in Nanyang of Henan Province, was selected as a study area, of which soil categories, elevation, slope, aspect, and normalized difference vegetation index (NDVI), and wetness (TCW) of tasseled cap transformation (TC) were worked out as parameters of the SOCD prediction model, based on a 1:50 000 soil map, Digital Elevatation Model, 30m in resolution and a 1990 thematic mapper(TM) image. A total of 439 soil profiles were cited as training dataset for SOCD mapping using SCLM, WLS, GWR, RF, OK and RK, separately, and another 49 soil profiles were used to verify accuracy of the maps. 【Result】Results show that soil genus and soil organic matter content of the topsoil layer is the most important and the second most important parameter, explaining jointly 57.5% of the SOCD variance, while terrain and remote sensing parameters jointly explain just less than 2%, and hence are very limited in contribution to SOCD mapping. However, SCLM makes use only of variables in soil category, like soil group, soil subgroup, soil genus, etc., while regression methods, like WLS and GWR, can only use numerical variables, like SOM and TCW, so none of these can achieve satisfactory prediction result. RF is based on both variables in soil category and numerical variables (SOM and TCW) and hence much better in SOCD prediction. The use of RK in prediction may end up in the following two situations. 1) The residues of WLS regression, GWR regression, and soil OM-and-TCW-based RF vary spatially and structurally to a varying extent, then regression kriging (RK) could improve the SOCD predictions of these models. 2) The residues of the predictions using SCLM, SOM+soil genus-based RF, SOM+soil genus+TCW based RF and all-variables-based RF vary spatially and randomly, for which the use of RK is meaningless. The cross-verified accuracy of WLS, GWL and soil OM-and-TCW-based RF increased in turn, however their RK ability predicting test data are reversed. And all prediction ability of RKs do not reach as high as the SOM+soil genus-based RF.【Conclusion】All the findings demonstrate that the ability of method to explain SOCD variability is the causa essentiae deciding the effect of SOCD mapping, and RK is not necessarily the fittest model because of the interactions in explanation ability between covariates and the spatial correlation.

    参考文献
    相似文献
    引证文献
引用本文

赵彦锋,李豪杰,陈 杰,孙志英,梁思源.基于土壤变异解释力的几种土壤制图方法的对比研究——以南阳市1m土体土壤有机碳密度制图为例[J].土壤学报,2018,55(1):43-53. DOI:10.11766/trxb201703270534 ZHAO Yanfeng, LI Haojie, CHEN Jie, SUN Zhiying, LIANG Siyuan. Comparison between Soil Mapping Approaches Based on Their Ability Explaining Soil Variability-A Case of Mapping Soil Organic Carbon Density of Soil (0~1m)in Nanyang District[J]. Acta Pedologica Sinica,2018,55(1):43-53.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2016-11-04
  • 最后修改日期:2017-08-25
  • 录用日期:2017-10-11
  • 在线发布日期: 2017-10-30
  • 出版日期: