土壤制图中多等级代表性采样与分层随机采样的对比研究

A COMPARATIVE STUDY OF MULTI-GRADE REPRESENTATIVE SAMPLING AND STRATIFIED RANDOM SAMPLING FOR SOIL MAPPING

采样设计是土壤地理研究中备受关注的重要问题。本文以区域尺度土壤属性制图为例,将多等级代表性采样与经典采样中的分层随机采样进行对比研究。以安徽宣城研究区的表层砂粒含量为目标要素,采集数量均为59个的两套样点,设计不同数量(46、58和59)的样点分组,采用两种制图方法进行制图并利用独立验证点进行评价。结果表明:1)无论是采用多元线性回归方法还是基于环境相似度的制图方法,在同等样点数量下,利用代表性样点所得土壤图精度均高于利用随机样点所得精度,并且利用少量代表性样点(46个)所得土壤图精度也高于利用多量随机样点(59个)所得精度;2)随着代表性较低样点的增加,土壤制图精度基本有一个提高的趋势,而采用随机样点所得土壤图的精度波动较大。因此,可认为多等级代表性采样方法是一种可用于区域尺度土壤调查的有效采样方法,且比分层随机采样高效、稳定。

Sampling design has long been a key issue of concerns in the study of Soil Geography. How to take samples efficiently is now an important problem that investigators or researchers are tackling. As more and more synergic environmen- tal data become easily available, they can also be used to assist designing of sampling so as to improve sampling efficiency. For that end, a multi-grade representative sampling method is developed. Using this method, investigators can catch patterns of soil spatial variability at different scales through designing different representative grades based on the relationship be- tween soil and its environmental covariates. This method has been testified as an effective sampling method in a case study of watershed scale soil mapping. However, the application of this method is very limited, especially for mapping regional soil with more complicated pedogenesis. It is also unknown how it is when comparing with classic sampling methods, such as stratified random sampling. In this paper the multi-grade representative sampling method is compared with the stratified ran- dom sampling method in regional scaled soil property mapping, in the following two aspects ： 1 ） designing of sampling sites, and result and accuracy of mapping; and 2） variation of accuracy of mapping with increased number of sampling sites. The case study was laid out in Xuancheng of Anhui Province, China, covering an area of 5900 km2. Seven environ- mental variables were selected, i.e. parent material （rock） , slope gradient, profile curvature, contour curvature, topo- graphic wetness index, annual average precipitation, and annual average temperature. The data of parent rock was used to stratify the study area. For representative sampling, FCM was employed to cluster the environmental variables of each stra- tum for designing representative sampling sites. Consequently a total of 59 sampling sites featuring three representative grades were defined. For random sampling, the number of sampling sites in each stratum was proportional to its area and also a total of 59 sampling sites were marked out. The two sets of sampling sites were sorted separately into three groups, each consisting of 46, 58 and 59 sites. For representative sampling, the first group of sampling sites was the highest in representativeness grade. The second group was the first group adding the sampling sites with the second highest represent- ative grade. And the third group was the total sample set, For random sampling, the first group of sampling sites was 46 sampling sites picked out randomly from the set. The second group was the first group adding 12 sampling sites randomly drawn from the remaining （59 -46） samples. And the third group was the total sample set. Soil surface layer sand content maps were plotted using two different soil mapping methods, i.e. multivariate linear regression and similarity-based map- ping. The resultant soil maps were evaluated with independent validation sites and RMSE as evaluation index. Results show ： 1 ） no matter which mapping method was used, in the case of the same sampling size, the representa- tive sampling method was lower than the random sampling method in RMSE and even the former with fewer sampling sites （46） was lower than the latter with more sampling sites （59） ; 2） with increased number of sampling sites that were lower in representativeness, RMSE displayed a basically declining trend, while in random sampling method, increased number of sampling sites would aggravate the fluctuation of RMSE. It is, therefore, held that the multi-grade representative sam- nlinc, method is more efficient and stable than the stratified random sampling.