Assessing geostatistical methods for presenting urban soil geochemical data from Coventry

The current shortest distance between sample points in urban soil surveys of the BGS GSUE (Geochemical Surveys in Urban Environments) project is around 350 metres. Geostatistical analysis using soil geochemical data from the survey of Stoke-on-Trent had shown that much of the variation occurs at shorter sampling intervals (less than 350 metres). This means that the uncertainties associated with estimating values at unsampled locations (using kriging) are likely to be relatively large. A project was designed to address three specific, and related, questions. First, what is the nature of the short-scale variability of major and trace elements in urban environments? Second, would recently published, robust geostatistical methods be more appropriate for producing interpolated maps of urban soil geochemistry in which there may be two processes operating; a background process and a quasi-point contaminant process? Third, is the current sampling resolution (adopted in urban surveys) appropriate? To address the first question we undertook a ‘nested’ survey at selected nodes within the urban area of Coventry where a standard GSUE survey had recently been undertaken. These samples were analysed for the same suite of major and trace elements as the standard survey, and also for their particle-size distribution (proportions of sand, silt and clay). In the case of the typical contaminant type elements (Pb, Sn, Sb, Cd, Ni, Cu and Zn) a large proportion (22-75%) of the variation occurred at spatial scales of 30 metres or less, compared to a range of major and trace elements and particle size classes. Without further chemical analysis, it is not possible to determine whether this variation is due to anthropogenic impacts (pollution) or natural variation. However, many of these elements are common environmental pollutants suggesting that their greater short-scale variability may result from human activities. The information on variability at short spatial scales collected in the nested survey enabled us to plot variograms for a range of elements in which almost all the spatial variation was captured. Robust geostatistical methods may be more appropriate when dealing with datasets in which there are a considerable number of outlying values. An assessment of conventional and robust geostatistical estimators was undertaken based on five elements with significantly skewed distributions (Cd, As, Pb, Zn and Ni). Based on the results of a cross-validation exercise, the conventional geostatistical estimator (that due to Matheron) was found to be optimal for estimating Cd, As, Pb and Ni. However, in the case of the data for Zn, in which there were a considerable number of outlying values, a robust estimator (Cressie-Hawkins) performed best. If optimal interpolation methods are to be used in mapping urban soil geochemistry, it is recommended that when a large number of outliers are present in a dataset, a comparison of robust and classical estimators is undertaken. In the case of Zn, another statistical technique was used to identify 29 spatial outliers - samples which appear to be the outcome of a quasi-point process. Interpolated maps were generated both with and without the spatial outliers to determine the scale of their impact on the background process. In the case of the latter, there was a significant difference in the distribution in the region of the highest values. In addition, a number of the ‘bullseye’ patterns which were associated with spatial outliers have been removed. This is a useful technique for separating the background process from the effects of a quasi-point process. However, it is computationally demanding and time-consuming. It remains to be seen whether a customer from a local authority would be prepared to pay for this level of skilled analysis in the preparation of contour maps. To determine whether robust geostatistcal methods should be applied to other urban data, there is a need to determine the skewness coefficients and the number of outliers in datasets from the other urban centres. This will provide a better understanding of whether the Coventry data is typical of the other centres for which data are available. Geostatistcal analysis of urban geochemical data to date has indicated that the current sampling resolution adopted in urban surveys does capture a varying proportion of the spatially correlated variation for most of the contaminant type elements. On the basis of the work presented here we would not advocate changing the current (4 samples per square kilometre) sampling resolution. However, when producing a continuous surface map, it is essential first to undertake exploratory data analysis and construct variograms of the data to assess the degree of autocorrelation prior to kriging. If little or no autocorrelation is present, it is preferable to present the data as proportional symbols because there is little or nothing to be gained from interpolation. When a continuous surface map has been produced it should be accompanied by a description of the method used to create it, and a statement that the contour intervals represent estimates, not true values. The nested survey also identified Cd (cadmium) contamination at the site of a series of allotments. Comparison with recently published soil guideline values suggests that it may represent a potentially significant risk to human health. This issue is currently being raised with the City Council. To increase the utility of the soil geochemical data collected in the urban environment, we recommend that it would be beneficial to focus on perhaps two or three key sites in an urban area where human exposure to contamination may be significant, such as allotments, children’s nurseries or groundwater protection zones. These sites could be selected in conjunction with the City/Local Council on the basis of land use information which may indicate the likely presence of historical contamination.