Professional Documents
Culture Documents
www.elsevier.com/locate/apgeochem
Abstract
Soil organic C (SOC) concentrations in topsoil samples taken at 678 sites in the grassland of Ireland were investi-
gated using statistics and geostatistics. SOC concentrations (WalkleyBlack method) follow a lognormal distribution,
with a median and geometric mean of 5.0%, and an arithmetic mean of 5.3%. Statistically signicant (P< 0.01) posi-
tive correlation between SOC and silt-plus-clay, and negative correlation between SOC and sand were observed, with
lower correlation (P=0.17) between SOC and pH. Lower SOC concentrations were associated with higher percentages
of land in tillage. In order to obtain a robust measurement of spatial structure, spatial outliers were detected, and
subsequently eliminated, using the local Morans I index. The spatial distribution of SOC concentrations based on
kriging interpolation showed coherent spatial patterns, with the highest values in the western coastal area, and rela-
tively low values in the inland and southeastern coastal areas; soils at higher elevation were also found to contain
higher SOC concentrations. These patterns are consistent with the distribution of rainfall within the country.
# 2003 Elsevier Ltd. All rights reserved.
1999). The data set from Brogan (1966) was obtained 1010 km2 sector of the national grid and sampled. Where
via a national soil survey, undertaken in 1964, which land cover by lake, forest, bog and mountain was less than
covers most parts of the Republic of Ireland. Despite 25%, 2 sites were selected; when greater than 75%, no
the time-lag since the data was generated, it was con- sample was taken. The 678 samples so obtained represent
sidered to still provide a valuable base against which approximately 50% of the total that could be accom-
subsequent and future measurements could be eval- modated at a frequency of 2 samples for each 1010 km2
uated. For this reason it has been used in this study. sector. The shortfall in the number of samples collected
can be accounted for by the low sampling-density in some
areas, e.g., in the west of the country, and in the Wicklow
2. Methods south Dublin areas which, though largely mountainous,
also included upland grassland sites. However, such sites
2.1. Sampling and analysis of SOC were often peaty with SOC in excess of 18% and were
excluded under the sampling protocol.
About 59% of the land surface of the Republic of The locations of the grassland sampling sites are
Ireland was recorded as being devoted to permanent pas- shown in Fig. 1. SOC concentrations were determined
ture in 1964 (CSO, 1965). Soils (010 cm depth) were sam- by the Walkley-Black (Cwb) method (Metson, 1956).
pled from under permanent pasture (Brogan, 1966). A Sand, silt-plus-clay contents and pH values were also
maximum of 2 sites were selected at random from each determined (Brogan, 1966).
Geostatistics (Matheron, 1963; Cressie, 1993; Webster Outliers, and especially spatial outliers, in the data set
and Oliver, 2001) uses the semi-variogram to measure can make the variogram exhibit erratic behaviour. Spa-
the spatial variability of a regionalized variable, and tial outliers are those values that are obviously dierent
provides the input parameters for the spatial interpola- from the values of their surrounding locations (Lalor
tion method of kriging, a term introduced by Matheron and Zhang, 2001). In this work, such outliers are iden-
in 1960 in honour of the work of the South African tied using an index, local Morans I (Anselin, 1995;
mining engineer D.G. Krige (1951). It relates the semi- Getis and Ord, 1996):
variogram, half the expected squared dierence between
zi z X
n
paired data values z(x) and z(x+h) to the distance, lag Ii 2
wij zj z 5
j1; j6i
h, by which locations are separated:
1
h E zx zx h2 1 where zi is the value of the variable z at location i; z is
2
the average value of z with the sample number of n; zj is
For discrete sampling sites, such as those that had the value of the variable z at all the other locations
been sampled in this study, the function is written in the (where j6i); 2 is the variance of variable z; and wij is a
form: weight, dened as the inverse of the distance dij between
locations i and j:
1 X
Nh
h zxi zxi h2 2 1
2Nh i1 wij 6
dij
where z(xi) is the value of the variable Z at location of Local Morans I can be standardised so that its sig-
xi, h is the lag, and N(h) is the number of pairs of nicance level can be tested based on the normal dis-
sample points separated by h. For irregular sampling, it tribution (Levine, 1999), when the distribution of the
is rare for the distance between the sample pairs to be raw data set is not too skewed. It should be noted that
exactly equal to h. Therefore, h is often represented by a when the raw data are too skewed, the normal approx-
distance interval. imation of the local Morans I may fail. When the stan-
A variogram plot is obtained by calculating values of dardized value of local Morans I is higher than the
the variogram at dierent lags. These values are then critical value of 1.96, it is concluded (P=0.05) that the
usually tted with a theoretical model: spherical, expo- sample under test is clustered with (or similar to) the
nential, or Gaussian (See Matheron, 1963; Cressie, surrounding samples; if lower than 1.96, it is con-
1993; or Webster and Oliver, 2001 for discussion). The cluded (P=0.05) that the sample under test is sig-
models provide information about the spatial structure nicantly dierent from the surrounding samples and is
as well as the input parameters for the kriging inter- considered to be a spatial outlier.
polation.
Kriging is regarded as an optimal spatial inter- 2.4. Data treatment with computer software
polation method, which is a type of weighted moving
average: The data sets were analysed using dierent software
X
n packages. The descriptive statistical parameters were
z^x0 li zxi 3 calculated with Microsoft Excel# and SPSS# (version
i1 10.0). Maps were produced with GIS software Arc-
View# (version 3.2) and its extension of Spatial Ana-
where z^x0 is the value to be estimated at the location lyst# (version 2). The geostatistics analyses were carried
of x0; z(xi) is the known value at the sampling site xi, out with GS+# (version 5.3) (Robertson, 2000) and
and li is a weight. There are n sites within the search Idrisi 32# (Release 2), and the local Morans I index was
neighbourhood around x0 used for the estimation, and calculated with CrimeStat# (version 1.0) (Levine, 1999).
the magnitude of n will depend on the size of the moving
search window and user denition. Kriging diers from
other methods (such as inverse distance weighted), in 3. Descriptive statistics
that the weight function li is no longer arbitrary, being
calculated from the parameters of the tted variogram 3.1. Probability distribution
model under the conditions of unbiasedness and mini-
mized estimation variance for the interpolation. Thus, Histograms of SOC concentrations determined by the
kriging is regarded as a best linear unbiased estimation WalkleyBlack method are shown in Fig. 2. That for the
(BLUE). raw data set has a long tail towards higher concentrations,
1632 D. McGrath, C. Zhang / Applied Geochemistry 18 (2003) 16291639
Fig. 2. Histograms of Cwb concentrations in soils of Ireland (n=678): (a) raw data; (b) logarithmically transformed data.
whereas that for the logarithmically-transformed dataset 5.0%. Due to the lognormal feature of Cwb, the geo-
can be satisfactorily modelled by a normal distribution. metric mean and median are recommended as its repre-
Table 1 shows the quantitative parameters of the sentative mean value, while the arithmetic mean
probability distribution and signicance level of the (average) should be discarded for this variable.
KolmogorovSmirnov test for conformance to a normal
distribution for the variables. The probability distribu- 3.3. Relationships between SOC and other factors
tion of Cwb is positively skewed and has a sharp peak
(positive kurtoses). The log-transformed data show Pearson (linear) correlation coecients between the 4
rather small skewnesses and kurtoses, and pass the KS variables were calculated. These are given in Table 3,
normality test at a signicance level of higher than 0.05. together with corresponding signicance levels.
On the other hand, contents of sand and silt-plus-clay
follow a normal distribution, and log-transformation
Table 2
makes the t to a normal distribution worse. pH values Basic statistic parameters of SOC concentrations and other
cannot pass either the normal or lognormal tests. How- related variables (%, except pH, n=678)
ever, its skewness and kurtosis are rather small. There-
fore, the log-transformed data sets of Cwb, and the raw Statistics Cwb Sand Silt+clay pH
data sets of sand, silt-plus-clay and pH were used for the Min 2.0 8 8 4.3
following multivariate analyses. 5% 3.0 21 36 5.1
25% 4.0 31 48 5.5
3.2. Mean values for SOC
Median 5.0 37 55 5.8
75% 6.1 44 63 6.2
Percentiles and commonly used estimators of location 95% 9.1 56 71 7.0
and spread were calculated (Table 2). The range of Cwb
varies from 2.0 to 17.8%, with the arithmetic mean of Max 17.8 87 88 7.8
5.3%. Both the geometric mean and median of Cwb are Average 5.3 37.7 54.9 5.9
GeoMean 5.0 36.1 53.7 5.8
Stdev 1.9 10.5 10.7 0.6
Table 1
Shape parameters of the probability distributions and sig-
nicance level of KolmogorovSmirnov test (KS p) (n=678)
Table 3
Statistics Cwb Sand Silt+clay pH Correlation coecients (lower-left side) and their signicance
levels (upper-right side) (n=678)
Raw data Skewness 1.66 0.32 0.28 0.70
Kurtosis 4.77 0.65 0.60 0.42 LnCwb Sand Silt+clay pH
KS p 0.00 0.17 0.73 0.00
LnCwb 0.00 0.00 0.17
Log-transformed Skewness 0.36 0.82 1.72 0.43 Sand 0.43 0.00 0.00
data Kurtosis 0.42 1.76 8.92 0.08 Silt+clay 0.22 0.91 0.00
KS p 0.11 0.01 0.01 0.00 pH 0.05 0.14 0.13
D. McGrath, C. Zhang / Applied Geochemistry 18 (2003) 16291639 1633
Fig. 3. Box-plot of SOC in dierent elevation groups (n=50, Fig. 4. Average SOC concentrations in each county as a func-
171, 301, 156 respectively). tion of proportion of agricultural land in tillage (n=27).
1634 D. McGrath, C. Zhang / Applied Geochemistry 18 (2003) 16291639
Fig. 5. SOC (Cwb) distribution in soils of Ireland and location of spatial outliers.
Fig. 7. Isotropic variogram model of LnCwb: (a) spatial outliers excluded (n=639, exponential model: C0=0.05; C0+C=0.11; a=40
km); (b) all samples (n=678).
As described above, in order to avoid introducing of the samples. Some outliers with abnormally lower
bias, spatial outliers were detected using standardized values than the majority of the samples occur in the
local Morans I (Levine, 1999); Samples with values western part, and two samples in the north-eastern
lower than 1.96 were dened as spatial outliers. The part.
calculation was based on the log-transformed Cwb data In order to obtain a robust variogram, the spatial
set. Altogether, 39 spatial outliers were detected (Fig. 5). outliers were excluded in the variogram calculation.
In the south-eastern part of the region, there are some However, all values, including the detected spatial out-
outliers with abnormally higher values than the majority liers, were included in the kriging calculation,
Fig. 8. Cross validation results showing the dierences between the estimated and actual values.
1636 D. McGrath, C. Zhang / Applied Geochemistry 18 (2003) 16291639
4.2. Spatial structure of SOC The range of 40 km in the model implies that the large-
scale spatial autocorrelation may extend to the eective
The variogram surface of LnCwb is illustrated in range of 120 km. Hence the present sampling density is
Fig. 6. sucient to reveal general patterns but short-range
It can be seen that the variogram is broadly isotropic, variations can only be disclosed by sampling at a higher
with the lowest values located in the centre, and it spatial density. The optimum sampling-scale for SOC in
increases in all the directions with the increase of lag. Irish soils is not known but it will be a compromise
Therefore, the experimental isotropic variogram for between what is most eective and what is feasible.
LnCwb was calculated (Fig. 7). For comparison, the Clearly the mean sampling-interval used here, ca. 7 km,
variogram for all samples (spatial outliers included) is is too great. Van Meirvenne et al. (1996) found in an
also shown in Fig. 7. investigation in Belgium that most of the change occur-
The experimental variogram has been tted with an red within about 4 km, so that the optimum sampling
exponential model. There is a signicant nugget eect of scale is almost certainly less than this.
0.05, which accounts for 45% of the total sill of 0.11. When compared with the variogram for all samples,
This shows that the small-scale variances are quite signicant improvement of the variogram is observed
strong. In this study, only grassland soils had been for the samples with the spatial outliers excluded. The
sampled. It is expected that the small-scale variances variogram for all samples is close to a nugget eect,
would be even stronger if all types of soils were sampled. which is mainly attributed to these spatial outliers.
4.3. Spatial distribution of SOC concentrations actual values regardless of sign) of LnCwb is 11.9%, and
the average of relative absolute errors is 15.6%. When
The parameters of the exponential model were used the data are back-transformed to Cwb, the median and
for kriging to produce the spatial distribution map of average value of the relative absolute errors are 18.8 and
SOC concentrations in soils of Ireland. A search region 24.2% respectively.
of 32 nearest-neighbours was applied in order to oset For the spatial interpolation, a cell size of 500500 m
the relatively high nugget eect. To test the eectiveness was chosen to divide the study area into a grid system
of the model, cross validation (i.e., back-calculation of containing 707 rows and 625 columns. Ordinary kriging
the observed values) was carried out with all the samples was used with a block size of 22 and a search neigh-
including spatial outliers. The results are shown in bourhood of 32 nearest-neighbours. The interpolated
Fig. 8. It can be seen that the under-estimated and over- values were then back transformed according lognormal
estimated points are irregularly distributed over the kriging. Fig. 9 shows the nal result of this spatial
study area. The spatial outliers detected using local interpolation process.
Morans I in the eastern part tend to be under-esti- The data range of the interpolated values is from 3.3
mated, while those in the western and northeast parts are to 8.2% Cwb. This is narrower than that of the raw data
over-estimated, which reconrms that they are outliers. set (Table 2), but is to be expected because of the
For the log-transformed data, the median of the relative smoothing eect of the spatial interpolation. However,
absolute errors (the dierence between the estimated and this smoothing eect helps to identify the general spatial
patterns and reduces both the local variations and the Research Fund 2002 of National University of Ireland,
negative eect of random errors. The spatial distri- Galway. The raw data sets were provided by TEA-
bution of SOC is generally consistent with elevation GASC. Helpful discussions with Prof. Mcheal O Cin-
(Fig. 10) and precipitation (Rohan, 1975; Collins and neide are acknowledged. The authors are grateful to Dr.
Cummins, 1996) in Ireland. Richard Howarth, Dr. Nils Gustavsson, and an anon-
The highest SOC concentrations are located in the ymous reviewer for their helpful comments and sugges-
western coastal area, where elevation and precipitation are tions to improve this paper.
the highest. In the northern part, there is a relatively high-
concentration area between counties Cavan, Roscommon,
Leitrim and Monaghan. On the eastern side, a small References
area of high concentrations is located in counties Dublin
Anselin, L., 1995. Local indicators of spatial association
and Wicklow. These areas also have relatively high ele- LISA. Geog. Anal. 27, 93115.
vation and precipitation. On the other hand, SOC con- Batjes, N.H., 1996. The total C and N in soils of the world.
centrations are rather low in the southeastern part, Eur. J. Soil Sci. 47, 151163.
which is consistent with the relatively low precipitation. Brogan, J.C., 1966. Organic carbon in Irish pasture soils. Irish
Even though the elevation is not low, the southeastern J. Agric. Res. 5, 169176.
part of Ireland is a rain-shadow area caused by the Burgess, T.M., Webster, R., 1980. Optimal interpolation and
mountains on the west side. Another area with isarithmic mapping of soil properties: II. Block kriging.
obviously low values is located in east Galway, south J. Soil Sci. 31, 333341.
Roscommon, west Longford and Westmeath. This area Chevallier, T., Voltz, M., Blanchart, E., Chotte, J.L., Eschen-
brenner, V., Mahieu, M., Albrecht, A., 2000. Spatial and
has low precipitation and low elevation. In west Kerry,
temporal changes of soil C after establishment of a pasture
the SOC concentration is not as high as expected, which on a long-term cultivated vertisol (Martinique). Geoderma
may be related to the low sampling density in this area. 94, 4358.
Finally, western Ireland generally has higher SOC con- Clark, I., Harper, W.V., 2000. Practical Geostatistics 2000.
centrations than eastern Ireland. More detailed quanti- Ecosse North America Llc, Columbus OH, USA.
tative comparison between SOC and precipitation and Collins, J.F., Cummins, T., 1996. Agroclimatic Atlas of Ire-
elevation may be carried out when more detailed sam- land. AGMETJoint Working Group on Applied Agri-
pling is available in the future. cultural Meteorology, Dublin.
Cressie, N.A.C., 1993. Statistics for Spatial Data, second ed.
John Wiley & Sons, New York.
CSO, 1965. Statistical Abstract of Ireland. Central Statistics
5. Conclusions
Oce, Dublin.
Dobermann, A., Goovaerts, P., Neue, H.U., 1997. Scale-
The SOC concentration (measured by the Walkley dependent correlations among soil properties in two tropical
Black method) in grassland of Ireland follows a log- lowland rice elds. Soil Sci. Soc. Am. J. 61, 14831496.
normal distribution, with a median value of 5.0%, Duncan, D.B., 1955. Multiple range and multiple F tests. Bio-
arithmetic mean of 5.3%, and geometric mean of 5.0%. metrika 11, 142.
Statistically signicant positive correlation exists Eswaran, H., Van den Berg, E., Reich, P., Kimble, J., 1995.
between SOC and silt-plus-clay; negative correlation Global soil carbon resources. In: Lal, R., Kimble, J., Levine,
between SOC and sand, and lower correlation between E., Stewwart, B.A (Eds.), Soils and Global Change. CRC
SOC and pH were found. Low SOC concentrations are Lewis Publishers, Boca Raton, FL, pp. 2743.
Frogbrook, Z.L., Oliver, M.A., 2001. Comparing the spatial
consistent with higher proportions of land in tillage. The
predictions of soil organic matter determined by two labora-
local Morans I index has been eective in this study in tory methods. Soil Use Manag. 17, 235244.
detecting spatial outliers. The spatial distribution of Getis, A., Ord, J.K., 1996. Local spatial statistics: an overview.
SOC concentrations in Ireland show a broad regional In: Longley, P., Batty, M. (Eds.), Spatial Analysis: Modelling
pattern, with the highest values in the west coastal areas, in a GIS Environment. GeoInformation International,
and relatively low values in the inland and south-eastern Cambridge, UK, pp. 261277.
coastal areas. This is generally consistent with the spa- Goovaerts, P., 1997. Geostatistics for Natural Resources Eval-
tial distribution of precipitation and elevation in the uation. Oxford University Press, New York.
country. Houghton, R.A., 1995. Changes in the storage of terrestrial
carbon since 1850. In: Lal, R., Kimble, J., Levine, E., Stew-
wart, B.A. (Eds.), Soils and Global Change. CRC Lewis
Publishers, Boca Raton, FL, pp. 4565.
Acknowledgements Jenny, H., 1980. The Soil Resource, Origin and Behavior.
Springer-Verlag, New York.
This study was sponsored by the Research Develop- Krige, D.G., 1951. A statistical approach to some basic mine
ment Fund of TEAGASC (Agriculture and Food valuation problems on the Witwatersrand. J. Chem. Metall.
Development Authority, Ireland) and the Millennium Min. Soc. S. Africa 52, 119139.
D. McGrath, C. Zhang / Applied Geochemistry 18 (2003) 16291639 1639
Lal, R., 2002. Soil carbon dynamics in cropland and rangeland. soil properties at dierent scales within three terraces of the
Environmental Pollution 116, 353362. Henare River (Spain). Cantena 33, 139153.
Lalor, G., Zhang, C.S., 2001. Multivariate outlier detection and Schimel, D.S., 1995. Terrestrial ecosystems and the carbon
remediation in geochemical databases. Sci. Total Environ. cycle. Global Change Biology 1, 7791.
281, 99109. Smith, P., Milne, R., Poulson, D.S., Smith, J.U., Falloon, P.,
Levine, N., 1999. CrimeStat (version 1.0): A Spatial Statistics Coleman, K., 2000. Revised estimates of the carbon migration
Program for the Analysis of Crime Incident Locations. Ned potential of UK agricultural land. Soil Use Manag. 16, 293295.
Levine & Associations, Annandale, VA. Stein, A., Brouwer, J., Bouma, J., 1997. Methods for compar-
Matheron, G., 1963. Principles of geostatistics. Econ. Geol. 58, ing spatial variability patterns of millet yield and soil data.
12461266. Soil Sci. Soc. Am. J. 61, 861870.
McGrath, D., 1980. Organic carbon levels in Irish soils. In: Van Meirvenne, M., Pannier, J., Hofman, G., Louwagie, G.,
Proc. Land Use Seminar on Soil Degradation, Wageningen, 1996. Regional characterization of the long-term change in
1315 October 1980. soil organic carbon under intensive agriculture. Soil Use
McGrath, D., McCormack, R.J., 1999. Signicance of heavy Manag. 12, 8694.
metal and organic micropollutants in soils. End of Project Webster, R., 1994. The development of Pedometrics. Geo-
Report. Teagasc, Ireland. derma 62, 115.
Metson, A.J., 1956. Methods of Chemical Analysis for Soil Webster, R., Oliver, M., 2001. Geostatistics for Environmental
Survey Samples. Dept. Sci. Indust. Res., New Zealand. 65-70. Scientists. John Wiley & Sons, Chichester.
Rohan, P.K., 1975. The Climate of Ireland. The Stationary Zhang, C.S., Selinus, O., Schedin, J., 1998. Statistical analyses
Oce, Dublin. on heavy metal contents in till and root samples in an area of
Robertson, G.P., 2000. Geostatistics for the Environmental southeastern Sweden. Sci. Total Environ. 212, 217232.
Sciences: GS+ Users Guide, Version 5. Gamma Design Zhang, C.S., Selinus, O., Wong, P., 2000. Spatial structures of
Software, MI. cobalt, lead, and zinc contents in tills in southeastern Swe-
Saldana, A., Stein, A., Zinck, J.A., 1998. Spatial variability of den. GFF (Trans. Geol. Soc. Stockholm) 122, 213217.