Professional Documents
Culture Documents
2009 SISand IKestimationof 3 Dsoiltextures
2009 SISand IKestimationof 3 Dsoiltextures
net/publication/262957019
CITATIONS READS
22 3,179
7 authors, including:
Li Baoguo Kelin hu
China Agricultural University China Agricultural University
240 PUBLICATIONS 6,229 CITATIONS 146 PUBLICATIONS 2,928 CITATIONS
Some of the authors of this publication are also working on these related projects:
Soil sustainable productivity under long-term chemical fertilization in rainfed area View project
All content following this page was uploaded by Kelin hu on 11 September 2015.
Abstract. The complex distribution characteristics of soil textures at a large or regional scale are difficult to understand
with the current state of knowledge and limited soil profile data. In this study, an indicator variogram was used to describe
the spatial structural characteristics of soil textures of 139 soil profiles. The profiles were 2 m deep with sampling intervals
of 0.05 m, from an area of 15 km2 in the North China Plain. The ratios of nugget-to-sill values (SH) of experimental
variograms of the soil profiles in the vertical direction were equal to 0, showing strong spatial auto-correlation. In contrast,
SH ratios of 0.48–0.81 in the horizontal direction, with sampling distances of ~300 m, showed weaker spatial auto-
correlation. Sequential indicator simulation (SIS) and indicator kriging (IK) methods were then used to simulate and
estimate the 3D spatial distribution of soil textures. The outcomes of the 2 methods were evaluated by the reproduction of
the histogram and variogram, and by mean absolute error of predictions. Simulated results conducted on dense and sparse
datasets showed that when denser sample data are used, complex patterns of soil textures can be captured and simulated
realisations can reproduce variograms with reasonable fluctuations. When data are sparse, a general pattern of major soil
textures still can be captured, with minor textures being poorly simulated or estimated. The results also showed that when
data are sufficient, the reproduction of the histogram and variogram by SIS was significantly better than by the IK method
for the predominant texture (clay). However, when data are sparse, there is little difference between the 2 methods.
Additional keywords: soil textures, spatial variability, 3D, SIS and IK.
SIS is increasingly preferred to kriging in cases where spatial characterise the spatial variability of different texture classes,
variation measured in the field must be preserved (Goovaerts and an SIS algorithm to model the 3-D distribution of a confining
2000). However, in practice, some studies show that the mean layer. However, they did not quantitatively validate the results
prediction error tends to be larger for simulated values than for of simulation and its corresponding variogram reproduction.
kriging estimates (Olea and Pawlowsky 1996; Goovaerts 1997, According to the study of Hengl et al. (2007), results for soil
1999). Therefore, selection of estimation or simulation to texture estimation may appear to be satisfactory visually but can
predict soil properties may involve trade-offs in terms of sometimes be misleading.
errors of the results, not only in prediction accuracy but also The objectives of this study are to explore the 3-D spatial
in the reproduction of spatial variability. structural characters of soil textures in an alluvial plain using
Soil textures have been assessed at regional scale through indicator geostatistics and to compare the performance of the
1-D simulation (Li et al. 1997, 1999) and 2-D estimation (Hengl SIS and IK methods by their reproduction of histogram and
et al. 2007). Such research provides useful prediction of the variogram, and accuracy of prediction. In addition, 2 datasets
heterogeneity of soil textures. However, 3-D results are required were used to demonstrate the data density sensitivity of the two
to provide detailed information and to allow visualisation of methods.
vadose zone complexity for applications in agricultural and
environmental management. Materials and methods
The description of the spatial variability of soil texture by IK
or SIS at regional scale is still developing. Grunwald et al. Study region
(2000) used Environmental Visualisation Software (EVS) to This research was conducted in an area of the plain of ~15 km2
export soil textures at differential scales. They used 2-D near the Quzhou Experiment Station, China Agricultural
horizontal ordinary kriging slices for interpolation; therefore, University, Hebei Province (368510 N, 115830 E) (Fig. 1). The
the variability in the vertical direction was not considered in the topography of the study region is mainly flat, only the south-west
3-D results. Hengl et al. (2007) used regression kriging to corner is low-lying land. Most of the area is a loamy depression
estimate the 2-D distribution of the textures of surface soils at in the Zhang River’s alluvial fan. The surface (the positive
regional scale, and Santanello et al. (2007) used remotely sensed direction in the vertical plane is upward) soil texture is primarily
estimates of soil moisture to infer soil texture. They introduced sandy loam, clay, and light loam, whereas the bottom is
auxiliary predictors which were useful when data were sparse. primarily sand and clay. The entire research region is
However, auxiliary information has still not been applied to farmland, cultivated to a depth of 0.2–0.3 m by small tractors
subsoil layers. Bierkens and Weerts (1994) used indicators to and oxen. The main crop rotations are wheat–maize,
Y
X Materials
Sand
N Sandy loam
Light loam
BeiJing Medium loam
TianJin Heavy loam
Study
area Clay
ShanXi
ShanDong
HeNan
7 0 7 14 km
Fig. 1. Location and soil profiles of study area in Quzhou county in Hebei, China.
624 Australian Journal of Soil Research Y. He et al.
wheat–cotton, and wheat–soybean. The soil organic matter by h. After calculation and variogram analysis, a suitable model
content is <1 g/kg in most of the area (Liu et al. 2005). can be derived.
Soil texture Sand Sandy loam Light loam Medium loam Heavy loam Clay
<0.01 mm particles (%) 0–10 10–20 20–30 30–45 45–60 >60
Simulation and estimation of 3-D soil textures Australian Journal of Soil Research 625
Indicator kriging for both IK and SIS, and the scatter plot of variogram v. distance
The aim of IK is to estimate the conditional cumulative was generated (Fig. 3). A theoretical variogram model was then
distribution function (ccdf) belonging to any category zk, used to fit the calculated values, and the model with the best-
conditional to data (n): fitting value (Isaaks and Srivastava 1989) was selected. The data
in the Y direction were sparse, and the spatial variability
I*ðu; zk Þ ¼ E * fIðu; zk ÞjðnÞg ¼ Prob * ðZðuÞ 2 zk jðnÞÞ ð4Þ structure in the Y and in the 458 (X–Y) directions was
similar to the X direction; we assumed that the variogram in
The algorithm IK assumes that the marginal probabilities the horizontal direction is symmetrical. This is similar to other
E * (ua; zk) are constant and known for all ua and zk (Goovaerts sedimentary variable studies, such as geology facies (Falivene
1997). et al. 2007).
For a categorical variable, the indicator variable for category The nugget-to-sill ratio (SH) designates the degree of spatial
k is defined as: heterogeneity arising from random components to that of the
total spatial heterogeneity (Robertson et al. 1993). A value close
1 if zðuÞ ¼ zk
iðu; zk Þ ¼ ð5Þ to 0 indicates that the variable has strong spatial auto-
0 otherwise relationship. Conversely, a value close to 1 indicates spatial
and the probability I * (u; k) for Z(u) belonging to a category is heterogeneity is dominated by randomness, or nugget effect.
estimated by simple kriging: From Table 2, we can see that soil texture layers in the vertical
direction have a strong spatial auto-relation (all SH = 0),
X
N
reflecting regular sedimentary deposition. However, soil
I*ðu; zk Þ EfIðu; zk Þg ¼ la ðIðua ; zk Þ EfIðua ; zk ÞgÞ textures in the horizontal direction exhibited some degree of
a ¼1
nugget effect, especially the light loam and medium loam
ð6Þ (>0.80), indicating random heterogeneity, and reflecting
The complicated formation conditions at the scale of sampling.
P n estimated probabilities must all be in [0, 1] and verify: The nugget effect is also called the variance of the minimal
k = 1 I *(u; zk) = 1. If not, they are corrected as follows:
practical error for sampling, which is a sum of all variances
(1) If I * (Z(u), zk) 2
= [0, 1] reset it to the closest boundary. If (Esbensen et al. 2007). For the soil texture data, the nugget effect
all the probability values are 0, no correction is made and may be caused by data errors, tillage, or data sparseness. In this
a warning is issued. study, the sampling interval may be the main factor for the
(2) Standardise the values so that they sum to 1: nugget effect in the horizontal direction, because the variance
with lags <300 m cannot be measured (Esbensen et al. 2007;
I*ðZðuÞ; zk Þ Hengl et al. 2007). If the correlation range is <300 m, the
I *corrected ðZðuÞ; zk Þ ¼ PK ð7Þ variogram becomes a pure nugget model. In the vertical
i ¼ 1 I*ðZðuÞ; zi Þ direction, the sample data are sufficient and the sampling
almost continuous. In this direction the variograms can be
The weights la in (4) were calculated from simple kriging estimated at even 1 pixel length lag (0.05m), which is why
system: the variograms have no nugget effect. However, the variograms
X
nðu 0Þ of soil textures in these 2 directions cannot be compared because
lk ðu0 ; zc ÞC I ðuk uj ; zc Þ ¼ C I ðuj u0 ; zc Þ of the great differences in their scales and sample intervals.
k¼1 The range is also part of the variogram. For lags within the
for j ¼ 1 to nðu0 Þ ð8Þ range, increments are progressively more auto-correlated, which
implies that sampling with these increment distances will
where n(u0) weights associated with the neighbouring data, zc capture the variation in the process with increasing reliability
denotes the corresponding indicators, CI (uk – uj) is the (Esbensen et al. 2007). Although all the ranges derived from
covariance between indicators at uk and uj, and CI (uk – uj) is fitted models were bigger than their corresponding sampling
the covariance between the sample point uj and the point being distances, for textures present at small proportions, such as light
estimated, with CI (h; zc) = E(zc)[1 – E(zc)] – g I (h; zc). Only loam and heavy loam in the horizontal direction (Fig. 3), the
limited data close to the location u0 being estimated can be ranges are insufficient. Therefore, only few pairs of indicator
efficiently used, therefore, the n(u0) << N (Goovaerts 1997). The data can be used for variogram computation, which is another
estimated value of variable characters can be drawn from the factor inducing a certain degree of nugget effect (Goovaerts
final probability distribution of IK results. 1997).
Vertical
0.30 0.40 Sandy loam 0.10 Light loam
Sand
0.08
0.30
0.20
0.06
0.20
0.04
0.10
Experimental variogram 0.10 0.02
MODEL
0.00 0.00 0.00
0.02 0.15
0.06
0.10
0.03 0.01
0.05
0.00 0.00 0.00
0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
Distance (m)
Fig. 3. Experimental indicator variograms (points) and corresponding fitted models (solid lines) for each texture
category based on dense data.
Table 2. Parameters of indicator variogram models in the vertical 0–1.6 m depth. The light loam and medium loam appeared in
direction with nested structures and in the horizontal direction both the upper and lower layers, but only to a small degree. In the
C0 is nugget, C is sill, a is range, and SH is nugget-to-sill ratio value south-west corner of the sampled area (Z = 1.7–2.0 m), clay is
(C0/(C0 + C)); all fitted models used were spherical present at the soil surface. The realisation is therefore visually
Soil texture C0 C1 a1(m) C2 a2 (m) SH consistent with the observed data.
(experimental) Figure 5a shows the SIS results based on the sparse dataset
(Fig. 2b). Obviously, the general pattern is captured in the
Vertical direction simulated map. However, due to the sparseness of the
Sand 0.00 0.13 0.5 0.100 1.2 0.00
conditioning data, the minor variables (such as light loam)
Sandy loam 0.00 0.12 0.5 0.080 1.3 0.00
Light loam 0.00 0.065 0.3 0.015 1.3 0.00
are underestimated, which is similar to the results of Li and
Medium loam 0.00 0.06 0.3 0.027 0.8 0.00 Zhang (2006).
Heavy loam 0.00 0.01 0.4 0.011 0.8 0.00
Clay 0.00 0.15 0.4 0.050 0.9 0.00 Three-dimensional IK
Horizontal direction A 3-D IK estimation was also conducted on a 170 50 80
Sand 0.13 0.10 2100 0.57 grid. The 3-D and cross-section results for IK are shown in
Sandy loam 0.12 0.08 2200 0.60 Fig. 6. In order to compare the SIS realisation with the IK
Light loam 0.065 0.015 1100 0.81
estimation, the results were derived from the estimated
Medium loam 0.06 0.027 2100 0.69
probability of IK. For convenience, the assigned value on
Heavy loam 0.01 0.011 950 0.48
Clay 0.15 0.05 2700 0.75 estimated grid was based on the maximum probability of
each soil texture, which is different from other reports
(Goovaerts et al. 1997; Castrignanò et al. 2000; Lin et al. 2002).
50 m by 0.025m); the total number of grids was ~680 000. The From Fig. 5, the results (based on dense datasets) reflected
realisation was based on the dense dataset. The cross-sections of the measured distribution. For example, sand layers are mostly
the SIS realisation (Fig. 4) showed that where the subsoil was distributed in the surface, and the sandy loam and clay layers are
clay, sandy loam dominated the upper layer (1.6–2.0 m). When mainly distributed at depth. A clay layer was shown at the
the bottom was sand and clay, the sand layer mainly appeared in surface in the south-west. In addition, from the realisations of
Simulation and estimation of 3-D soil textures Australian Journal of Soil Research 627
Materials Materials
Sand Sand
Sandy loam Sandy loam
Light loam Light loam
Medium loam Medium loam
2.00 Heavy loam 2.00 Heavy loam
Clay Clay
1.00 1.00
0.00 0.00
0
0
500
500
1000
1000
1500
1500
2000
2000
2500
2500
3000
Z Z
3000
3500
2000
3500
4000
2000
4500
4000
1000
5000
4500
1000
Y Y
5500
5000
6000
5500
6000
X X
(a) 3-D IK estimation
(a) 3-D SIS realisation
2000
2000
1000
1000
0
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
2.00 2.00
1.00 1.00
0.00 0.00
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
Fig. 4. 3-D realisation by SIS and its corresponding cross-section Fig. 5. 3-D estimation by IK and its corresponding cross-section
conditioned on the dense dataset. conditioned on the dense dataset.
SIS and IK, they had similar general trends, although the The histograms of SIS and IK are shown in Fig. 7 and
distribution of SIS realisation seems less ordered and the IK Table 3. In comparison with the frequency distribution of texture
realisation was smoothed. classes derived from the data (Fig. 2), the overall proportions
Similar to results of SIS, when using the sparse dataset, the were reproduced reasonably well by both SIS and IK (Fig. 7)
general pattern can be captured in the estimated map (see conditioned on the dense dataset. For the predominant textures,
Fig. 8b) and the minor textures were also relatively when using the dense dataset, SIS is better than IK (e.g. the
underestimated (e.g. light loam) in IK. MAPE of clay for SIS is 0.4, whereas for IK it is 17.9). However,
when using the sparse dataset, there is no obvious difference
Model validation between SIS and IK. In addition, the textures present in small
Reproduction of histogram and variogram proportions (light loam and medium loam) were underestimated
by both SIS and IK when using the dense dataset and the case
The histograms and variograms were calculated on grid was serious when using the sparse dataset (Table 4, Fig. 8). This
points (32 10 40 grids, grid size is 175 m by 250 m by indicates that the reproduction of such small variables is difficult
0.05 m) from the SIS and IK results. This scale is similar to in these methods if there are many category variables, as also
the measured data. The histograms and variograms of simulated found by Regli et al. (2004).
and experimental results were compared by absolute percentage The variogram reproduction for SIS and IK is shown in
error (APE): Tables 5, 6; Figs 3, 9). When using the dense dataset, there
At A i was little difference in the horizontal directions between SIS
APE ¼ ð9Þ
At and IK. However, in the vertical direction, the MAPE for IK
was much bigger than for SIS for clay (i.e. predominant
The difference between measured value At and predicted Ai is textures). When using the sparse dataset, the variograms
divided by the measured value At. were not well reproduced (except for clay). In addition, when
628 Australian Journal of Soil Research Y. He et al.
0.50
Measured
0.40 SIS
IK
Frequency 0.30
0.20
0.10
0.00
Sand Sandy loam Light loam Medium loam Heavy loam Clay
Soil texture
Fig. 6. Histogram reproduction of the frequency distribution of the 6 textures by SIS and IK conditioned
on the dense dataset.
Vertical
0.40 0.10 Light loam
Sand Sandy loam
0.40
0.08
0.30
0.30
0.06
0.20 0.20 0.04
SIS
0.10 IK 0.10 0.02
MODEL
0.00 0.00 0.00
0.09 Medium 0.03 Heavy loam 0.25 Clay
loam 0.20
0.06 0.02
0.15
Indicator variogram
0.03 0.10
0.01
0.05
0.00 0.00 0.00
0.0 0.3 0.6 0.9 1.2 1.5 0.0 0.3 0.6 0.9 1.2 1.5 0.0 0.3 0.6 0.9 1.2 1.5
Horizontal
0.30 Sand 0.25 Sandy loam 0.15 Light loam
0.20 0.12
0.20
0.15 0.09
0.10 0.06
0.10
0.05 0.03
0.00 0.00 0.00
Fig. 7. Indicator variogram reproduction by SIS realisation and IK estimation based on sparse dataset.
Soil texture type: Sand Sandy loam Light loam Medium loam Heavy loam Clay
Frequency of histogram Measured 0.308 0.287 0.075 0.078 0.015 0.230
SIS 0.376 0.276 0.052 0.049 0.015 0.231
IK 0.410 0.296 0.048 0.050 0.018 0.195
APE SIS 18.1 4.00 44.2 59.2 0.00 0.400
IK 24.9 3.00 56.3 56.0 16.7 18.0
using the dense dataset, SIS does not work well for textures dataset, indicating that when large-range continuity is present
present in small proportions (light loam, medium loam, and the SIS does not provide good variogram reproduction
heavy loam) and the case was even worse when using the sparse (Goovaerts 1999; Bastante et al. 2008) and sensitive to data
Simulation and estimation of 3-D soil textures Australian Journal of Soil Research 629
Table 4. Mean absolute percentage error of SIS and IK indicator variogram reproduction
Soil texture type: Sand Sandy loam Light loam Medium loam Heavy loam Clay
Vertical SIS 12.1 19.2 16.4 85.6 41.5 2.40
IK 16.5 22.1 92.3 101.7 25.0 30.5
Horizontal SIS 9.60 10.2 8.50 15.5 9.40 6.80
IK 12.3 9.7 13.8 10.0 5.70 6.80
(a) (b)
Materials
Sand
2000 2000
Sandy loam
Light loam
1000 1000 Medium loam
Heavy loam
0 0 Clay
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
Fig. 8. (a) SIS simulation and (b) IK estimation based on the sparse dataset (Z = 2 m).
Table 5. Absolute percentage error of SIS and IK indicator variogram reproduction conditioned on the sparse dataset
Soil texture type: Sand Sandy loam Light loam Medium loam Heavy loam Clay
Vertical SIS 15.2 21.9 185.6 452.1 80.7 9.8
IK 9.0 17.4 341.5 215.7 83.8 8.8
Horizontal SIS 12.6 20.2 164.4 244.6 30.9 5.0
IK 13.5 24.0 304.6 331.9 19.8 3.3
Table 6. Absolute percentage error of SIS and IK histogram reproduction conditioned on the sparse dataset
Soil texture type: Sand Sandy loam Light loam Medium loam Heavy loam Clay
Frequency of histogram Measured 0.308 0.287 0.075 0.078 0.015 0.230
SIS 0.348 0.325 0.027 0.027 0.013 0.260
IK 0.333 0.350 0.017 0.028 0.014 0.258
APE SIS 11.5 11.7 177.8 188.9 15.4 11.5
IK 7.5 18.0 341.2 178.6 7.1 10.9
0.50
Measured
0.40 SIS
Frequency
IK
0.30
0.20
0.10
0.00
Sand Sandy loam Light loam Medium loam Heavy loam Clay
Soil texture
Fig. 9. Histogram reproduction of the frequency distribution of the 6 textures by SIS and IK conditioned
on the sparse dataset.
630 Australian Journal of Soil Research Y. He et al.
Leij FJ, Dane JH, van Genuchten MT (1991) Mathematical analysis of one- Regli C, Rosenthaler L, Huggenberger P (2004) GEOSSAV: a simulation
dimensional solute in a layered soil profile. Soil Science Society of tool for subsurface applications. Computers & Geosciences 30, 221–238.
America Journal 55, 944–953. doi: 10.1016/j.cageo.2003.10.007
Li WD, Li BG, Shi YC (1999) Markov-chain simulation of soil textural Remy N, Boucher A, Wu JB (2008) ‘Applied Geostatistics with SGeMS.’
profiles. Geoderma 92, 37–53. doi: 10.1016/S0016-7061(99)00024-5 (Cambridge University Press: Cambridge, UK)
Li WD, Li BG, Shi YC, Tang DY (1997) Application of the Markov chain Robertson GP, Crum JR, Ellis BG (1993) The spatial variability of soil
theory to describe spatial distribution of textural layers. Soil Science resources following long-term disturbance. Oecologia 96, 451–456.
162, 672–683. doi: 10.1097/00010694-199709000-00009 doi: 10.1007/BF00320501
Li WD, Zhang C (2006) A generalized Markov Chain Approach for Santanello JA Jr, Peters-Lidard CD, Garcia ME, Mocko DM, Tischler MA,
conditional simulation of categorical variables from grid samples. Moran MS, Thoma DP (2007) Using remotely-sensed estimates of soil
Transactions in GIS 10(4), 651–669. doi: 10.1111/j.1467-9671. moisture to infer soil texture and hydraulic properties across a semi-arid
2006.01017.x watershed. Remote Sensing of Environment 110, 79–97. doi: 10.1016/
Lin YP, Chang TK, Shih CW, Tseng CH (2002) Factorial and indicator j.rse.2007.02.007
kriging methods using a geographic information system to delineate Weerts HJT, Bierkens MFP (1993) Geostatistical analysis of overbank
spatial variation and pollution sources of soil heavy metals. deposits of anastomosing and meanding fluvial systems – Rhine
Environmental Geology 42, 900–909. doi: 10.1007/s00254-002-0600-5 Meuse Delta, The Netherlands. Sedimentary Geology 85, 221–232.
Liu YH, Yu ZR, Zhang FR, Song CM, Liu Y (2005) Dynamic change of soil doi: 10.1016/0037-0738(93)90085-J
organic matter and its affecting factors at county level. Plant Nutrition Zhou LZ, Selim HM (2001) Solute transport in layered soils: nonlinear and
and Fertilizer Science 11(3), 294–301. kinetic reactivity. Soil Science Society of America Journal 65,
Olea RA, Pawlowsky V (1996) Compensating for estimation smoothing in 1056–1064.
kriging. Mathematical Geology 28, 407–417. doi: 10.1007/BF02083653
Plante AF, Conant RT, Stewart CE, Paustian K, Six J (2006) Impact of soil
texture on the distribution of soil organic matter in physical and chemical
fractions. Soil Science Society of America Journal 70, 287–296.
doi: 10.2136/sssaj2004.0363 Manuscript received 29 September 2008, accepted 6 May 2009
http://www.publish.csiro.au/journals/ajsr