See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/51185057

The assessment of spatial distribution of soil salinity risk using neural

network

Article  in  Environmental Monitoring and Assessment · June 2011

DOI: 10.1007/s10661-011-2132-5 · Source: PubMed

CITATIONS READS

41 182

2 authors:

Akmal Akramkhanov Paul L. G. Vlek

International Center for Agricultural Research in the Dry Areas (ICARDA) University of Bonn
39 PUBLICATIONS   615 CITATIONS    370 PUBLICATIONS   12,094 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

GLOWA Volta View project

Soil quality degradation assessment View project

All content following this page was uploaded by Akmal Akramkhanov on 22 April 2015.

The user has requested enhancement of the downloaded file.

Environ Monit Assess
DOI 10.1007/s10661-011-2132-5
The assessment of spatial distribution of soil salinity risk
using neural network
Akmal Akramkhanov · Paul L. G. Vlek
Received: 14 September 2010 / Accepted: 5 May 2011
© Springer Science+Business Media B.V. 2011
Abstract Soil salinity in the Aral Sea Basin is one
of the major limiting factors of sustainable crop
production. Leaching of the salts before planting
season is usually a prerequisite for crop estab-
lishment and predetermined water amounts are
applied uniformly to fields often without discern-
ing salinity levels. The use of predetermined wa-
ter amounts for leaching perhaps partly emanate
from the inability of conventional soil salinity
surveys (based on collection of soil samples, lab-
oratory analyses) to generate timely and high-
resolution salinity maps. This paper has an ob-
jective to estimate the spatial distribution of soil
salinity based on readily or cheaply obtainable en-
vironmental parameters (terrain indices, remote
sensing data, distance to drains, and long-term
groundwater observation data) using a neural net-
work model. The farm-scale (∼15 km2 ) results
were used to upscale soil salinity to a district area
(∼300 km2 ). The use of environmental attributes
and soil salinity relationships to upscale the spatial
distribution of soil salinity from farm to district
scale resulted in the estimation of essentially sim-
ilar average soil salinity values (estimated 0.94
A. Akramkhanov (B) · P. L. G. Vlek
Center for Development Research (ZEF),
University of Bonn, Walter-Flex-Str. 3,
Bonn, Germany 53113
e-mail: akmal@zef.uzpak.uz, api001@yahoo.com
vs. 1.04 dS m−1 ). Visual comparison of the maps
suggests that the estimated map had soil salinity
that was uniform in distribution. The upscaling
proved to be satisfactory; depending on critical
salinity threshold values, around 70–90% of loca-
tions were correctly estimated.
Keywords Irrigated agriculture · Upscaling ·
Validation · Spatial variation ·
Environmental correlation
Introduction
Soil salinity in the Aral Sea Basin is one of the ma-
jor limiting factors of sustainable crop production.
Leaching of the salts before planting season is
usually a prerequisite for crop establishment and
predetermined water amounts are applied uni-
formly to fields often without discerning salinity
levels. The use of predetermined water amounts
for leaching perhaps partly emanate from the in-
ability of conventional soil salinity surveys (based
on collection of soil samples, laboratory analy-
ses) to generate timely and high-resolution salinity
maps. The spatial distribution of soil salinity over
the landscape depends on many environmental
factors (Evans and Caccetta 2000; McKenzie and
Ryan 1999). While intensive soil sampling can be
eliminated or significantly reduced by an elec-
tromagnetic (EM) induction tool that measures

electrical conductivity of the soil (Lesch et al.
1995; Rhoades et al. 1999), there are few studies
on relationship between environmental variables
and EM values to estimate soil salinity in irrigated
areas.
Similar to many other successful attempts
(Lavado et al. 2006; Leij et al. 2004; Moore et al.
1993) to develop analogues of conventional sur-
veys by terrain attributes, this study builds on
readily available or easily extractable data to ac-
count for the spatial distribution of soil salinity.
The focus here was whether it would be possible
to develop models using the available existing
data sets regardless their intrinsic shortcomings. If
so, it would help decision making under the con-
ditions encountered by farmers or water resource
managers.
The salinity distribution was modeled using the
identified factors at the farm scale (Akramkhanov
2005) and this model was used to upscale the salin-
ity distribution to the district level. Soil salinity is
a dynamic soil property particularly in irrigated
areas of Khorezm region of Uzbekistan. Emphasis
was made to test the feasibility of the model to
estimate soil salinity during summer period, when
most irrigation takes place and soil salinity in the
area is subject to variation. The model developed
in such variable environment will be robust and
applicable for other time periods.
The complex nature of the environmental vari-
ables and assumptions in classic statistical analysis
often make the use of statistical theories difficult.
An alternative is the use of artificial neural
network (ANN), the computerized information
processing model, the development of which was
motivated by the way the brain works. It is able
to modify its structure by learning from the data
rather than being programmed, and is thus able to
solve complicated relations (Principe et al. 2000).
In salinity studies, ANN have been used for a
variety of purposes, including the prediction of a
salt build-up in the crop root zone (Patel et al.
2002), dielectric constant–soil water relationships
(Persson et al. 2002), and river water salinity fore-
casting (Bowden et al. 2002; Maier and Dandy
1999).
In studies with a spatial domain, Eklund et al.
(1998) applied a neural network algorithm for
inductive inference from GIS layers to clas-
Environ Monit Assess
sify soil salinity risk potential. Within the soil-
landscape dimension, Zhu (2000) showed that the
neural network approach reveals much greater
spatial detail and has higher quality than that
derived from the conventional soil map. Similarly,
McBratney et al. (2000) gave an example of clay
content estimation where neural networks per-
formed among the best. In all these studies, as well
as in the majority of other studies, feed-forward
networks have been used. Feed-forward neural
networks are composed of layers of neurons, in
which the input layer of neurons is connected to
the output layer of neurons. The training process
is undertaken by changing the weights such that a
desired input–output relationship is realized.
Multilayer perceptrons (MLPs) are layered
feed-forward networks typically trained with sta-
tic backpropagation (Principe et al. 2000). With
backpropagation, the input data is repeatedly pre-
sented to the neural network. With each presenta-
tion, the output of the neural network is compared
to the desired output and an error is computed.
This error then is fed back (backpropagated) to
the neural network and used to adjust the weights
such that the error decreases with each iteration
and the neural model gets closer and closer to pro-
ducing the desired output (this process is known
as “training”). Principe et al. (2000) emphasize
that their main advantage is in being easy to use,
and the key disadvantages are that they train
slowly and require a large amount of training
data (typically three times more training samples
than network weights). However, MLPs with a
sufficient number of hidden units can approximate
any continuous function to a pre-specified accu-
racy; in other words, MLP networks are universal
approximators (Cherkassky and Mulier 1998).
In this study, ANN is used as an alternative
to regression techniques for prediction of the soil
salinity distribution. More than 400 cases of field-
collected data are used for model development.
The overall research objective of the study was
to develop a model to estimate soil salinity distri-
bution at the farm and district level with an em-
phasis on using readily available and cheaply ob-
tainable measurements, so that real-time decision
making can be supported both on-farm (by farm-
ers) and off-farm (by policy makers) in Khorezm,
Uzbekistan.
Environ Monit Assess
Materials and methods
Area description
The data for training of the model at the farm
scale was collected in 2002 on the research farm
(41◦ 36 N, 60◦ 31 E) of the Urgench State Uni-
versity and the surrounding farm southwest of
the city of Khiva, in Khorezm, Uzbekistan. The
sampling area (ca. 15 km2 ) consists of sands, sandy
loams, and silt loams at equal proportions. Most
of the area is cultivated land. However, bare or
abandoned land was also included in the study
to possibly increase the variability in soil salinity.
More details of the farm-scale description can be
found in Akramkhanov (2005) and Akramkhanov
et al. (2008). Resulting trained neural network
model from farm scale in 2002 is applied to up-
scaling area in 2003.
The upscaling area is a district. Khiva district is
one of 11 districts of the Khorezm region and is
situated in the southwest of the region. The total
area of upscaling in this study is approximately
300 km2 (Fig. 1), but data is gathered from sur-
rounding areas as well to avoid edge effect dur-
ing interpolation analysis. A range of attributes
present at farm scale are also represented at the
Fig. 1 Location of study
area for model
development and district
area for upscaling
district scale, with the difference that rice is grown
in many areas of the district, whereas on the
study farm rice was not grown during the sampling
year.
The climate in Khorezm is arid with an annual
precipitation of about 100 mm ranging from 35 to
170 mm during dry and wet years, respectively.
About 70% of it falls in the winter and spring.
The topography of the district is similar to that of
the farm, i.e., flat, with elevation points normally
distributed ranging from 85 to 110 m above sea
level (mean 93 m and standard deviation 2.9). The
landscape is dissected by an extensive network
of drains and collectors (Fig. 1). The soils in the
district are classified by the local classification sys-
tem as desert zone meadow-oasis soils and by the
World Reference Base as Cambisols (Abdullaev
2003). At finer resolution some soils match
Solonchak description but depending on location,
once leached, fall into a complex of arenosols–
fluvisols–anthrosols descriptions. These soils have
moderately high humus content (0.8–1.4%) com-
pared to the rest of the soils in this desert zone.
Since the land in the area has been constantly
under development, many studies distinguish soils
by irrigation duration, basically referring to them
as soils of ancient or recent development.

Soil salinity and environmental variables
Soil salinity (CMv) was estimated by electromag-
netic induction device (CM-138) which measures
apparent electric conductivity (ECa ) of a soil
profile down to 1.5 m depth in vertical mode.
The following variables from selected environ-
mental factors were derived from varying sources
for the farm and district scale inputs. The soil
texture for the farm scale input was analyzed on
samples collected during a survey from top 30 cm.
We assume that topsoil is representative of the
1 m layer. The soil texture for the district level
was estimated using the dataset from 1996 soil sur-
vey provided by Soil Science Institute where the
average depth of the top horizons of 51 sampling
sites was 31 cm. For the ease of transferability of
the results in the local context, clay was assumed
to consist of soil separates with a size of less
than 0.001 mm and silt to be between 0.001 and
0.05 mm (Russian classification scheme).
Elevation data from a 1:10,000-topographic
map were obtained to create a digital elevation
model (DEM). Terrain indices were calculated
from a 30 × 30-m raster-based DEM. The fol-
lowing terrain indices were calculated using the
software DiGeM 2.0 (Olaf Conrad, Gottingen,
Germany): aspect (as), slope (sl), profile curva-
ture (profc), plan curvature (planc), curvature
(curv) (Zevenbergen and Thorne 1987); diver-
gence/convergence (dc) indices, flow accumula-
tion (upslope contributing area, ua), wetness (wt),
and erosivity (based on universal soil loss equa-
tion, ls) (Moore et al. 1993). Additionally, curva-
ture (curv7) and the terrain characterization in-
dices (tci) were calculated according to Park et al.
(2001). Elevation values in meters (alt) above mean
sea level and solar radiation (solar) were also used.
Agricultural maps (scale 1:10,000 and 1:25,000)
were used to obtain the drain water collector net-
work. The layers were digitized and the nearest dis-
tance from drain collector (dcoll) to sampling point
was obtained using ArcView 3.2 (ESRI Inc., USA).
Remote sensing parameters were obtained
from a Landsat 7 satellite image acquired on the
July 12, 2002. Three variables were selected, the
ratio vegetation index which is known to delineate
reduced reflectance due to salinity (Wang et al.
2002) and the raw bands 3 and 5.
Environ Monit Assess
Data on groundwater table depth and its salin-
ity were obtained from over 500 observation wells
installed in the study area by the Hydrological
Melioration Expedition of the Khorezm Depart-
ment of Land and Water Resources. The average
value of groundwater table depth and salinity in
groundwater observation wells collected in July
for the period from 1990 to 2002 were averaged
and used for interpolation and feeding into the
model. It was assumed that long-term average
data represent the spatial distribution of ground-
water table depth and salinity better than single-
year observations.
ANN simulation model characteristics
The PC-based software package NeuroSolutions
4.31 (NeuroDimension, Inc.) was used in this
study for the artificial neural network operation.
The package is integrated into a spreadsheet. The
steps outlined in Maier and Dandy (1999) and
Bowden et al. (2002) were used as a guide in this
work. The default values suggested in the software
were used unless stated otherwise.
The dataset of 441 cases was split into three
subsets, namely training, cross-validation, and
testing. The training subset is used to train the
network, the cross-validation subset is used to
avoid neural network overtraining (training is ter-
minated when mean squared error (MSE) of the
cross-validation subset increases), and the testing
subset is used to test the built model on unseen
data. The number of epochs (iterations) was set
to 10,000 and the network was trained using cross-
validation ten times. Bowden et al. (2002) showed
that data division into subsets is important for
the model performance, and three subsets should
represent the same population. For example, in-
dependent subsets were split provided that each
included the lowest and the highest values for each
variable. Out of 441 cases, the training subset con-
sisted of 265 cases, cross-validation 66 cases, and
the testing subset 110 cases, the detailed statistics
are given in Akramkhanov (2005). Some variables
could not be divided into equal subsets, which can
be attributed to singular or rare events present in
the dataset. However, in general, the statistics of
the subset are in good agreement and appear to
come from the same population.
Environ Monit Assess
Fig. 2 Relationship
between ECa measured
with electromagnetic
induction device CM-138
in vertical mode and ECe
of the soil for various soil
textures
A multi-layer perceptron with one hidden layer
with 12 processing elements was built with the
help of a wizard provided in the software. Non-
linearity was introduced by the hyperbolic tan-
gent function (TanhAxon), and the momentum
learning rule was used for updating the weights.
Predictions of the neural network were assessed
with the correlation coefficient R between mea-
sured and estimated values and mean squared
error. The best weights during these runs were
applied for the testing subset. The best weights
were then applied to the production (dataset to be
predicted) subset.
In addition, sensitivity analysis was performed
to identify those input variables that had more of
an effect on estimating soil salinity. It provides
a measure of the relative importance among the
inputs of the neural model and illustrates how the
model output varies in response to input variation.
The first input is varied between its mean ±1
standard deviations while all other inputs are fixed
at their respective means. The network output is
computed for 50 steps above and below the mean.
This process is repeated for each input. A report
Table 1 Neural network results of training; estimates between measured and estimated soil salinity values (dS m−1 )
Subset Measured
CMv mean
Training 0.51
Cross-validation 0.46
Testing 0.51
is generated that summarizes the variation of each
output with respect to the variation in each input.
Validation survey of the district scale map
The study area was represented by 1,755 points
randomly generated by the Arcview sample gen-
erator extension (DNR 2004). This number of
points was arbitrarily chosen to cover the whole
area of the district, and it was expected to provide
a sufficient number of points. The trained model
was applied to these 1,755 points with the same
attributes used in training but derived for the
district area. The resulting map from predicted
values of soil salinity was interpolated using the
inverse distance weighting in ArcMap.
The validation of the generated salinity map
was done by ground-truthing at 315 randomly
selected locations. Sampling sites were randomly
selected but with a view to logistics and accessi-
bility and were required to cover the whole area
of the delimited district area. Their coordinates
were uploaded into a handheld GPS in advance
and navigation by GPS was used to locate the
Estimated Correlation MSE
CMv mean coefficient, R
0.51 0.82 0.04
0.51 0.85 0.03
0.53 0.83 0.03
 Environ Monit Assess

1.8 1.4 1.4

Training subset Cross-validation subset Testing subset
1.6 1.2 1.2
predicted CMv, dS m-1

predicted CMv, dS m-1

predicted CMv, dS m-1

1.4
1 1
1.2
1 0.8 0.8
0.8 0.6 0.6
0.6
0.4 0.4
0.4
0.2 0.2 0.2
0 0 0
0 0.5 1 1.5 2 2.5 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
measured CMv, dS m-1 measured CMv, dS m-1 measured CMv, dS m-1

Fig. 3 Scatter plots between measured and estimated CMv for different subsets in the neural network

sampling points on the ground. The timeframe and low mean square error between measured
for the validation survey was scheduled to be in and estimated soil salinity values. MSE values are
summer, running for about 4–5 weeks and start- comparable between subsets, which indicate that
ing in July. Vertical electromagnetic conductivity the division into subsets yielded relatively uniform
measurement (CMv, dS m−1 ) was used in the samples in these subsets.
analysis. Sensitivity analysis showed that the response
To compare estimated and measured CMv, we of the model was highly dependent on the fol-
used relationship presented in Fig. 2. For the lowing terrain variables, listed in decreasing order
study area the conversion for loamy texture of of importance: curvature (curv), plan curvature
Rhoades et al. (1999) and Bennett et al. (1995) (planc), profile curvature (profc), and solar radi-
is appropriate, where CMv of 0.6, 1.0, 1.5, and ation (solar). It appears that soil salinity is largely
2.3 dS m−1 roughly correspond to ECe of 4, 6, 8, influenced by the micro-topography, convexity, or
and 12 dS m−1 , respectively. Soils with ECe above concavity, which tends to influence surface water
4 dS m−1 are considered saline. Since soils in the retention.
study area are a mix of light and medium texture,
transition range of critical salinity CMv values
above which soils are considered saline could be Validation results
within 0.6–0.9 dS m−1 . Therefore, we classify CMv
values into saline and not saline levels in esti- The statistics of the estimated and measured
mated and measured datasets using three criti- soil salinity datasets are summarized in Table 2.
cal salinity threshold values such as 0.7, 0.8, and The mean values for both, estimated and mea-
0.9 dS m−1 . For each critical salinity threshold we sured sets, are close to each other (0.94 and
compare estimated with measured CMv classified 1.04 dS m−1 ); MSE and RMSE are 0.05 and 0.25,
as saline (i.e., >0.7 dS m−1 ) and not saline (i.e., respectively. Despite the fact that the average
<0.7 dS m−1 ) classes. We calculate precision of
correctly classified saline values at each selected
critical salinity value.
Table 2 Statistics of trained, measured, and estimated val-
ues of soil salinity (dS m−1 )

Results Trained Estimated Measured

Mean 0.51 0.94 1.04
Neural network model training performance Median 0.51 0.93 1.02
Mode 0.20 1.86 0.96
SD 0.33 0.11 0.19
Statistics of the performance of the neural net- Skewness 0.79 2.21 0.67
work model are summarized in Table 1. The scat- Minimum 0.03 0.41 0.59
ter plots of the measured and estimated salinity Maximum 1.92 1.86 1.78
values for each subset are shown in Fig. 3. The Percentiles 25 0.20 0.88 0.91
results indicate that the model performed well 50 0.51 0.93 1.02
taking into account high correlation coefficients 75 0.73 0.99 1.14
Table 3 Comparison table of estimated and measured salinity maps using three critical threshold values 0.7, 0.8, and 0.9 dS m−1
Threshold Measured Predicted
Environ Monit Assess

0.7 dS m−1 White – not saline Dark – saline

0.8 dS m−1

0.9 dS m−1

Fig. 4 Precision of classifying saline soils of estimated
maps based on critical salinity threshold values. Threshold
values of 0.7, 0.8, and 0.9 discussed in text are marked
value of the training subset (0.51 dS m−1 ) was
considerably lower than that of the estimated set
(0.94 dS m−1 ), the model performed well, i.e., it
predicted the average salinity for the district close
to the average value of the measured set.
According to Rhoades et al. (1999) and Bennett
et al. (1995) average CMv of estimated and mea-
sured datasets indicate saline soils in the study
area. Spatial distribution of saline and not saline
soils in estimated and measured maps is presented
in Table 3 using three critical salinity threshold
values. Light color indicates not saline soils, dark
color—saline soils. Change in precision using crit-
ical salinity threshold values is shown in Fig. 4.
If we assume that critical salinity threshold
above which soils are considered saline to be equal
to 0.7 dS m−1 then the precision of estimating
saline soils is close to 90%. High precision can also
be seen from two maps in Fig. 4, where both es-
timated and measured maps are uniformly saline
throughout.
With the increase in the critical salinity thresh-
old to 0.8 and 0.9 dS m−1 precision is reduced
to around 80% and 70%, respectively. Estimated
maps have more areas classified as saline com-
pared to measured dataset.
Discussion
The scatter plots (Fig. 3) of all three subsets
used in training of the model appear to have
Environ Monit Assess
one feature: deviation in the lower and higher
range between measured and estimated values.
The neural network seems to overestimate values
in the low salinity range and to underestimate
those in the higher range. Underestimations of soil
salinity, especially in the vicinity of peak values,
could be caused by the lack of samples with high
values in the training set for the model to learn.
The modeling procedure has to deal with several
issues, namely change of scales and the complex
relationship among the variables involved.
Data fed into the model come from a vari-
ety of sources and often represent the proxies
of the true covariates obtained by interpolation.
Interpolation has been extensively studied, and
errors resulting from interpolation itself can be
quite substantial. For example, the RMSE for
different kriging methods of soil salinity examined
by Triantafilis et al. (2001) varied from 1 to 1.3
in conditions very similar to this study area. Also,
interpolation is known to smooth the surface and
the resulting dataset represents less the low and
high values. Perhaps the lack of such low and high
values in the number of environmental covari-
ates included in the model training contributed to
over- or under-estimation of soil salinity.
The strong influence of the terrain attributes
indicated by the sensitivity analysis is encour-
aging, because it supports the concept of the
soil–landscape relationship in this study. It ap-
pears that soil salinity variability is mainly due to
the spatial variability of curvatures (curv, planc,
profc), which tend to control water movement
and solar radiation. The neural network was more
sensitive to these micro-scale controlling factors
than to those that are obvious at a macro-scale.
Kachanoski et al. (1985) observed that A-horizon
variability was due to spatial variability of surface
curvature, but the relationship was complicated
due to the presence of positive and negative cor-
relations over different frequency ranges.
The explanation of spatial distribution of soil
salinity by terrain information, i.e., shape, is at-
tractive for empirical model building, because it
does not require detailed soil surveys over the
landscape. The major drawback of such models,
however, is that they are scale specific or range
(of values it had been trained on) specific, which
is also noticeable in the scatter plots (Fig. 3).
Environ Monit Assess
The difference between average salinity val-
ues of training and measured datasets stands out.
Accurate prediction of the average soil salinity
perhaps was due to the presence of the lowest and
highest salinity ranges in the training subset dur-
ing the model training. This ability of the neural
network to model complicated functions demon-
strates that they are excellent approximators.
The use of environmental attributes and the
soil salinity relationship to upscale spatial distri-
bution of soil salinity from farm to district scale
essentially estimated similar mean soil salinity val-
ues (0.94 vs. 1.04 dS m−1 ). The neural network
model used to spatially estimate soil salinity val-
ues at selected points was able to approximate soil
salinity despite the possible errors in intermediate
processing of data. This shows that the model
based on the included variables can effectively
estimate soil salinity. The major advantage of this
approach is that available data is used.
Since soil texture change is transitional rather
than abrupt the application of one critical salinity
threshold value might not be appropriate. There-
fore, the precision of the estimated map in identi-
fying saline soils is around 70–90% depending on
the critical salinity threshold value used. From the
experience in the area and other surveys with EM
devices threshold values of 0.8 and 0.9 dS m−1 will
certainly indicate saline soils, although with lower
precision.
As technology and models are developing
rapidly, many methodologies emphasize the
flexibility and adeptness to modifications. The
model can be fine-tuned by incorporation of other
related factors, for instance, geodesic data such as
the extent of impermeable layers. By incorporat-
ing new correlating factors, further complication
of the model should be avoided. By fine-tuning a
“window of opportunity” (Bouma and Hoosbeek
1996) is rather created for improvement of the
model by replacing with either a better explaining
factor or a more easily obtainable factor.
Hoosbeek and Bouma (1998) emphasize that
to be able to define suitable indicators for certain
soil and land qualities, it is important to have
insight into the characteristics of an indicator. As
previously attempted by many others, the Jenny
model was simplified by the assumption that other
parameters are constant and was based largely on
terrain and texture, and on additional parameters
from remote sensing and drainage networks. The
emphasis was on terrain attributes because, as
Moore et al. (1993) and McKenzie et al. (2000)
proved, these provide a much better basis for
visualizing other data layers and allow generating
sets of data that have value in their own right (e.g.,
slope and solar radiation), even when the exact
form of the relationship is unknown.
This study demonstrates that the ANN ap-
proach is suitable for upscaling purposes, and re-
sults are valid within a given range of soil salinity.
Timing of measurement is important, and as the
study was conducted during the summer period,
further application of this model to other dis-
tricts of the Khorezm region and validation should
involve the summer period only. Application of
the results to other seasons must be carefully as-
sessed before making any generalizations about
soil salinity distribution using the model presented
here. For this, conducting field survey and retrain-
ing model with dataset obtained in, for example,
autumn.
The influence of soil and irrigation manage-
ment on soil salinity is a dynamic feature and
should be included in models of soil salinity dis-
tribution. Capturing and parameterizing this dy-
namic feature into a model could potentially im-
prove model performance. Therefore, further re-
search is needed to quantify these parameters,
which can then be included in the further devel-
opment of such models.
Conclusions
The environmental correlation model built for the
farm-scale estimated soil salinity using a neural
network with a correlation coefficient between es-
timated and measured soil salinity of 0.83. Explicit
prediction of salinity was satisfactory taking into
account contrasting scales of soil salinity varia-
tion and environmental data derived from varying
scales and unknown accuracy.
The use of environmental attributes and soil
salinity relationship to upscale spatial distribution
of soil salinity from farm to district scale essen-
tially estimated identical mean soil salinity values
(0.94 vs. 1.04 dS m−1 ). Precision of estimation

of spatial distribution of saline soils was in the
range of 70–90% depending on the critical salinity
threshold applied.
In general, the results indicate that significant
correlations between quantifiable terrain at-
tributes and soil salinity exist. This study shows
that such a relationship was successfully used to
estimate soil salinity at farm scale solely based
on quantified environmental variables. Therefore,
the environmental variables must be available for
the study area in high resolution or easily measur-
able.
Acknowledgement This study was funded by the German
Ministry for Education and Research (BMBF; project
number 0339970A).
References
Abdullaev, U. (2003). Republic of Uzbekistan. Land degra-
dation assessment in drylands (LADA). Research and
Design Institute Uzgipromeliovodkhoz, pp. 43.
Akramkhanov, A. (2005). The spatial distribution of soil
salinity: Detection and prediction. ZEF Series in Ecol-
ogy and Development No. 32, Bonn, Germany, Uni-
versity of Bonn.
Akramkhanov, A., Sommer, R., Martius, C., Hendrickx,
J. M. H., & Vlek, P. L. G. (2008). Comparison and sen-
sitivity of measurement techniques for spatial distrib-
ution of soil salinity. Irrigation and Drainage Systems,
22, 115–126.
Bennett, D. L., George, R. J., & Ryder, A. (1995). Soil
salinity assessment using the EM38: Field operating
instructions and data interpretation. Miscellaneous
Publication 4/95, Western Australia, Department of
Agriculture.
Bouma, J., & Hoosbeek, M. R. (1996). The contribution
and importance of soil scientists in interdisciplinary
studies dealing with land. In R. J. Wagenet, & J.
Bouma (Eds.), The role of soil science in interdscipli-
nary research (pp. 1–15). SSSA Special Publication 45.
Bowden, G. J., Maier, H. R., & Dandy, G. C. (2002). Opti-
mal division of data for neural network models in wa-
ter resources applications. Water Resources Research,
38, 1–11.
Cherkassky, V., & Mulier, F. (1998). Learning from data:
Concepts, theory, and methods. U.S.A.: Wiley.
Department for Natural Resources (2004). DNR random
sample generator. Release 2.2. Minnesota: Department
for Natural Resources.
Eklund, P. W., Kirkby, S. D., & Salim, A. (1998). Data
mining and soil salinity analysis. International Journal
of Geographical Information Science, 12, 247–268.
Environ Monit Assess
Evans, F. H., & Caccetta, P. A. (2000). Broad-scale spatial
prediction of areas at risk from dryland salinity. Car-
tography, 29, 33–40.
Hoosbeek, M. R., & Bouma, J. (1998). Obtaining soil and
land quality indicators using research chains and geo-
statistical methods. Nutrient Cycling in Agroecosys-
tems, 50, 35–50.
Kachanoski, R. G., Rolston, D. E., & de Jong, E. (1985).
Spatial and spectral relationships of soil properties
and microtopography: I. Density and thickness of A
horizon. Soil Science Society of America Journal, 49,
804–812.
Lavado, F., Maneta, M., & Schnabel, S. (2006). Prediction
of near-surface soil moisture at large scale by digital
terrain modelling and neural networks. Environmental
Monitoring and Assessment, 121(1–3), 211–230.
Leij, F. J., Romano, N., Palladino, M., Schaap, M. G., &
Coppola, A. (2004). Topographic attributes to pre-
dict soil hydraulic properties along a hillslope transect.
Water Resources Research, 40, W02407.
Lesch, S. M., Strauss, D. J., & Rhoades, J. D. (1995).
Spatial prediction of soil salinity using electromag-
netic induction techniques. 2. An efficient spatial sam-
pling algorithm suitable for multiple linear regression
model identification and estimation. Water Resources
Research, 31, 387–398.
Maier, H. R., & Dandy, G. C. (1999). Empirical com-
parison of various methods for training feed-forward
neural networks for salinity forecasting. Water Re-
sources Research, 35, 2591–2596.
McBratney, A. B., Odeh, I. O. A., Bishop, T. F. A.,
Dunbar, M. S., & Shatar, T. M. (2000). An overview
of pedometric techniques for use in soil survey. Geo-
derma, 97, 293–327.
McKenzie, N. J., & Ryan, P. J. (1999). Spatial prediction of
soil properties using environmental correlation. Geo-
derma, 89, 67–94.
McKenzie, N. J., Gessler, P. E., Ryan, P. J., & O’Connell,
D. A. (2000). The role of terrain analysis in soil map-
ping. In J. P. Wilson, & J. C. Gallant (Eds.), Terrain
analysis: Principles and applications (p. 245–265). New
York: Wiley.
Moore, I. D., Gessler, P. E., Nielsen, G. A., & Peterson,
G. A. (1993). Soil attribute prediction using terrain
analysis. Soil Science Society of America Journal, 57,
443–452.
Park, S. J., McSweeney, K., & Lowery, B. (2001). Iden-
tification of the spatial distribution of soils using
a process-based terrain characterization. Geoderma,
103, 249–272.
Patel, R. M., Prasher, S. O., Goel, P. K., & Bassi, R. (2002).
Soil salinity prediction using artificial neural networks.
Journal of the American Water Resources Association,
38, 91–100.
Persson, M., Sivakumar, B., Berndtsson, R., Jacobsen,
O. H., & Schjonning, P. (2002). Predicting the dielec-
tric constant–water content relationship using artificial
neural networks. Soil Science Society of America Jour-
nal, 66, 1424–1429.
Environ Monit Assess
Principe, J. C., Euliano, N. R., & Lefebvre, C. W. (2000).
Neural and adaptive systems: Fundamentals through
simulations. New York: Wiley.
Rhoades, J. D., Chanduvi, F., & Lesch, S. (1999). Soil salin-
ity assessment: Methods and interpretation of electri-
cal conductivity measurements. FAO Irrigation and
Drainage Paper 57.
Triantafilis, J., Odeh, I. O. A., & McBratney, A. B. (2001).
Five geostatistical models to predict soil salinity from
electromagnetic induction data across irrigated cotton.
Soil Science Society of America Journal, 65, 869–878.
View publication stats
Wang, D., Wilson, C., & Shannon, M. C. (2002). Inter-
pretation of salinity and irrigation effects on soybean
canopy reflectance in visible and near-infrared spec-
trum domain. International Journal of Remote Sensing,
23, 811–824.
Zevenbergen, L. W., & Thorne, C. R. (1987). Quantita-
tive analysis of land surface topography. Earth Surface
Processes and Landforms, 12, 47–56.
Zhu, A. X. (2000). Mapping soil landscape as spatial con-
tinua: The neural network approach. Water Resources
Research, 36, 663–677.