You are on page 1of 10

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION


VOL. 38, NO. 1 AMERICAN WATER RESOURCES ASSOCIATION FEBRUARY 2002

SOIL SALINITY PREDICTION USING ARTIFICIAL NEURAL NETWORKS'

R. M. Patel, S. 0. Prasher, P K Goel, and R. Bassi2

ABSTRACT: This study explores the applicability of Artificial Neu- If the decision fails to yield the desired result, a
ral Networks (ANNs) for predicting salt build-up in the crop root second alternative is considered. This process contin-
zone. ANN models were developed with salinity data from field
lysimeters subirrigated with brackish water. Different ANN archi- ues until a satisfactory result is achieved. The final
tectures were explored by varying the number of processing ele- decision and the governing situation are added to the
ments (PEs) (from 1 to 30) for replicate data from a 0.4 m water knowledge base for future use, and if necessary, earli-
table, 0.8 m water table, and both 0.4 and 0.8 m water table lysime- er data is either modified or deleted to accommodate
ters. Different ANN models were developed by using individual the new examples. The processing elements (PEs), or
replicate treatment values as well as the mean value for each treat-
ment. For replicate data, the models with twenty, seven, and six computational units, are arranged in different layers,
PEs were found to be the best for the water tables at 0.4 m, 0.8 m and can combine one or more inputs and modify them
and both water tables combined, respectively. The correlation coeffi- according to a transfer function to produce one or
cients between observed salinity and ANN predicted salinity of the more outputs.
test data with these models were 0.89, 0.91, and 0.89, respectively. ANN models have been effectively employed in
The performance of the ANNs developed using mean salinity values
of the replicates was found to be similar to those with replicate many areas, ranging from economics to science. ANN
data. Not only was there agreement between observed and ANN models have considerable potential for modeling some
predicted salinity values, the results clearly indicated the potential aspects of agroecosystems, which are often complex,
use of ANN models for predicting salt build-up in soil profile at a non-linear, and ill defined. The use of ANN technology
specific site. has shown promising results in dealing with several
(KEY TERMS: artificial neural networks; soil salinity; subirriga-
tion; green peppers.) types of agricultural problems. The first type of appli-
cation is machine vision and image analysis for the
identification and classification of agricultural prod-
ucts, such as weed recognition, apple and wheat ker-
INTRODUCTION nel color classification, potato inspection, and
pistachio grading (Deck et al., 1995; Ghazanfari et
al., 1996; Nakano, 1997; Wang et al., 1999; Yang et
Artificial neural networks (ANNs) are computa- al., 1999). Secondly, ANNs have been used in decision
tional algorithms. Their architecture and functions support systems for management and process control,
are derived from the structure of biological neural e.g. management of water pollution control,
networks and the theory of adaptive learning. Biologi- greenhouse environment control, agricultural input
cal neural networks (such as human brains) respond recommendation, and resource management (Kok et
to a situation through highly complex processes based al., 1994; Gimblett and Ball, 1995; Broner and Corn-
on learning from examples. New conditions are stock, 1997; Seginer, 1997; Wen and Lee, 1998).
interpreted in terms of lessons learned from the Thirdly, ANNs have been used for simulating and pre-
examples: a decision is selected from the available dicting various agricultural and environmental
options, and then the expected outcome is evaluated. problems (Chakraborty et al., 1992; Hsu et al., 1995;

'Paper No. 00136 of the Journal of the American Water Resources Association. Discussions are open until October 1, 2002.
'Respectively, Postdoctoral Fellow, Professor, Graduate Student, and Former Graduate Student, Agricultural and Biosystems Engineering,
Macdonald Campus of McGill University, 21111 Lakeshore Road, Ste-Anne de Bellevue, Quebec, Canada H9X 3V9 (E-MaillPrasher: shiv.
prasher@mcgill.ca).

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 91 JAWRA


Patel, Prasher, Goel, and Bassi

Jayawardena and Fernando, 1995; Batchelor et al., (2000) developed ANN models for salinity prediction
1997; Clair and Ehrman, 1998). in lysimeters. The lysimeters were packed with a
The potential utility of ANNs for solving complex sandy soil. The soil was salinized before planting.
hydrological processes is still being explored. A num- Potatoes (Solanum tuberosum L.) were seeded in this
ber of approaches are being investigated for forecast- saline soil. The potatoes were subirrigated with
ing events and modeling the movement of water in brackish water, and the salinity at various depths in
soil systems. Information about the hydrological the root zone was recorded throughout the season. In
behavior of such systems is acquired by conducting general, the results from Patel et al. (2000) were quite
field-level surveys and experiments under specific promising, although in certain instances, the models
conditions. However, the processes involved in such either overpredicted or underpredicted salinity. How-
systems are very complex, and it is quite difficult to ever, it is not very clear which are the scenarios
describe them explicitly in mathematical terms. The where the ANN overpredicts or underpredicts salinity.
process-based models that have been developed The ANN could model salt build-up in saline soil.
require many input parameter values, for some of However, salt build-up patterns in non-saline soil
which it is difficult to determine the limits to their subirrigated with brackish water could be different
applicability. ANN models have proven to be very from saline soil, and an investigation of the capability
effective in similar situations in other areas of investi- of the ANNs is needed. Therefore, this study was
gation. Maier and Dandy (1996), for example, high- directed to evaluate the potential of ANN to simulate
lighted the usefulness of ANN models and presented salt build-up in nonsaline lysimeters, subirrigated
a case study for forecasting salinity in a river. They with brackish water, and planted with green peppers
reviewed the difference between ANNs and tradition- (Capsicum annum Grossum). There was an attempt
al (statistical and physically-based models) forecast- to check whether the model simulations were satisfac-
ing methods. Pachepsky et al. (1996) compared the tory in different scenarios of varying water tables and
accuracy of ANNs and statistical regression methods irrigation water salinity. A number of different ANN
for the estimation of water retention in soils based on architectures were tried, and the model effectiveness
soil texture and bulk density data and found ANNs to was evaluated on the basis of the Root Mean Square
be superior. Shukia et al. (1996) explored the possibil- (RMS) error, a parameter most commonly used to
ity of using ANNs for simulating water table depth evaluate the training efficiency of ANN models. Final-
for transient drainage design. They obtained very ly, simulated salinity values were compared with the
encouraging results and concluded that ANNs might observed values, and various statistics, such as arith-
be applied to the real time control of drainage sys- metic mean, absolute arithmetic mean, standard devi-
tems, in light of their reliability and high speed in ation, and correlation coefficient, for testing and
simulating water table conditions, as compared to tra- training data were determined for the evaluation of
ditional mathematical models. Similarly, Yang et al. the developed models.
(1995) successfully related DRAINMOD outputs with
ANN.
Water and solute flow through soil is highly com-
plex. Therefore, limited success has been achieved MATERIALS AND METHODS
in mathematical modeling of the involved physio-
chemical processes. Study of these processes is fur- Salt Build-up Data for ANN Models
ther complicated under growing crop conditions.
Therefore, the prediction of salt build-up in the soil Data for the development of ANN models was
profile under such conditions is more difficult. Very obtained from a field lysimeter experiment. Green
few studies have been reported on the mathematical
modeling of salt transport in the soil profile, and field pepper (Capsicum annum Grossum, 'Bellboy') was
grown in field lysimeters that were one meter high
validation of these models has been limited. Comfort and 0.45 m in diameter (Patel, 1997). Brackish water,
et al. (1993) observed that models tested with field with salinity levels of 1, 5, and 9 dS/m, was applied
data have generated salinity values that vary consid- with a subirrigation system. Two water table depths
erably from the observed values. ANNs do not need to (0.4 and 0.8 m from the soil surface) were maintained.
know the exact form of the mathematical equations as
In addition, four levels of fertilizers were applied.
they learn the phenomena from the observed data. Thus, three levels of irrigation water salinity, two
Moreover, the spatial and temporal variability of vari- water table depths, and four levels of fertilizers,
ous physical and chemical parameters makes such resulted in a total of 24 treatments. With three repli-
problems even more difficult to model. In such cases,
cates per treatment, there were 72 lysimeters used in
ANN models seem to have good potential. Patel et al.
this study.

JAWRA 92 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION


Soil Salinity Prediction using Artificial Neural Networks

In the lysimeters with 0.4 m water table depth, soil In the experiment, Patel (1997) reported different
solution salinity was recorded at 0.1 and 0.3 m from salt build-up trends under different water table treat-
the soil surface. In the lysimeters with 0.8 m water ments. Irrigation water salinity had a greater effect
table depth, the salinity was measured at 0.1, 0.3, 0.5, near the water table and less of an impact on salt
and 0.7 m. The salinity measurements were made 20, build-up near the soil surface. Initially, different irri-
29, 38, 43, 50, 57, 64, and 72 days after planting the gation water salinities had little impact on soil salini-
green peppers. Patel (1997) reported that in this ty, but at later stages, higher irrigation water salinity
experiment, fertilizer levels did not influence salt resulted in greater salt build-up in the soil column.
build-up in the soil. Therefore, the data for the fertil- Furthermore, irrigation water salinity had a signifi-
izer levels was not considered as an input parameter, cant impact on salt build-up at the 0.1 and 0.3 m
and the observations with different fertilizers were depths in lysimeters with a 0.4 m water table. Salt
also considered as replicates in this study. build-up in the lysimeters with a 0.8 m water table
Earlier work (Patel, 1997) had also shown that salt was not dependent on the irrigation water salinity at
build-up in soil profile depends on water table depth, the 0.1 and 0.3 m depths. However, the salt build-up
subirrigation water salinity, depth of salinity mea- at 0.5 and 0.7 m depths was found to be dependent on
surement, and the number of days elapsed after irrigation water salinity in the latter case. Therefore,
planting. Therefore, these independent variables separate ANNs were developed for both the 0.4 and
(water table depth, subirrigation water salinity, depth 0.8 m water tables. To develop a more general model,
of salinity measurement, and days elapsed after suitable for both water tables, an ANN was also
planting), and the dependent variable (soil salinity) trained using the data from both water table treat-
were treated as input and output parameters, respec- ments.
tively. Salt build-up in soil is also governed by many Although the models were developed using the
other agroclimatic parameters and physical soil prop- replicate data from 0.4 m and 0.8 m water table treat-
erties. For example, rooting depth is an important ments separately as well as both treatments com-
parameter influencing salt buildup. In the proposed bined, upon close scrutiny of the experimental data, it
ANN modeling, it has been considered in an implicit was found that there was considerable variation in
way due to choice of our input parameter, water table salinity among the replicates. Therefore, in the sec-
depth. The ANN model used only four inputs, in com- ond approach, instead of using replicate data, the
parison to other modeling approaches, such as mean values of all replicates were used to develop the
DRAINMOD (Workman et el., 1994) and SWAP 1993 ANN model.
(Van den Broek et el., 1994), which require extensive
input parameters. However, whereas the ANN model
is only applicable to the given soil and for the given Development of ANN Models
agroclimatic conditions, the advantage of the procedu-
ral mathematical models, like DRAINMOD and Neuraiworks Neural Netware Professional IT/PLUS
SWAP 1993, is that they can be used for any soil and (Neuralware, 1993) was used for developing the ANN
for nearly any geographic location. Thus, although models used in this study. All networks were of the
ANN models are applicable to the conditions of the feed-forward type. The standard back-propagation
experiments where the data is used in their develop- learning algorithm was selected because of its utility
ment, it is, of course, possible to develop a more com- in modeling nonlinear processes (Lacroix et al., 1997).
prehensive ANN model by incorporating a greater The most widely used normalized-cumulative d.elta
number of soil and meteorological input parameters, learning rule and the Sigmoidal transfer function
provided such data is available. were selected. To test the model at different stages of
The randomized data were divided into two sets: training, the "Best Save" function of the Neuralworks
one for training the ANNs and the other for testing. software was used. This function retains the network
Of the total, 80 percent of the data was used to make which gives the best result until the training process
a training data file. The remaining 20 percent of the results in a better network. The default values were
data was used for testing the models during the train- used for all other parameters. The Root Mean Square
ing process. The developed model was then used to (RMS) error method was used to determine the train-
predict the salinity values of testing data. It is diffi- ing effectiveness of the ANN model. This software
cult to evaluate the performance of the ANN in vari- first scales the supplied data into a certain range,
ous subirrigation scenarios using limited numbers of computes the outputs, calculates the RMS error, and
testing data. Therefore, salinity values were also pre- then transforms the scaled output values back to the
dicted for training data to generate a complete original scale.
dataset.

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 93 JAWRA


Patel, Prasher, Goel, and Bassi

Different ANN architectures, with a single hidden RESULTS AND DISCUSSION


layer with different numbers of PEs (from 1 to 30)
were considered. During the training phase, the ANN Selection of the ANN Model with Replicate Data
was provided with a training data file, containing all
for a Specific Water Table Depth
input values as well as the desired output values. The
ANN was trained for a maximum of 200,000 training
cycles. The minimum RMS error value for each model ANN models were developed for specific water
was recorded. The model with the minimum RMS table depths as well as for a generalized water table
error value was then used to predict the salinity for condition. The best model for each condition was fur-
all input data. ther evaluated to compare the observed and simulat-
In the second approach, an attempt was made to ed soil salinities. As mentioned earlier, the RMS error
develop ANN models with mean observed salinity val- was used in the selection of the model. The RMS error
ues rather than using individual replicate salinity values recorded for different ANN models are given in
values. By doing so, the number of records was Figure 1. In the case of ANNs for the 0.4 m water
reduced from 576 to 48 for the 0.4 m water table table depth with replicate data, the RMS error
treatment, and from 1152 to 96 for the 0.8 m water values varied from 0.0674 to 0.0602. Although there
table treatment. Thus, for both water table treat- were local variations based on number of PEs in the
ments combined, the number of records was reduced RMS error values, in general, the RMS error
from 1728 to 144. decreased with increasing number of PEs, up to 20

0.10

Replicate data
0.08
0.4 m WI
0.8mWT
BothWT
0
Mean data
0.06 •..• O.4mWT
a .
."- •---• 0.8mWT
£ b
O.4 •O •-•a•- BothWT

£ . .. £ A
a

0.04 ..
.p 0

.0

0.02
0 5 10 15 20 25 30
Number of Processing Elements

Figure 1. RMS Error Values for Different ANN Models Using Replicate and Mean Data With
Water Tables at Either 0.4 m or 0.8 m, and the Combined Data for Both Water Tables.

JAWRA 94 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION


Soil Salinity Prediction using Artificial Neural Networks

PEs. Additional PEs did not reduce the error. Thus, (difference between observed and predicted salinity)
for the 0.4 m water table depth, the minimum RMS were 0.0024 and 0.0539 dS/m for training and testing
error of 0.0602 was observed for the model with 20 data, respectively. For the 0.8 m water table, these
PEs in the hidden layer. The RMS error for the ANN errors were 0.0112 and 0.0298 dS/m. Thus, these
models developed for the 0.8 m water table depth errors were also reasonably low. Similarly, the corre-
ranged between 0.0766 and 0.0576 (Figure 1). The sponding absolute arithmetic mean errors for the 0.4
minimum RMS error (0.0576) was obtained for a m water table data were 0.4725 and 0.4514 dS/m, and
model with seven PEs in the hidden layer. These for the 0.8 m water table were 0.4935 and 0.4850
"best" models were used to predict the salinity in the dS/m. In light of the observed salinity range (0.52 to
soil profile. The observed and predicted salinity val- 7.21 dS/m), and considering the inherent variability
ues for 0.4 m water table are shown in Figure 2, and in the field, these errors are rather small.
the same for the 0.8 m water table are shown in Fig-
ure 3. A very good agreement between the predicted
and observed salinity values was observed for both General ANN model with Replicate Data
water tables, as evident from Figures 2 and 3. In the for Any Water Table Depth
case of the 0.4 m water table, the correlation coeffi-
cients were 0.86 and 0.89 for training and testing In the previous step, different ANN models were
data, respectively (Table 1). The corresponding corre- developed for each water table depth and were found
lation coefficients for the 0.8 m water table were 0.88 to do a satisfactory job in simulating soil salinities at
and 0.91 m (Table 1). It is clear from Figures 2 and 3 different depths. Next, efforts were made to develop a
that the simulation of the testing data was similar to generalized ANN model that would work for e:ither
that of the training data, which could be confirmed water table depth. The RMS error varied in this case
with the similarity of correlation coefficients of train- from 0.0736 to 0.0591 for different number of PEs
ing and testing data. Some of the other statistics of (Figure 1). A gradual decrease in RMS error values
the "best" models are presented in Table 1. In the case was observed with increasing numbers of PEs, up to
of the 0.4 m water table, the arithmetic mean errors

6
E -0
C,)
o o 00 0 •...
0 •a
0 •.OC [Testing
In
• d..d b... •• Training

0- •0• .0 • S.
'1)
4-'
C.)
•.—S 0
• Ito • .•o.
•• . • • . 00 0 •.
2 • •.....
• •-

• •

•.
0

• 0.0 0..
oPu.. ••0
0 I

0 2 4 6 8
Observed salinity dS/m

Figure 2. Observed and ANN Predicted Salinity in 0.4 m Water Table.

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 95 JAWRA


Patel, Prasher, Goel, and Bassi

0 OS
0 S 0
6
(/.)

0 • • . . 0••. 0 • 00
.00 . •. 0
0 Testing
-4. 0 •.0. .
0
• 0V00
0
• Training
C)
— ..QUb. %I

..._q..
0
• 0. •.p.10
— ••. DO
%d
o.
•00g • .1

2 ,t. . 12•. °'• •

.dsa

0
0 2 4 6 8
Observed salinity dSIm

Figure 3. Observed and ANN Predicted Salinity in 0.8 m Water Table.

TABLE 1. Statistics for Results from Best Models in Different Categories.

Replicate Data Water Tables Mean Data Water Tables


Data 0.4 m 0.8 m Combined 0.4 m 0.8 m Combined
PEs 20 7 6 24 17 14
RMS Error 0.0602 0.0576 0.0591 0.0218 0.0493 0.0401
Arithmetic Mean Error Training 0.0024 0.0112 0.0031 0.0080 0.0051 0.0035
Testing 0.0539 0.0298 0.0202 -0.0411 0.0527 -0.0228
Absolute Arithmetic Mean Error Training 0.4725 0.4935 0.5072 0. 1881 0.3325 0.3439
Testing 0.4514 0.4850 0.4943 0.1581 0.3974 0.3042
Standard Deviation Error Training 0.6802 0.6645 0.6937 0.2560 0.4124 0.4252
Testing 0.6291 0.6414 0.6600 0.1805 0.4983 0.4059
Correlation Coefficient Training 0.8597 0.8812 0.8679 0.9806 0.9508 0.9421
Testing 0.8875 0.9134 0.8940 0.9912 0.9430 0.9636
Total 0.8689 0.8888 0.8738 0.9805 0.9490 0.9472

NOTE: RMS error, arithmetic mean error, absolute arithmetic mean error and standard deviation are in dS/m.

six PEs (Figure 1). The addition of more PEs in the model, the salinity values were predicted for training
hidden layer did not result in any further reduction of and testing data. The predicted salinity values from
RMS error values. Thus, the ANN model with 6 PEs the ANN model and the observed values were in good
was found to be the best in this case (RMS error value agreement for both training and testing data (arith-
of 0.0591). To check the overall performance of the metic mean error of 0.0031 and 0.0202 dS/m, absolute
JAWRA 96 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION
Soil Salinity Prediction using Artificial Neural Networks

arithmetic mean error of 0.5072 and 0.4943 dSIm, and of ANN models, it is still difficult to comment on its
correlation coefficient of 0.80 and 0.89, for training capability to simulate salt build-up trends, which are
and testing data, respectively) (Table 1). On compar- observed to be different in different ranges. Pate], et
ing the salinity predicted by the generalized ANN al. (2000) also reported that in spite of the overall
model with that predicted by the ANNs developed in good performance of ANN models, differences in sim-
the previous step, for individual water tables, it is ulated models and observed values in certain ranges
clear from Figure 4 that they are very close to each was found to be more pronounced, however, further
other (correlation coefficient 0.99). The paired t-test of clarification was not provided. Therefore, the ability
this data revealed no significant difference (P<0.05) of ANNs to distinguish between these different trends
between simulation results. This implies that the under different scenarios was evaluated by plotting
ANN developed with combined data adequately simu- the observed and predicted salinity values (Figure 5).
lates salinity with varying water table depths. It is evident from Figure 5 that the observed and sim-
As mentioned earlier, there were different salt ulated salinities are in close agreement. Thus, the
build-up patterns depending upon the water table salt build-up trends at various depths for lysimeters
depths, subirrigation water salinity, depth of salinity under both water tables is well simulated by the ANN
measurement, and days of salinity measurement model (Figures 5a through 5f).
(Patel, 1997). Correlation coefficients between the
observed and simulated salinity (0.86 to 0.91), abso-
lute arithmetic mean error (0.1820 to 0.5046 dS/m), ANN Models With Mean Salinity Data
arithmetic mean error (-0.0022 to 0.0149 dS/m), and
standard deviation (0.2413 to 0.6869 dS/m) indicated Inherent variability in the observed data is evident
a good agreement between observed and predicted from Figure 5 and is demonstrated by the dispersion
salinity values (Table 1). Although, the above statisti- (standard deviation) in observed salinity for diffe:rent
cal analysis establishes the overall good performance replicates under a particular set of conditions. This

6
Ec
.4
c
o

0 2 4 6 8

Predicted salinity (dS/m)


Separate water table data

Figure 4. Comparison of Salinity Predicted by Models Developed With Separate Water Table
Data and the Model Developed With the Combined Data for Both Water Tables.

JOURNAL OF THE AMERICAN WATER RESOURCES AssociATioN 97 JAWRA


Patel, Prasher, Goel, and Bassi

.u. 8 ,
a. At 0.1 m depth with 0.4 m water table
8 At O.3 dgth table

6:,
qp
O.mL1jaj
6

4 4
.2 2 0
(I) ,
0 20 40 60 80 0 20 40 60 80

rv
8 c. At 0.1 m depth with O.8rnwater table
4fiuLlffuj
8 d. At 0.3 mdepth with 0.8 m water table
i uj qj
6 6

4 4

2 2

0 0
0 20 40 60 80 0 20 40 60 80

e. At 0.Srndepth with 0.8mwater table f. At 0.7 m depth with 0.8 m water table
8 8'

6: 6'

4 4

2'
= Do 2

0 I I

0 20 40 60 80 0 20 40 60 80
Days after planting Days after planting

Observed 0 1 dS/m 0 5 dS/m 9 dS/m


Predicted ldS/m • 5dS/m 9 dS/m
Standard deviation ldS/m o 5dS/m U 9 dS/m
(observed)

Figure 5. Observed and Predicted Soil Solution Salinity in the Lysimeters With
Water Tables at 0.4 m and 0.8 m, Subirrigated With 1, 5, and 9 dSIm Water.

variability can be eliminated using a mean value. bined. The minimum RMS error values obtained
Therefore, ANN models were developed using the using the mean salinity values were 0.0218 (24 PEs),
mean salinity values for the 0.4 m and 0.8 m water 0.0493 (17 PEs), and 0.401 (14 PEs) (Figure 1, Table
tables separately and with both water tables corn- 1). The correlation coefficients between the observed
JAWRA 98 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION
Soil Salinity Prediction using Artificial Neural Networks

and simulated values were 0.98, 0.95 and 0.94 for types of replicate data used for training the ANNs.
training data of the three cases (Table 1), and the cor- The ANN models developed separately for each water
responding values for testing data were 0.99, 0.94, table treatment did not perform better than the model
and 0.96 for testing data (Table 1). These values for developed for the combined dataset. The results of
the entire data set were 0.98, 0.95, and 0.95. These this study also indicate that either replicates data or
values apparently indicate that the ANNs developed the mean data for the replicates of the given treat-
with the mean data are more accurate than those ments can be used for the development of effective
developed with the data obtained from individual ANN models.
replicates (Table 1). However, a closer scrutiny of the
results shows that this is not the case. An ANN model
predicts a single salinity value for a given set of condi- LITERATURE CITED
tions (i.e., for a particular treatment with 12 repli-
cates). In this case there were 12 different input Batchelor, W. D., X. B. Yand, and A. T. Tschanz, 1997. Development
values and only one ANN predicted output. Therefore, of a Neural Network for Soybean Rust Epidemics. Transactions
of the ASAE 40(1):247-252.
although salinity values from a number of replicates Broner I. and S. Comstock, 1997. Combining Expert Systems and
may be supplied for training, the mean observed Neural Networks for Learning Site Specific Conditions. Comput-
salinity value for a given set of conditions could be ers and Electronics in Agriculture Journal 19:37-53.
used for the comparison with that of the predicted Chakraborty, K., K Mehrotra, C. K. Mohan, and S. Ranka, 1992.
salinity value. Thus, for ANNs developed using repli- Forecasting the Behavior of Multivariate Time Series Using
Neural Networks. Neural Networks 5(6):961-970.
cate data, the correlation between the observed mean Clair, T. A., and J. M. Ehrman, 1998. Using Neural Networks to
salinity values and predicted salinity values was also Assess the Influence of Changing Seasonal Climates in Modify-
determined for the entire dataset. In this case, the ing Discharge, Dissolved Organic Carbon, and Nitrogen Export
correlation coefficient values were found to be 0.98, in Eastern Canadian Rivers. Water Resour. Res. 34(3):447-455.
0.95, and 0.94 for the models developed using the Comfort, S. D., W. P. Inskeep, and R. H. Lockerman, 1993.
Observed and Simulated Transport of a Conservative Tracer
replicate data from the 0.4 m water table, 0.8 m water Under Line-source Irrigation. J. Environment Quality 22(3):554-
table, and the combine dataset, respectively (the cor- 561.
relation coefficient for the entire replicate data set Deck, S. H., C. T. Morrow, P. H. Heinemann, and H. J. Sommer III,
were 0.87, 0.89, and 0.87) (Table 1). These values are 1995. Comparison of a Neural Network and Traditional Classifi-
almost the same as those obtained with the ANNs er for Machine Vision Inspection of Potatoes. Applied Engineer-
ing in Agriculture 11(2):319-326.
developed using the mean salinity data. Therefore, Ghazanfari, A., J. Irudayaraj, and A. Kausalik, 1996. Grading Pis-
ANN models, which were developed using either tachio Nuts Using a Neural Network Approach. Transactions of
replicate data or mean values, predicted similar salin- the ASAE 39(6):2319-2324.
ity values. Gimblett, R. H. and G. L. Ball, 1995. Neural Network Architectures
for Monitoring and Simulating Changes in Forest Resource
Management. Al Applications 9(2):103-123.
Hsu, K.-L., H. V. Gupta, and S. Sorooshian, 1995. Artificial Neural
Network Modelling of the Rainfall-Runoff Process. Water
CONCLUSIONS Resour. Res. 31(1O):2517-2530.
Jayawardena, A. W. and D. A. K. Fernando, 1995. Artificial Neural
Artificial neural networks were successfully Networks in Hydro-Meteorological Modeling. Developments in
Neural Networks and Evolutionary Computing for Civil and
applied to predict soil salinity values in lysimeters Structural Engineering, Civil-Comp Press, Edinburg, United
irrigated with brackish water through subirrigation. Kingdom, pp. 115-120.
The depth of the water table, the salinity level of the Kok, R., R. Lacroix, G. Clark, and E. Taillefer, 1994. Limitation of a
irrigation water, the depth of measurement of the soil Procedural Greenhouse Model With an Artificial Neural Net-
salinity, and the number of days after the initiation of work. Canadian Agricultural Engineering 36(2):117-126.
Lacroix, R., F. Salehi, X. Z. Yang, and K. M. Wade, 1997. Effects of
irrigation were all considered as input parameters. Data Processing on the Performance of Artificial Neural Net-
Feed-forward, back propagation ANN models with a works for Dairy Yield Prediction and Cow Culling Classification.
varying number of processing elements were devel- Transactions of the ASAE 40(3):839-846.
oped. The RMS error values from ANNs using salinity Maier, H. R. and G. C. Dandy, 1996. The Use of Artificial Neural
Networks for the Prediction of Water Quality Parameters. Water
data from replicate lysimeters, ranged from 0.0674 to Resour. Research 32(4):1013-1022.
0.0602, 0.0766 to 0.0576, and 0.0736 to 0.0591, Nakano, K., 1997. Application of Neural Networks to the Color
respectively, for datasets corresponding to the water Grading of Apples. Computer and Electronics in Agriculture
table depths of 0.4 m and 0.8 m, and a combined 18(2,3): 105-116.
dataset for both depths. The correlation coefficients NeuralWare, 1993. Neural Computing: A Technology Handbook for
Professional Il/Plus and NeuralWorks Explorer. NeuralWare
values between observed salinity values and those Inc., Pittsburgh, Pennsylvania.
predicted by the ANN models with the minimum
RMS error were 0.86, 0.88, and 0.87 for the three
JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 99 JAWRA
Patel, Prasher, Goel, and Bassi

Pachepsky, Y. A., D. Timlin, and G.Varallyay, 1996. Artificial Neu-


ral Networks to Estimate Soil Water Retention from Easily
Measurable Data. Soil Sci. Soc. Am. J. 60:727-733.
Patel, R. M., 1997. Subirrigation with Brackish Water. Unpublished
Ph.D. Thesis, McGill University, Montreal, Canada.
Patel, R. M., 5. 0. Prasher, P. K. Goel, and R. B. Bonnell, 2000.
A Neural Network Approach to Predict Soil Salinity. ICID J.
49(2):67-78.
Seginer, I., 1997. Some Artificial Neural Network Applications to
Greenhouse Environment Control. Computers and Electronics
in Agriculture 18(2,3):167-180.
Shukla, M. B., R. Kok, S. 0. Prasher, G. Clark, and R. Lacroix,
1996. Use of Artificial Neural Networks in Transient Drainage
Design. Transactions of the ASAE 39(1):119-124.
Van den Broek, J. A. Elbers, J. Huygen, P. Kabat, J. G. Wresseling,
J. C. van Dam, and R. A. Feddes, 1994. SWAP 1993 — Input
Instructions Manual. Vakgroep Waterhuishouding, Nieuwe
Kanaal, Wageningen.
Wang, D., F. E. Dowell, and R. E. Lacey, 1999. Single Wheat Kernel
Color Classification Using Neural Networks. Transactions of the
ASAE 42(1):233-240.
Wen, C.-G. and C-S. Lee, 1998. A Neural Network Approach to
Multiobjective Optimization for Water Quality Management in a
River Basin. Water Resour. Research 34(3):427-436.
Workman, S. R., J. E. Parsons, G. M. Chescheir, R. W. Skaggs, and
J. F. Rice, 1994. Drainmod User's Guide. USDA, Natural
Resources Conservation Service, Washington, D.C.
Yang, C.-C., S. 0. Prasher, and J. A. Landry, 1999. Use of Artificial
Neural Networks to Recognise Weeds in a Corn Field. Journée
d'information Scientifique et Technique en genie agroalimen-
taire, Saint-Hyacinthe, Quebec, Canada, pp. 60-65. Conseil des
productions végétales du Québec, Inc.
Yang, C.-C., S. 0. Prasher and R. Lacroix, 1995. Applications of
Artificial Neural Networks to Land Drainage Engineering.
Transactions of the ASAE 39(2):525-533.

JAWRA 100 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

You might also like