Environmental Modelling & Software 63 (2015) 73e86

Contents lists available at ScienceDirect

Environmental Modelling & Software

journal homepage: www.elsevier.com/locate/envsoft

A Fuzzy Decision Support System for irrigation and water

conservation in agriculture
E. Giusti a, b, S. Marsili-Libelli a, *
a
b

Department of Information Engineering (DINFO), School of Engineering, University of Florence, Italy

ATI srl, Empoli, Italy

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 18 April 2014
Received in revised form
24 September 2014
Accepted 25 September 2014
Available online

Since agriculture is the major water consumer, web services have been developed to provide the farmers
with considerate irrigation suggestions. This study improves an existing irrigation web service, based on
the IRRINET model, by describing a protocol for the eld implementation of a fully automated irrigation
system. We demonstrate a Fuzzy Decision Support System to improve the irrigation, given the information on the crop and site characteristics. It combines a predictive model of soil moisture and an
inference system computing the most appropriate irrigation action to keep this above a prescribed safe
level. Three crops were used for testing the system: corn, kiwi, and potato. This Fuzzy Decision Support
System (FDSS) favourably compared with an existing agricultural model and data-base (IRRINET). The
sensitivity of the FDSS was tested with random rainfall and also in this extended case the water saving
was conrmed.
2014 Elsevier Ltd. All rights reserved.

Keywords:
Decision support systems
Irrigation management
Climate changes
Fuzzy inference systems
Fuzzy modelling

1. Introduction
Water conservation in agriculture is becoming an increasingly
important issue in the Mediterranean countries, in view of the
current changes both in climate and in the agricultural practices.
There is a great need for irrigation system modernization to cope
with the increasing value and decreasing availability of this commodity. Decision Support Systems (DSS) are now recognized as a
fundamental tool in environmental management and planning.
Since Guariso et al. (1985) rst introduced DSS many studies have
reported advances in their use for the management of water
resource. McIntosh et al. (2011) reviewed the experience of a global
group of environmental decision support systems (EDSS) developers, emphasizing the key challenges and structures in EDSS
development. The integration between large-scale models and
socio-economic and environmental issues in the EDSS context was
considered by van Delden et al. (2011) and by Lehmann and Finger
(2014), who developed a bio-economic model to optimize management decisions in potato production in the Broye catchment
(Switzerland). The importance of involving stakeholders in the
irrigation management was considered in MONIDRI (Fais et al.,
2004) advocating participatory planning, whereas the relevance

* Corresponding author.
E-mail addresses: stefano.marsililibelli@uni.it, Sig-zzig@libero.it (S. MarsiliLibelli).
http://dx.doi.org/10.1016/j.envsoft.2014.09.020
1364-8152/ 2014 Elsevier Ltd. All rights reserved.

of farmers as valuable sources of local knowledge was underlined

by David et al. (2012). Integrated DSS have been developed for the
Mediterranean countries (IRRIGATE, Merot and Bergez, 2010),
southern Italy (Portoghese et al., 2013), and north-west China (Ge
et al., 2013). State-of-the-art information technology tools were
adopted in dealing with the smart management of the water
resource: the multi-agent approach for spatial modelling was used
by van Oel et al. (2010), whereas the fuzzy approach in the denition of a multi-criteria decision framework was adopted by Chen
and Paydar (2012). On the control side, an application of model
predictive control has been reported (McCarthy et al., 2014),
requiring very detailed information on weather, soil and crop to
calibrate a complex crop model.
The reference model used in our study is IRRINET (Rossi et al.,
2004; Mannini et al., 2013) that assists farmers in their irrigation
scheduling. The web service engineered around the IRRINET model
provides the farmers with an irrigation suggestion based on the
kind of crop, soil characteristics, plot location, and meteorological
data (temperature, precipitation, etc.). The advice, issued on a daily
basis, suggests the amount of irrigation to be applied to the plot in
order to keep the soil water content in a safe band between the eld
capacity and the wilting point, providing just the right amount of
water for full crop development. The software assumes that the
previous advice has been fully implemented and therefore it
operates on an open-loop basis, i.e. without incorporating any information on whether any irrigation (suggested or otherwise) was
actually applied.

74

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

List of symbols and abbreviations

DT
mi

Symbol
a,b,d
c
ETc
Irr
I
m
N
P
RI
Tsum
U
Ulow
Uhigh
vi

Abbreviations
mfs
Membership functions
CER
Canale Emiliano Romagnolo
DSS
Decision Support System
EDSS
Environmental Decision Support Systems
FCM
Fuzzy C-Means (Bezdek, 1981)
FDSS
Fuzzy Decision Support System
IPI
Irrigation Performance Index
Nmfs
Number of membership functions
PP
Phenophase
RF
Rain Forecast
RTU
Remote Terminal Unit
SCADA Supervisory Control And Data Acquisition
VAF
Variance Accounted For

Fuzzy model consequent parameters

Number of clusters
Crop Evapotranspiration (mm/d)
Irrigation (mm)
Irrigation singleton (mm)
Fuzzy exponent
Number of experimental observations
Vector of consequent model parameters
Sum of daily rain and irrigation (mm)
Growing Degree Days (GDD) ( C d)
Soil moisture (mm)
Soil moisture lower threshold (mm)
Soil moisture upper threshold (mm)
Centroid of i-th cluster

The IRRINET software is managed by CER (Canale Emiliano

Romagnolo), a water consortium in the Emilia-Romagna region,
Northern Italy, and the irrigation advices it provides are based on a
very complex mathematical model developed by CER itself (Rossi
et al., 2004; Mannini et al., 2013) that describes the water balance in the soil based on the PenmaneMonteith equation. This
model estimates the daily water needs for a given crop based on
past irrigations, soil characterization, and current meteorological
conditions. The IRRINET model is context-dependent and requires
real-time data for computing the next irrigation, therefore it cannot
be used outside its native environment nor can it be operated in an
off-line mode to generate alternate irrigation strategies. In the
present context we use this model as a tested data source, which
together with the eld data (precipitation, temperature, etc.),
provides the necessary benchmark for DSS calibration and testing.
1.1. The IRRISAVE project
Recently a joint venture by A.T.I. srl and the Department of Information Engineering of the University of Florence (DINFO) was
set-up to develop a prototype irrigation system, called IRRISAVE,
conceived to overcome the present limitations of IRRINET:

The implementation of the advice is left to the goodwill of the
farmer, since no automatic actuation is provided;

Each advice assumes that the previous one was implemented. In
the medium to long range this may lead to a considerable
divergence between the actual and expected crop condition if
some suggestions were not put in practice;

The advice is based on the current meteorological conditions,
disregarding micro-scale climate or any short range rain
forecast.
To mend the above shortcomings, the IRRISAVE irrigation system should provide a fully automated system of which the soil
moisture model and the irrigation decision maker form the software core.
Fig. 1 illustrates the development stages of the IRRISAVE
project, the rst two of which are described in this paper. In Step
1 an independent soil moisture model is calibrated with the
IRRINET database and its internal model. In Step 2 a set of irrigation rules is designed to automate the irrigation suggestions
and assess the relative water saving by comparison with the
IRRINET advice. In Step 3, yet to be implemented, the FDSS will

Daily variation of Tsum ( C d)

Degree of activation

be deployed on the eld, closing the loop from the irrigation

advice back to the FDSS. The algorithmic parts of the project
described here were developed in the Matlab platform, using the
Fuzzy Toolbox for developing the inference system producing the
irrigation advice.
2. Structure of the decision support system
The inner structure of the DSS of Fig. 1 is detailed in Fig. 2. It is
composed of three main parts:

The predictive soil moisture model, calibrated with data produced
by the IRRINET model using the same inputs drawn from its
agro-meteorological database;

The irrigation inference system, consisting of a fuzzy inference
system to decide the timing and amount of irrigation on the
basis of the crop phenophase, previous irrigations, and the
prescribed soil moisture thresholds;

The irrigation performance index (IPI), consisting of the sum of
the past irrigations

IPI

Irrk;

(1)

to which additional energy costs may be added to take into account

the pumping overhead costs. This renement can only be added
after the kind of irrigation (sprinkler, ridge and furrow, etc.) has
been specied as each one has its own peculiarity, efciencies and
additional operating costs.
In the sequel each module is described in details and the performance of the FDSS is compared against IRRINET in terms of
consistency of the advice and water savings. The possibility that the
computed advice may impair crop productivity by providing a suboptimal irrigation is ruled out by the fact that IRRISAVE adopts the
same safe soil moisture thresholds that were tested by IRRINET in
many years of successful farming practice (Mannini et al., 2013).
2.1. Analysis of available crop data
The IRRISAVE soil moisture model was calibrated and validated
using the large IRRINET proprietary database including both
vegetable and fruit crops. The database consists of a very large

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

75

Fig. 1. Development stages of the IRRISAVE project e In Step 1 an independent soil moisture model is calibrated with the IRRINET database and internal model. In Step 2 a set of
irrigation rules is designed to automate the irrigation advices and assess the water saving obtained by comparison with the IRRINET scheme. In the Step 3, yet to be implemented,
the FDSS is ported onto the eld, closing the loop from the irrigation advice back to the FDSS.

collection of many differing open-eld crops including climatic

data (temperature, solar radiation, wind, rain, etc.) and agricultural
data such as irrigation, soil composition, crop characteristics, site
location, etc. for a total of 45 input variables. The model output is
the soil moisture (U), expressed in mm. Normally this quantity is
expressed either in weight or volume percent (%) or as the force
with which water is retained between the soil particles (Allen et al.,
1998). This water potential is measured by the negative pressure
(suction) required to extract water from the pores and is expressed
as centibars (cbar) or kilo Pascals (kPa).

The peculiarity of the IRRINET model is the production of a daily

irrigation advice and this requires the soil moisture units to be
compatible with the irrigation amount. The non-standard millimetre unit arise from the internal one-dimensional water balance
model, which determines the amount of water in the root layer
expressed as height of water (mm), which is equivalent to the soilwater content calculated as volume percentage (volume water/
volume soil) for the root depth. The model calculates a root growth
rate at daily steps as a function of crop, water content of deep soil
layer, and evapotranspiration.

Fig. 2. Structure of the decision support system based on a predictive soil moisture model calibrated with the IRRINET-generated agricultural data, an irrigation inference system,
and a performance index to assess the effectiveness of each irrigation advice.

76

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

Of the large collection of variables used by the IRRINET model a

subset of primary inputs was extracted, including the growing
degree days (GDD) (Tsum,  C d), crop evapotranspiration (ETc, mm/
d) as dened by Allen et al. (1998), and the total water input to the
crop, dened as the sum of daily natural precipitation (Rain, mm)
and irrigation (Irr, mm), whereas the model output is the soil
moisture (U, mm). An additional input variable is the crop current
phenophase (PP) providing information of the crop development
and in particular delimiting the active time-horizon during which
calibration data were acquired.
2.2. A fuzzy model for predicting the soil moisture
The IRRINET model could not be incorporated in the DSS scheme
of Fig. 2 because it is conceived to operate in strict coordination
with its proprietary on-line database and lacks the exibility
required to accept synthetic irrigation data as inputs. On the contrary, the model specically developed for this study has a simple
and exible structure and accepts off-line synthetic data to
generate irrigation scenarios, mimicking the response of the IRRINET model for each crop selected in its database. Further, given the
complexity of the underlying processes involved in the water balance in the soilerooteplant continuum and the wide variability of
the cultural parameters depending on the specic crop and plot
location, a fuzzy model was found to be a viable alternative, provided it is calibrated for each specic agricultural situation, instead
of using a deterministic complex relations such as those considered
by Paredes et al. (2014) and McCarthy et al. (2014). Fuzzy models
(see. E.g. Yager and Filev, 1994) have been in use for many decades
and have won a wide acceptance among data-driven methods for
their relatively intelligible structure. In fact the logical rules forming the core of the model can be easily understood and discussed
with end users and stakeholders. Of course, lacking an inner
mechanistic structure, fuzzy model are as good as the data used to
set them up and their domain of applicability is limited to the range
of the training data. In our case this implies that the model will
incorporate information about the kind of crop, soil composition,
plot location and will not be transferable to a differing context
without a proper re-calibration.

The proposed fuzzy model is shown in Fig. 3 and can be regarded as an extension of the classical Sugeno approach (Takagi and
Sugeno, 1985; Yager and Filev, 1994; Babuska, 1998a) with the antecedents being clusters representing typical crop conditions to
which local linear models are associated as consequents. The global
model response is then obtained by defuzzication, consisting of
the average of the consequent linear models weighted by their
degrees of membership. This mixed cluster/linear consequents
approach was proposed by Babuska (1998a) and Abony (2003) and
previously applied to the composting process (Giusti and MarsiliLibelli, 2010) by adapting the public-domain software developed
by Babuska (1998b).
As shown in Fig. 3 the model has three inputs, sampled at daily
intervals during the active phenophase horizon (tjjj 1,,N). The
Growth Degree Days (Tsum), the actual water applied to the crop
(RI), and the evapotraspiration (ETc) form the input vector
z(tj) [Tsum(tj), RI(tj) Rain(tj) Irr(tj), ETc(tj)] whose values are
partitioned into c relevant operating regions (clusters) according to
the Fuzzy C-Means (FCM) algorithm (Bezdek, 1981; Abony, 2003).
These clusters are identied by their labels and their centroids
i
Zi Tsum
; RIi ; ETic  ji 1; ; c, dened by the coordinates of each
cluster centre. The degree of membership of a generic input vector
z(tj) to each cluster Zi determines the degree of activation (mi) of the
corresponding consequent linear model, which contains past input/
output samples, taking the form of Eq. (2)

 
 
b t
Ri : if z tj Zi then U
i j








ai;0 U tj1 bi;0 Tsum tj2 bi;1 Tsum tj3 bi;2 ETc tj2






bi;3 ETc tj3 bi;4 RI tj1 bi;5 RI tj2 di i
1; :::; c:
(2)
b t of each rule (Ri) is computed by the algebraic
The output U
i j
part of Eq. (2), whose coefcients have yet to be determined. The
b t is then obtained as the average of the individual
model output U
j
b t weighted by their respective degrees of
model outputs U
i j
membership mi

Fig. 3. Structure of the soil moisture predictive model.

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

 
b t
U
j

Pc

 
b t
mi $ U
i j
:
Pc
i1 mi

i1

(3)

The parameters P {ai,o,bi,o,,bi,5,diji 1,,c} of the consequent

models and the number of antecedent clusters c are estimated by
nonlinear minimization of the sum of squared differences
(Babuska, 1998b) between the observed soil moisture U(tj) and its
b t over the N simulation days, i.e.
predicted value U
j

b arg min
P
P

N   
 2
X
b t
U tj  U
j

(4)

j1

The goodness-of t of the model was assessed by computing the

variance-accounted-for (VAF) dened as

PN    b  2 C
B
B
C
j1 U tj  U tj
C
VAF 100B
2 C;
PN   
B1 
@
A
U
U
t

j
j1

(5)

where Uis the average soil moisture. Equation (5) compares the
b in the numerator to the
model error variability Utj  Ut
j
intrinsic data variability Utj  U in the denominator. The summation in Eq. (5) is extended to all the N observations in the active
time-horizon.

2.3. Fitting the fuzzy model to the IRRINET crop data

Three widespread crops were selected to calibrate the fuzzy
model Eqs. (2)e(3) to the data extracted from the IRRINET database: corn (Zea mays L.), kiwi (Actinidia chinensis Planch) and potato
(Solanum tuberosum L.). The data were drawn from plots grown in

77

the Ferrara Province, Northern Italy, in 2008, as shown in Fig. 4,

where agro-meteorological data were recorded. The validation was
performed using data from the same plot and a different year, using
the experimental inputs for the pertinent year.
The following gures and tables show the model t to the corn
data (Fig. 5 and Table 1), using the actual precipitation and IRRINET
irrigations. Similar ts are shown for the kiwi (Fig. 6 and Table 2)
and for the potato (Fig. 7 and Table 3). In all cases the simulation
horizon begins with phenophase 2 (emergence) and ends with the
last phenophase of farming interest, depending on the crop. It can
be seen that the performance (VAF) degradation from calibration to
validation is about 5%, which can be regarded as acceptable
considering the large climatic differences between the two years
2006 and 2008.

2.4. Structure of the Fuzzy Decision Support System (FDSS)

The predictive soil moisture model just developed is now
incorporated into the Decision Support System (DSS). The same
motivations supporting the fuzzy choice for the model apply to the
DSS as well, for which the fuzzy approach consists of a set of easily
understandable rules that can be discussed with end users and
stakeholders. Hereafter the DSS will be referred to as a Fuzzy Decision Support System (FDSS). Its goal is to keep the soil moisture
above a prescribed threshold, inherited from IRRISAVE to ensure
full agricultural production, by using the minimum quantity of
water resulting from Eq. (1). Irrigation efciency and energy costs
may be added to IPI at a later stage, when the characteristics of the
irrigation system will be more precisely dened. The FDSS task is to
keep the soil moisture above a prescribed safe threshold by coping
with the variations in evapotranspiration ETc, precipitation, and
irrigation.
To assess its efciency, the FDSS is run in the same IRRINET
conditions and the irrigation totals of Eq. (1) are compared. In this
test the soil moisture is kept at the same threshold value of IRRINET,

Fig. 4. Growth area of the crops used in this study, in the Ferrara province, north-eastern Italy. The dark dots indicate the location of the rain gauges, whereas the hollow dots refer
to the crop locations.

78

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

Fig. 5. Calibration (left) and validation (right) of the fuzzy model Eqs. (2)e(3) with the Corn data from the IRRINET database. VAF as a goodness-of-t measure is computed according to eq. (5).

Table 1
Calibration and validation results for the Corn fuzzy cluster model. VAF as a goodness-of-t measure is computed according to eq. (5).
Crop: Corn
VAFcalib 99.0608
VAFvalid 93.3315
DVAF 5.78%
Antecedents

Location: Berra
Calibration year: 2008; validation year: 2006

Consequent model parameters

Label

Centroids

ao

b3

b4

b5

z1 (L)

Tsum 443.1945
RI 2.5798
ETc 3.2284
Tsum 655.5432
RI 2.4756
ETc 3.8597
Tsum 1980.5421
RI 1.9579
ETc 1.6125

0.970

1.105

1.123

3.191

1.963

0.054

0.104

6.884

1.022

0.837

0.843

2.906

1.144

2.129

0.337

7.589

0.993

0.0619

0.062

0.131

0.652

1.040

0.083

1.194

z2 (M)

z3 (H)

bo

b1

at least in the central phenological period, whereas at the beginning and at the end of the growth season, the FDSS uses a variable
threshold to better adapt the irrigation to the changing crop
phenology.
The inner structure of the FDSS is shown in Fig. 8. In addition to
the soil moisture model of Sect. 2.2. the decision process is split in
two hierarchical levels: rst the need for irrigation is checked, then
the its amount is determined.

b2

2.4.1. First level: irrigation needs

At the upper level, the decision to irrigate is based on the
following joint conditions.

  
 
 


 


b t U t
AND Irr tj1 0 AND Irr tj2 0 ;
U
j
th j

(6)

with tj being the current simulation day. The rst condition ensures
that the soil moisture does not drop below the minimum threshold

Fig. 6. Calibration (left) and validation (right) of the fuzzy model Eqs. (2)e(3) with the Kiwi data from the IRRINET database. VAF as a goodness-of-t measure is computed according to eq. (5).

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

79

Table 2
Calibration and validation results for the Kiwi fuzzy cluster model. VAF as a goodness-of-t measure is computed according to eq. (5).
Crop: Kiwi
VAFcalib 97.8389
VAFvalid 92.6242
DVAF 5.33%

Location: Berra
Calibration year: 2008; validation year: 2006

Antecedents

Consequent model parameters

Label

Centroids

ao

b1

b2

b4

b5

z1 (L)

Tsum 736.5458
RI 1.8308
ETc 3.6901
Tsum 1881.9935
RI 20.1291
ETc 4.3558
Tsum 2394.8524
RI 0.2053
ETc 3.6378

0.948

0.081

0.081

1.480

0.266

0.875

0.065

11.878

0.971

0.066

0.065

1.093

0.396

1.046

0.237

4.442

0.022

0.669

0.288

1.666

0.034

1.283

z2 (M)

z3 (H)

1.000

bo

0.0224

Uth, whereas the other two prevent repeated irrigation on two days
running. The soil moisture threshold value Uth is crop-specic and
is dened as a time-varying prole, depending on the crop
phenology, as follows:

if PP  3

then

if 3 < PP  4

then

if 4 < PP  PPn  1

then

if PPn  1 < PP  PPn

then

b3

phenophases. They are motivated by the varying water demand

during the crop development and by water conservation considerations in relation to water stress (Stone et al., 2001; Garcia y Garcia
et al., 2009; Shivakumar et al., 2011; Aydinsakira et al., 2013).

 
Uth tj Ulow
 

Uhigh  Ulow 
Uth tj Ulow
tj  t3
t4  t3
 
Uth tj Uhigh
 

Ufinal  Uhigh 
Uth tj Uhigh 
tj  tn1
tn  tn1

where n is the last phenophase of interest (i.e. until harvesting).

During phenophase 2 the threshold is kept at its low value Ulow, from
which it is gradually increased to Uhigh, the value inherited from
IRRINET, until phenophase 4 is reached. This high value is maintained until the crop enters the penultimate phenophase PP(n1),
and it is then gradually decreased to Unal when the crop approaches
the last phenophase PP(n). These threshold variations are an
improvement over IRRINET, which only uses Uhigh during the central

(7)

2.4.2. Second level: irrigation amount

When the conditions in Eq. (6) are jointly met the amount of
irrigation for the current day tj is decided on the basis of the
following quantities derived from the local data:

1:

Daily variation of Tsum

 


DTj Tsum tj  Tsum tj1

(8)

Fig. 7. Calibration (left) and validation (right) of the fuzzy model Eqs. (2)e(3) with the potato data from the IRRINET database. VAF as a goodness-of-t measure is computed
according to eq. (5).

80

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

Table 3
Calibration and validation results for the Potato fuzzy cluster model. VAF as a goodness-of-t measure is computed according to eq. (5).
Crop: Potato
VAFcalib 98.6849
VAFvalid 93.7375
DVAF 5.01%

Location: Ciarle-Mirabello
Calibration year: 2008; validation year: 2006

Antecedents

Consequent model parameters

Label

Centroids

ao

b1

b2

b4

b5

z1 (L)

Tsum 851.8959
RI 4.1583
ETc 2.5932
Tsum 1721.2683
RI 4.9185
ETc 0.6128
Tsum 3282.4682
RI 0.5287
ETc 1.4991

0.769

0.247

0.245

0.857

0.154

0.471

0.070

9.729

0.985

0.081

0.079

0.578

0.603

1.380

0.124

3.637

0.981

0.078

0.076

1.087

0.055

0.158

0.056

8.781

z2 (M)

z3 (H)

2:

bo

Cumulative rain forecast for the next two days

2
X

RFj

 
Rain tji

(9)

b3

  Pn m $Irr
i
i1 i
I tj P
n
i1 mi

(12)

i0

3:

Current evapotranspiration

 
ETcj ETc tj

(10)

The FDSS consists of a set of fuzzy rules with the following

structure.







Ri : IF DTj is Ti AND RFj isRi AND ETcj is Ei THEN Irri Iri
i 1;;n;
(11)
where Ti, Ri, and Ei are time e invariant fuzzy sets whose membership
functions (mfs) will be dened later, and Iri are the possible (singleton)
irrigation values (Iri 10;30;60). The combined antecedent degree of
membership of the i-th rule (mi), obtained by AND-ing the individual
degrees, represents the relative weight of the corresponding singleton
(Iri). From Eq. (11) the actual irrigation is then obtained by
defuzzication

The antecedents mfs for the three input variables of Eqs.

(8e10) were obtained by clustering the observed behaviours
with the Fuzzy C-Means (FCM) algorithm (Bezdek, 1981;
Babuska, 1998a), as shown in of Fig. 9. Since analytically
dened mfs are required in the fuzzy algorithm, these shapes
were approximated by the mathematical mfs of Eqs. (13e15) by
adjusting their parameters.
Low (L): Z-shaped mf, with two parameters (a, b)

8
1
>
>
>


>
>
xa 2
>
>
>
12
>

 <
ba
zmf x; a; b


>
>
xb 2
>
>
2
>
>
ba
>
>
>
:
0

x<a
ax

ab
2

ab
<x  b
2
x>b

Fig. 8. Integration of the predictive soil moisture model of Fig. 3 in the Fuzzy Decision Support System.

(13)

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

81

Fig. 9. Numerical membership functions of the three antecedent variables obtained by clustering the operational data from the IRRINET database.

Medium (M): Double Gaussian mf, with four parameters.

(m1,s1,m2,s2)

8
>
xm1 2
>
>

>
>
2
>
2s
>
1
x  m1
>e
<
 >

gauss2mf x;m1 ;s1 ;m2 ;s2
1
m1 <x  m2 m1 <m2
>
>
>
>
2
> xm2
>
>

>
2
>
:e
2s
2
x>m2
(14)
High (H): S-shaped membership function, with two parameters
(a, b)

8
0
>
>
>


>
>
xa 2
>
>
>
2
>
<
ba
smf x; a; b


>
>
xb 2
>
>
1

2
>
>
ba
>
>
>
:
1

x<a
ax

ab
2

(15)

ab
<x  b
2
x>b

It can be seen that the two side-functions zmf and smf are
complementary to one another, i.e. smf 1  zmf . The parameters
of the mfs Eqs. 13e15 were adjusted in two stages. In the rst stage
the parameters were chosen to minimize the sum of squared errors
with the data-generated memberships of Fig. 9. In this way the
intermediate humps of the numerical mfs, due to the constraint
P
m 1, inherent to the FCM clustering procedure, were eliminated. In the second stage, these analytical mfs were used as initial

Table 4
Optimized parameters of the nal membership functions and singleton
consequents.
Antecedents

Consequent

DT

RF

ETc

4.35
20.60

3.21
11.10
2.82
13.01

8.89
20.50

1.09
34.36

3.79 17.90
11.63 46.42
6.14
23.90

Ir

1.38
4.80

0.61 0.95
2.81 6.02
1.30
3.77

10

30

60

conditions for a further tuning where the goal was the minimization of the water quantity, given the model prediction and the soil
moisture threshold of Eq. (7) as constraints. The parameters of the
nal mfs are listed in Table 4 whereas the three groups (datagenerated, tted, nal) of mfs are shown in Fig. 10.
2.4.3. Denition of the FDSS inference rules
At each simulation day tj the irrigation logic computes the
amount of irrigation depending on the three input variables
dened by Eqs. (8e10). The FDSS decision rules were obtained by
N N

starting with the complete set of Nmfsant cons 331 81 combinations and eliminating inconsistent or contradictory rules. A
further decimation was then introduced by eliminating the rules
that during the simulations were globally activated by less than 5%.
This produced the rule set of Table 5, with the last rule being an
override option stating that when the rain forecast is High the
irrigation should be Low, irrespective of all the other conditions.
This reduced set of rules should be regarded as a preliminary
example to demonstrate the conceptual viability of the method at
this early stage of the project. They are neither denitive nor
exhaustive and are expected to be further tested and enhanced
during the eld experimentation, possibly making them cropspecic.
2.4.4. Sensitivity of the FDSS scheme to rain perturbations
To test the robustness of the FDSS scheme, a set of random
precipitations were synthesized on the basis of the rainfall
observed at the sites shown in Fig. 4 in the same calibration year
(2008). For each day we computed the rain mean and variance over
the nine gauging stations and tted a Gaussian distribution to those
data, as shown in Fig. 11. During each simulation, a random sample,
drawn from the distribution of the current day, was introduced into
the model as the perturbed rain. A total of 10,000 simulations were
performed for each crop, producing in the results of Fig. 12.
Only the amount of rainfall was randomized without introducing a corresponding time jitter by varying the timing of the rain
event. Likewise no randomization of the growing degree days
(GDD) was introduced because this would reect into the ETc and
hence into the phenophases. In fact, introducing too many degrees
of freedom into the simulations would make the results difcult to
interpret.
The comparison between IRRINET and FDSS is summarized in
Table 6, that also includes the single simulations with the observed

82

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

Fig. 10. Evolution of the antecedent membership functions, from their initial shape identied from the data, to the data-generated clusters of the input variables of Fig. 9, until the
nal shape is reached on the basis of the overall irrigation performance.

Table 5
FDSS Decision rules for irrigation. The qualiers L. M. H refer to the quantities and
variables of Table 4 and not to those of Tables 1e3.
n

Rule

1
2
3
4
5
6
7
8
9
10
11
12

If
If
If
If
If
If
If
If
If
If
If
If

(DT is L) and (RF is L) and (ETc is L) then (Ir is L)

(DT is L) and (RF is L) and (ETc is M) then (Ir is M)
(DT is L) and (RF is L) and (ETc is H) then (Ir is M)
(DT is M) and (RF is L) and (ETc is L) then (Ir is L)
(DT is M) and (RF is L) and (ETc is M) then (Ir is M)
(DT is M) and (RF is L) and (ETc is H) then (Ir is H)
(DT is H) and (RF is L) and (ETc is L) then (Ir is M)
(DT is H) and (RF is L) and (ETc is M) then (Ir is M)
(DT is H) and (RF is L) and (ETc is H) then (Ir is H)
(DT is H) and (RF is M) and (ETc is M) then (Ir is M)
(DT is H) and (RF is M) and (ETc is H) then (Ir is H)
(RF is H) then (Ir is L)

rainfalls, previously considered. A one-sample t-test conrmed that

the difference between the IRRINET advice and the average FDSS
suggestions of Fig. 12 is statistically signicant in all cases. The FDSS
produces on average a water saving, which is larger in the randomized case for corn and kiwi, whereas for potato the saving is
greater in the single simulation case. These results can be justied
by observing the histograms of Fig. 12 where for corn and kiwi there
is a partial superposition of the randomized rain results with the
IRRINET values, implying that for some realizations of the rain
random process the FDSS water consumption may be greater than
that of IRRINET. Also, the histograms of corn and potato are bimodal
with a fraction of the realizations producing a water consumption
fairly near the IRRINET value, though the majority remains
consistently below that value.
3. Discussion
There is a widespread debate on the usefulness and acceptance
of model-based Environmental Decision Support Systems (EDSS).

Fig. 11. Daily precipitation distributions during 2008 at the sites of Fig. 4. The dots represent the daily spatial average value and the bars extends for 1 std. dev. The shadowed
curves represent the tted Gaussian distributions from which the daily rainfall value was sampled and used in the perturbed simulations.

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

83

Fig. 12. Comparison of IRRINET and FDSS irrigation schemes where the latter was applied with 10,000 independent realizations of the randomized precipitation process. (a) Corn;
(b) Kiwi; (c) Potato.

McIntosh et al. (2011) identied two kinds of key challenges in

EDSS development: how to involve the end users and how to
evaluate the EDSS effectiveness. In the present context, the rst
challenge is non-existent because the study was jointly invited by
the web service provider and by the automation enterprise
providing the middleware for the DSS implementation. The project
goal was the transition from a manually implemented advice to a
fully automated solution, possibly improving the water conservation efciency. In this sense, the FDSS presented here is aimed
primarily at being credible rather than optimal, credibility
meaning adequacy relative to an intended use, as aptly dened by
Aumann (2011). Further, the acceptability of this solution should be
assessed in terms of outcomes rather than outputs, as suggested by Matthews et al. (2011). In this case a desirable outcome
would be the achievement of a full automation relieving the farmer
from the manual implementation of the advice (Step 3 in Fig. 1). The
fuzzy approach ensures that the rules determining the irrigation
are always comprehensible, accessible, and adjustable to the specic irrigation equipment in order to achieve more water saving.

For example the rules of Table 5 could be made visible and editable
by the irrigation manager, who can adapt them to the specic crop
needs. As said, the rules listed in Table 5 are neither exhaustive nor
universal, but just an initial seed of knowledge that should be
enhanced with the experience gained through eld practice.
As the project moves forward to its eld implementation further
interactions and renements will be necessary (van Delden et al.,
2011). At this early stage, however, our goal is to test the feasibility of automating the irrigation advice on a simulation platform
(Matlab), with automatic computation of the irrigation at the
faucet. Additional efciency factors related to the type of irrigation
system (e.g. sprinkler, ridge and furrow, etc.) and crop coefcients
could be introduced only after the water distribution system has
been specied. In this sense the approach advocated by Inman et al.
(2011) of evaluating the EDSS based on the perception of groups of
end-users is not yet applicable here, given the early stage of IRRISAVE development, in which our primary goal is to demonstrate
that IRRISAVE could overcome the limitations of IRRINET
mentioned in Sect. 1.1. Following the development of Fig. 1, once the

Table 6
Comparison between IRRINET and FDSS irrigation schemes in terms of water saving, both with observed and randomized precipitations. The FDSS suggestion beats the IRRINET
advice in all cases but one (kiwi validation). The test and val labels stand for test and validation runs. The check sign in the t-test column indicates that the difference between
IRRISAVE and the average FDSS irrigation is statistically signicant.
Crop

Corn
Kiwi
Potato

Observed precipitations

Randomized precipitations

IRRINET irrigation

FDSS irrigation

% Saving

FDSS irrigation

% Average saving

One-sample t-test

test
val
test
val
test
val

226.42
234.79
413.63
4.02.81
138.17
193.15

5.65
1.13
4.24
3.26
34.55
24.61

215.19 15.61

10.33

407.26 19.41

5.71

172.87 17.84

18.12

239.97
237.48
431.93
390.09
211.12
256.20

84

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

Fig. 13. Comparison between the FDSS and IRRINET irrigation polices for the corn crop in both the test (a) and the validation (b) years. The numbers along the variable soil moisture
threshold (dashed line) indicate the beginning of each phenophase.

eld deployment is considered and the link with IRRINET severed,

the model in the FDSS will be recalibrated and integrated with realtime weather and soil moisture data to improve its accuracy.
To test the FDSS at this stage, its model was rst calibrated and
validated with eld data from the IRRINET database. In the process
the VAF criterion of Eq. (5) degrading little more than 5% from
calibration to validation. After integrating the model with the fuzzy
decision maker as in Fig. 2, the irrigation suggestions of the FDSS
were compared to those produced by IRRINET. For a fair comparison the same crop and site were considered, with the difference
that the FDSS decisions were based on the inputs dened by Eqs.
(8e10) and on a time-varying threshold for each phenophase as
described by Eq. (7), instead of a constant threshold as in IRRINET.
To corroborate the analysis even further, the sensitivity of the FDSS
was tested by randomizing the precipitation for the same year
using spatially averaged samples.
It may be argued that water saving was the only concern in the
FDSS design and that the relation between water quantity and crop
productivity was overlooked. In fact our prime objective was to
demonstrate to possibility of designing a fully automated irrigation
system. As far as water saving (our second goal) is concerned, our
system took IRRINET as a reference, which was previously optimized in terms of crop productivity with more than a decade of
eld testing. Since we adopted the same soil moisture thresholds

used by that model, we are condent that the suggestions produced

by FDSS do not impair crop production and that any excess water
would have a negligible effect on productivity.
Using the crop data already used for model calibration might be
considered unfair from the modelling viewpoint, but it is required
for the direct comparison of the two irrigation schemes. Table 6
summarizes the irrigation results in all cases (calibration year,
validation year, randomized rain) showing that only in one case the
FDSS slightly underperforms IRRINET, whereas in all other instances some water saving is achieved. The FDSS irrigation advice is
based on several factors not considered by IRRINET, such as rain
forecast, past irrigations, and current phenophase. The variable
threshold has its maximum set at the IRRINET level to ensure the
full crop development. However this value is limited to the central
phenophases only, and a gradual decrease in introduced in the early
and nal phenophases.
The assessment for corn is shown in Fig. 13. It can be seen that in
the test simulation (Fig. 13a) the soil moisture follows the
increasing threshold during the transition from PP 3 to PP 4,
whereas IRRINET violates this constraint in this period, perhaps
because this scheme does not consider intermediate thresholds. In
details, the comparison for the test run (Fig. 13a) shows that the
FDSS produces a preliminary small irrigation at day 39 to avoid a
violation of the rising threshold. The second irrigation at day 62 is

Fig. 14. Comparison between the FDSS and IRRINET irrigation polices for the Kiwi crop in both the test and the validation years. The numbers along the variable soil moisture
threshold (dashed line) indicate the beginning of each phenophase.

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

85

Fig. 15. Comparison between the FDSS and IRRINET irrigation polices for the Potato crop in both the test and the validation years. The numbers along the variable soil moisture
threshold (dashed line) indicate the beginning of each phenophase.

not matched by IRRINET and is activated again by the rising

threshold. Later, the IRRINET irrigation at day 96 is not replicated by
FDSS, producing instead more suggestions at days 102, 116, 132, and
168, of decreasing magnitude. As a result the FDSS operation produces a water saving of 13.55 water units. In the validation run
(Fig. 13b) the FDSS shows some minor threshold violations during
the central phenophases (4e7) with a negligible water saving, but
then also IRRINET has a violation at day 83. The reduced performance of FDSS in the validation run can be explained with the
strong climatic differences between years 2006 and 2008 and the
need for a more specic tuning of the rules of Table 5 for a waterintensive crop such as corn.
The comparison for kiwi is shown in Fig. 14, where it can be seen
that in the validation run the FDSS slightly underperforms IRRINET,
while producing some water saving in the test run. Kiwi appears to
be a difcult crop, needing frequent watering and requiring a more
specic set of rules than the basic set of Table 5. For the potato crop
Fig. 15 shows that a substantial saving is achieved by the FDSS in
both cases. As a concluding remark, it should be considered that the
same set of rules (Table 5) was applied to very different crops,
whereas crop-specic rules taking into account the differing
phenology and farming practices would surely yield better results
and will be considered in the future developments.
4. Conclusion
This paper has presented an irrigation decision support system
based on a fuzzy inference system (FDSS), which incorporates a
fuzzy soil moisture model. Its development was motivated by the
need to enhance the performance of an existing web-based irrigation advisory system, in view of an independent automation with
dedicated eld hardware.
The fuzzy model inside the FDSS yields an estimate of the soil
moisture based on the growing degree days (GDD), total water
applied to the crop, and crop evapotranspiration (ETc). It was calibrated with agricultural data from the IRRINET database for several
widely grown crops (corn, kiwi, and potato) in the Ferrara province
during 2008 and validated with similar crops and sites for the year
2006. Based on the model forecast the FDSS decides whether an
irrigation is needed and determines its amount by a set of rules
involving the variation of growing degree days (DTsum), the two-day
ahead rain forecast (RF), and the crop evapotranspiration (ETc).
The performance of the FDSS has been tested by comparing its
irrigation suggestions to those produced by IRRINET and in almost
all cases it compared favourably with the latter. To assess the

robustness of the scheme, the analysis was extended to a wider

scenario which considered randomized precipitations in place of
the real ones. In this extended contest, the advantage of FDSS over
IRRINET was conrmed, and in several cases even increased.
The next stage of development, already under way, will be the
porting of the FDSS onto the eld, where it will be connected with a
dedicated sensor/actuator network, required for the complete
automation of the irrigation process.
Acknowledgement
The authors gratefully acknowledge the cooperation Dr. Giancarlo Giannerini, Altavia srl, and of Dr. Tommaso Letterio and Dr.
Roberto Genovesi, Canale Emiliano Romagnolo, for the helpful advices and for granting access to IRRINET model and crop database.
The authors also wish to thank the anonymous reviewer #1 for the
many valuable suggestions.
References
user, Boston, p. 273.
Abony, J., 2003. Fuzzy Model Identication for Control. Birkha
Allen, R.G., Pereira, L.S., Raes, D., Smith, M., 1998. Crop Evapotranspiration: Guidelines for Computing Crop Water Requirements. FAO Irrigation and Drainage
Paper 56, Rome, Italy, ISBN 92-5-104219-5.
Aumann, C.A., 2011. Constructing model credibility in the context of policy
appraisal. Environ. Model. Softw. 26, 258e265.
Aydinsakira, K., Erdala, S., Buyuktasb, D., Bastugb, R., Tokera, R., 2013. The inuence
of regular decit irrigation applications on water use, yield, and quality components of two corn (Zea mays L.) genotypes. Agric. Water Manag. 128, 65e71.
Babuska, R., 1998a. Fuzzy Modeling for Control. Kluver Publ. Co., Amsterdam, p. 260.
Babuska, R., 1998b. Fuzzy Modeling and Identication Toolbox for Use with MATLAB. Lecture Notes, Available on the Net at. http://lcewww.et.tudelft.nl/
babuska.
Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective Function Algorithms.
Plenum Press, New York, p. 256.
Chen, Y., Paydar, Z., 2012. Evaluation of potential irrigation expansion using a spatial
fuzzy multi-criteria decision framework. Environ. Model. Softw. 38, 147e157.
David, M.O., Rob, D.F., Michael, W., Chris, J.H., Heathwaite, A.L., Dave, R.C., 2012.
Valuing local knowledge as a source of expert data: farmer engagement and the
design of decision support systems. Environ. Model. Softw. 36, 76e85.
Fais, A., Nino, P., Dono, G., Severini, S., Bazzof, P., Napoli, R., Giannerini, G.,
Turchi, A., Giovacchini, A., Minelli, U., Perini, L., 2004. MONIDRI - a participatory
IDSS for water use management in agriculture at river basin level. In: IFAC
Workshop on Modelling and Control for Participatory Planning and Managing
Water Systems, Venice, 29th Sept.e1st Oct. 2004.
Garcia y Garcia, A., Guerra, L.C., Hoogenboom, G., 2009. Water use and water use
efciency of sweet corn under different weather conditions and soil moisture
regimes. Agric. Water Manag. 96, 1369e1376.
Ge, Y., Li, X., Huang, C., Nan, N., 2013. A decision support system for irrigation water
allocation along the middle reaches of the Heihe River Basin, Northwest China.
Environ. Model. Softw. 47, 182e192.
Giusti, E., Marsili-Libelli, S., 2010. Fuzzy modelling of the composting process. Environ. Model. Softw. 25, 641e647.

86

E. Giusti, S. Marsili-Libelli / Environmental Modelling & Software 63 (2015) 73e86

Guariso, G., Rinaldi, S., Soncini-Sessa, 1985. Decision support system for water
management: the Lake Como case-study. Eur. J. Oper. Res. 21, 295e306.
Inman, D., Blind, M., Ribarova, I., Krause, A., Roosenschoon, O., Kassahun, A.,
Scholten, H., Arampatzis, G., Abrami, G., McIntosh, B., Jeffrey, P., 2011. Perceived
effectiveness of environmental decision support systems in participatory
planning: evidence from small groups of end-users. Environ. Model. Softw. 26,
302e309.
Lehmann, N., Finger, R., 2014. Economic and environmental assessment of irrigation
water policies: a bioeconomic simulation study. Environ. Model. Softw. 51,
112e122.
Mannini, P., Genovesi, R., Letterio, T., 2013. IRRINET: large scale DSS application for
on-farm irrigation scheduling. Procedia Env. Sci. 19, 823e829.
Matthews, K.B., Rivington, M., Blackstock, K.L., McCrum, G., Buchan, K., Miller, D.G.,
2011. Raising the bar? The challenges of evaluating the outcomes of environmental modelling and software. Environ. Model. Softw. 26, 247e257.
McCarthy, A.C., Hancock, N.H., Raine, S.R., 2014. Simulation of irrigation control
strategies for cotton using model predictive control within the VARIwise
simulation framework. Comput. Electron. Agric. 101, 135e147.
McIntosh, B.S., Ascough II, J.C., Twery, M., Chew, J., Elmahdl, A., Haase, D.,
Harou, J.J., Hepting, D., Cuddy, S., Jakeman, A.J., Chen, S., Kassahun, A.,
Lautenbach, S., Mathews, K., Merrit, W., Quinn, N.W., Rodriguez-Roda, I.,
Sieber, S., Stavenga, M., Sulis, A., et al., 2011. Environmental decision support
systems (EDSS) development e challenges and best practices. Environ. Model.
Softw. 26, 1389e1402.
Merot, A., Bergez, J.-E., 2010. IRRIGATE: a dynamic integrated model combining a
knowledge-based model and mechanistic biophysical models for border irrigation management. Environ. Model. Softw. 25, 421e432.

Paredes, P., Rodrigues, G.C., Alves, I., Pereira, L.S., 2014. Partitioning evapotranspiration, yield prediction and economic returns of maize under various irrigation
management strategies. Agric. Water Manag. 135, 27e39.
Portoghese, I., D'Agostino, D., Giordano, R., Scardigno, A., Apollonio, C., Vurro, M., 2013.
An integrated modelling tool to evaluate the acceptability of irrigation constraint
measures for groundwater protection. Environ. Model. Softw. 46, 90e103.
Rossi, F., Nardino, M., Mannini, P., Genovesi, R., 2004. IRRINET Emilia Romagna:
online decision support on irrigation. Online agrometeological applications
with decision support on the farm level. Cost Action 2004 (718), 99e102.
Shivakumar, H.K., Ramachandrappa, B.K., Mudalagiriyappa, H.V.N., 2011. Effect of
phenophase based irrigation schedules on growth, yield and quality of baby
corn (Zea mays L.). Agric. Sci. 2, 267e272.
Stone, P.J., Wilson, D.R., Reid, J.B., Gillespie, R.N., 2001. Water decit effects on sweet
corn. I. Water use, radiation use efciency, growth, and yield. Aust. J. Agric. Res.
52, 103e113.
Takagi, T., Sugeno, M., 1985. Fuzzy identication of systems and its applications to
modeling and control. IEEE Trans. Syst. Man Cybern. 15, 116e132.
van Delden, H., Seppelt, R., White, R., Jakeman, A.J., 2011. A methodology for the
design and development of integrated models for policy support. Environ.
Model. Softw. 26, 266e279.
van Oel, P.R., Krol, M.S., Hoekstra, A.Y., Taddei, R.R., 2010. Feedback mechanisms
between water availability and water use in a semi-arid river basin: a spatially
explicit multi-agent simulation approach. Environ. Model. Softw. 25, 433e443.
Yager, R.R., Filev, D.P., 1994. Essentials of Fuzzy Modeling and Control. John Wiley &
Sons, p. 408.