Case Study

A Strategy to Assess Water Meter Performance

Clara Cordeiro 1; Ana Borges 2; and M. Rosário Ramos 3

Abstract: Apparent water losses can be problematic to water companies’ revenues. This type of loss is very difficult to detect and quantify
and is often associated with water meter anomalies. This study was motivated by a water company’s challenge that links a decrease in water
consumption to water meters’ malfunction. The aim is to develop a strategy to detect decreasing water usage patterns, contributing to meter
performance assessment. The basis of the approach is a combination of statistical methods. First, the time series of billed water consumption
is decomposed using Seasonal-Trend decomposition based on Loess. Next, breakpoint analysis is performed on the seasonally adjusted time
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.

series. After that, the Mann–Kendall test and Sen’s slope estimator are used to analyze periods of progressive decrease changes in water
consumption, defined by breakpoints. A quantitative indicator of this change is proposed. The strategy was successfully applied to eight-time
series of water consumption from the Algarve, Portugal. DOI: 10.1061/(ASCE)WR.1943-5452.0001492. This work is made available under
the terms of the Creative Commons Attribution 4.0 International license, https://creativecommons.org/licenses/by/4.0/.

Introduction meters is crucial in reducing any uncertainty affecting the water

A recurring problem that troubles water companies is that of timely
detection of malfunctioning water meters. Indeed, defective water
meters impair the management of water supply and diminish water
companies’ revenues. The management of nonrevenue water (NRW)
is one key issue for improving water use efficiency, reducing gaps
between water supply and demand (Ncube and Taigbenu 2019).
Water losses in the distribution system can be categorized as either
real (leakage) or apparent (commercial) losses (Mutikanga et al.
2011). Water meter under-registration resulted in apparent losses
and lost revenue (Moahloli et al. 2019; Fourie et al. 2020), being
apparent losses one of the components of NRW. As Criminisi et al.
(2009) explain, apparent losses are caused by unauthorized con-
sumption and meter inaccuracies, corresponding not to physical
but rather to financial losses. Since they are water volumes taken
from the network and consumed but not accounted for, as a con-
sequence, the company will be jeopardized by the unaccounted
water volume, resulting in an important impact on the utility’s
water and economic balances.
A long-known fact is that reducing apparent losses caused by
meter inaccuracies can result in substantial short-term increases in
utility revenue and lead to increasingly equitable service charges
for long-term water consumers (Richards et al. 2010; Kadenge
et al. 2020). The accurate measurement of the water collected by
1
Assistant Professor, Faculdade Ciências e Tecnologia, Universidade do
Algarve, Faro 8005-139, Portugal; Centro de Estatística e Aplicações,
Faculdade de Ciências, Universidade de Lisboa, Lisboa 1749-016, Portugal
(corresponding author). ORCID: https://orcid.org/0000-0002-1026-6078.
Email: ccordei@ualg.pt
2
Assistant Professor, Centro de Inovação e Investigação em Ciências
Empresariais e Sistemas de Informação, Escola Superior de Tecnologia
e Gestão, Politécnico do Porto, Felgueiras, Porto 4610-156, Portugal.
ORCID: https://orcid.org/0000-0003-4244-5393
3
Assistant Professor, Dept. of Sciences and Technology, Universidade
Aberta, Rua da Escola Politécnica, 147, Lisboa 1269-001, Portugal; Center
of Mathematics, Fundamental Applications and Operations Research,
Faculdade de Ciências, Universidade de Lisboa, Lisboa 1749-016, Portugal.
Note. This manuscript was submitted on July 29, 2020; approved on
September 13, 2021; published online on November 18, 2021. Discussion
period open until April 18, 2022; separate discussions must be submitted
for individual papers. This paper is part of the Journal of Water Resources
Planning and Management, © ASCE, ISSN 0733-9496.
balance and has significant technical and economic implications
(Pacheco et al. 2020). A cubic meter consumed but not measured
reduces the company revenue in quantity equal to the selling price of
the last cubic meter of water consumed by that consumer (Arregui
et al. 2018).
The water companies’ awareness for the responsible use of
water has gained importance, with climate changes emphasizing
this need. In the context of water scarcity, Oviedo-Ocaña et al.
(2020) alert to the importance of assessing losses in water distri-
bution systems since the compensation of water losses represents
an increase in the source’s water supply. Enhancing that water use
efficiency and conservation are priority alternatives to ensure, for
example, universal access to drinking water and reducing the num-
ber of people suffering from water scarcity. Furthermore, Pacheco
et al. (2020) emphasize that water measurement will become an
even more critical aspect in the near future because of the increase
in the world’s water supplies due to population growth. In this con-
text, one form of water companies to be aware of its important role
is to steer the effective water consumption and guarantee an effi-
cient manutention of its equipment, namely the water meters.
Current research confirms that several factors affect a water
meter’s accuracy, particularly water meterage and total registered
volume (Moahloli et al. 2019; Pacheco et al. 2020). Detecting
the point in time in which water meters should be replaced implies
that they can be substituted proactively to minimize the impact of
water consumption nonregistration and under-registration on NRW
(Moahloli et al. 2019). There seems to be no specific determination
and/or agreement on a water meter lifespan or optimal replacement
period. Water meter producers only give a limited warranty against
manufacturing defects but do not specify the meters’ lifetime
(Moahloli et al. 2019). It is believed that the premature replacement
of water meters will result in a higher average life-cycle cost due to
the initial fixed costs. However, if a meter is replaced too late, a
significant loss of revenue caused by meter nonregistration and
under-registration will also increase the average life-cycle cost
(Fontanazza et al. 2015). The main studies on this problem have
adopted approaches based on laboratory studies, which may impli-
cate two significant issues: (1) the high costs inherent to the labora-
tory, and (2) the conditions in which the tests are taken upon are far
from the real scenario. As Moahloli et al. (2019) recently identified,
several methods can be used to determine water meter accuracy,

© ASCE 05021027-1 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2022, 148(2): 05021027

 such as laboratory testing or installing a master meter in series to the
meter on the ground and the net present value analysis. However,
although accurate, the authors also indicate that the first two meth-
ods tend to be expensive and time-consuming. Malfunctioning water
meters impair water supply management and diminish water com-
panies’ earnings, justifying the need for a detailed data analysis of
water consumption.
Past research on water consumption in Portugal has shown that
the Algarve region has the highest value (Reynaud 2015). The offi-
cial statistics on water distributed/consumed per inhabitant for 2017
[from FEMS (2021)] reveal higher values for this region. The aware-
ness toward the responsible use of water has increased significantly
and, in recent years, periods of drought attributed to climate change
emphasized that need. Aware of this problem, water utilities such as
Infraquinta and the Loulé Municipality feel the need to improve and
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.
develop mechanisms for predictive water planning based on data
analysis. In particular, a mechanism able to detect abnormal de-
creasing trends in water consumption, which they believe is asso-
ciated with a possible water meter malfunction.
This paper aims to develop a procedure capable of detecting
decreasing water consumption changes, contributing to meter per-
formance assessment. The decrease in water consumption (pro-
gressive or sudden) is, as explained by Monedero et al. (2016),
associated with water meter malfunction—one of the possible
causes for apparent losses. Thus, it should be investigated by water
utilities in order to mitigate this type of loss. The proposed ap-
proach intends to contribute and provide a helpful tool for the water
utilities within decision analysis on water meter replacement. The
strategy developed uses several statistical methods to detect signifi-
cant decreases (changes) in water consumption after removing the
seasonal pattern.
In this case study, an empirical analysis of billed water con-
sumption data from seven households and a hotel from the Algarve,
Portugal, was performed. Analyzing data on the real conditions on
which the water meters operate should provide more precise results
than those obtained through laboratory analysis. Time series de-
composition is frequently used in many areas, such as in electricity
consumption to evaluate electrical meter data [e.g., Miller and
Meggers (2017) and Miller (2019)], in hydrological studies to
understand groundwater behavior [e.g., Lafare et al. (2016)], and
to predict or detect changes in water consumption [e.g., Ohana-
Levi et al. (2020), Quesnel and Ajami (2017), Hester and Larson
(2016), and Gelažanskas and Gamage (2015)]. Seasonal-Trend
Decomposition of time series based on Loess [(Seasonal-Trend de-
composition by Loess (STL)] (Cleveland et al. 1990) has been suc-
cessfully used in studies of water consumption (Hester and Larson
2016). Researchers often use the STL default smoothing parameters
because specifying these parameters can be subjective rather than
objective. To overcome this issue, the procedure stl.fit (Cristina et al.
2016) is used to obtain a more accurate decomposition based on the
error measure mean absolute error (MAE).
In the context of water consumption analysis, some recent stud-
ies have adopted the methodology of time series decomposition and
breakpoint detection, such as Quesnel and Ajami (2017) and Hester
and Larson (2016). The latter points out that decomposition meth-
ods have been applied in several hydrological studies, and structural
change models have been used in water quality. Also, the opinion of
these authors is that the breakpoint analysis appears underutilized in
water management. Therefore, the procedure developed in this case
study uses breakpoint analysis (Zeileis et al. 2003) to find points
in time at which statistical patterns change after the seasonal pat-
tern was removed. These breakpoints are associated with segments
where the increases or decreases of the trend are more pronounced.
© ASCE 05021027-2
J. Water Resour. Plann. Manage., 2022, 148(2): 05021027
Furthermore, two new formulas estimate the maximum number of
breakpoints and the minimum number of observations in a segment.
In the analysis of hydrological and meteorological variables,
(Gocic and Trajkovic 2013) and (Sharma et al. 2016) conclude that
there is an interaction between trends and breakpoints. The signifi-
cance of the trend and its decreasing rate are analyzed through
Mann–Kendall testing [Mann (1945) and Kendall (1975)] and
Sen’s Slope (Sen 1968). The negative slope means that there is a
decrease in water consumption. The water company must investi-
gate this situation because it could be related to an anomaly in the
water meter, and it might be the ideal time to replace it. A water
consumption indicator named relative magnitude of the change
(RMC) is proposed to provide some guidance in this decision anal-
ysis. The idea is to quantify the change of the decrease in water
consumption for each consumer and identify which ones the water
company should investigate.
The paper is organized as follows: the first section presents a
brief description of the problem and the dataset analyzed. The next
section explains the methodologies applied, followed by a descrip-
tion of the water consumption analysis’s main results on a case
study of seven households and one hotel, ending with the discus-
sion and conclusion and future work.
Methodologies
In this section, the methods that constitute the strategy proposed are
described in detail.
STL Decomposition
A time series is a sequence of observations indexed by time t,
t ¼ 1; : : : ; n, usually ordered in equally spaced intervals. Often,
time series exhibit seasonal behavior and adequate control for a sea-
sonal component is essential before using any statistical model. Let
consider that a time series is a combination of three components:
trend-cycle, which represents an upward or downward movement
over the time horizon; seasonality, which is a repetitive pattern over
time; and the irregular, which is remaining after the other compo-
nents have been removed (Hyndman and Athanasopoulos 2018).
Both trend and seasonality components are potential confounding
features in analysis, so their identification and removal are impor-
tant. Additionally, this characteristic impacts the computation of the
trend-cycle term T t that bears the time series’s interannual variations
but excluding possible variations related to interannual changes in
seasonality.
Seasonal-Trend Decomposition of time series based on Loess
(STL) is used to decompose the time series fY t ; t ¼ 1; : : : ; ng into
a trend (T t ), seasonal (St ), and irregular (I t ) components using non-
parametric regression and is described as
Y t ¼ T t þ St þ I t ; t ¼ 1; : : : ; n ð1Þ
STL was chosen over other decomposition methods in the liter-
ature because it has attractive modeling features. Such as a nonlinear
trend component, the seasonal component is allowed to change over
time and is robust in the presence of outliers. Note that these
observations are problematic since it influences the estimation of
parameters (e.g., in ARIMA models) that will mislead any statistical
conclusion. STL is an iterative nonparametric procedure that repeat-
edly uses a Loess smoother to refine and improve estimates of the
trend and seasonal components [see Cleveland et al. (1990) for more
details]. In common with all nonparametric regression methods,
STL requires the subjective selection of smoothing parameters.
The two main parameters are the seasonal (s.window) and trend
J. Water Resour. Plann. Manage.
 (t.window) window widths. The Loess smoothing parameters must
be defined in advance and set according to the user’s knowledge
about the data or as default values. The function stl() (R Core
Team 2021) is commonly used setting by default the s.window
parameter as periodic. The algorithm proposed in Cristina et al.
(2016), stl.fit() (Cordeiro 2016), overcomes this drawback. It per-
forms an automatic selection of the smoothing parameters based on
minimizing an error measure, obtained through the function accu-
racy() in package forecast (Hyndman et al. 2021). The algorithm
selects the best STL model, which has the smallest error measure
achieved with a specific combination of s.window and t.window.
The root mean square error is an error measure widely used to
access the model performance and predictability. However, its
sensitivity to extreme observations and outliers can mislead data
analysis. The MAE is the error measure chosen to be used in this
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.
work because it is less sensitive to outliers (Hyndman and Koehler
2006) when compared with the root mean square error, and it
presents several advantages as explained in Willmott and Matsuura
(2005). In the forthcoming analysis, the seasonally adjusted time
series
Y t ¼ Y t − St
t ¼ 1; : : : ; n, is considered.
Detect Structural Changes
Some structural disturbances can be observed in the dynamic struc-
ture of the time series. Is important to stay alert for breakpoints,
points in time at which statistical patterns change, as they represent
periods of the time horizon where the increases or decreases of the
trend are more pronounced. The R package strucchange (Zeileis
et al. 2002) is used to detect structural breaks to test for structural
changes in linear regression models, estimating the number of seg-
ments (m) and the set of the breakpoints bp ¼ ft1 ; t2 ; : : : ; tm−1 g,
minimizing the Bayesian information criterion (BIC) and the
residual sum of squares (RSS) (Zeileis et al. 2003). Additionally,
this procedure has a bandwidth parameter (h) that determines the
minimal segment size (h × n observations). In function breakpoints
() (Zeileis et al. 2002) the default bandwidth parameter is h ¼ 0.15
but, it is up to the user to change h to a more suitable value based on
the knowledge of the data. For example, Verbesselt et al. (2010) use
a minimum of 12% of data (23 observations) for a nine-year time
horizon; in Awty-Carroll et al. (2019), this minimum was two years
(46 observations); and in Hester and Larson (2016), the minimum
was three years (36 observations). In line with the previous refer-
ences, the present study contributes with two new expressions for
obtaining the minimum length between consecutive breaks
frequency
h¼
n points (bp );
and the maximum number of breaks
n 0.5
breaks ¼ ¼ ð4Þ
2  frequency h use another decomposition method instead, e.g., classical decom-
where n = length of the time series and frequency is the seasonal
pattern’s length.
Detect Decreasing Trends
Research on an increasing or decreasing trend in time series is an
important issue. It is usually based on the parametric test for the
linear model and in the Mann–Kendall (MK) test [Mann (1945)
© ASCE
J. Water Resour. Plann. Manage., 2022, 148(2): 05021027
and Kendall (1975)] as a nonparametric counterpart. The test takes
Kendall’s Tau coefficient to assess the correlation between the var-
iable of interest and time. Related to MK’s rationale, Sen (1968)
developed a nonparametric estimator for the slope of a linear trend
in the sample of N pairs of data. The Theil–Sen (TS) slope provides
the magnitude of the trend through a robust linear regression. Also,
MK and TS are more flexible in the assumptions than parametric
t-test, accommodate departures from normality, small sample sizes,
and are robust in the presence of outliers [Helsel and Hirsch (2002)].
MK does not depend on the data range and can find a monotonous
trend over nonlinear behavior. So, is well-suited for a trend esti-
mated through STL. Also, it is possible to consider the autocor-
relation by adjusting the variance of the test statistic when the
assumption of independence fails. To investigate a decrease in the
segments, MK and TS are applied on Y  and an estimate of the mag-
nitude of the trend (or slope) is obtained. If the result points to a
significant and negative trend, the breakpoint is relevant and is in-
cluded in the new set bp , where bp ⊆ bp. This analysis is per-
formed using the functions mk.test() and sens.slope(), available in
the R package trend (Pohlert 2020).
ð2Þ
Relative Magnitude of the Change
The RMC is proposed to assess water consumption behavior
change. It is calculated through a ratio that compares the consump-
tion pattern before and after a breakpoint (see Fig. 1). The idea of
RMC is in line with the same principle as in Gocic and Trajkovic
(2013) and is defined by
slpafter − slpbefore
RMC ¼ ð5Þ
jslpbefore j
where slpafter and slpbefore = nonparametric Sen’s Slopes at a
breakpoint tk ∈ bp . This unitless indicator measures the relative
change, or the relative difference, between the slope after (slpafter )
and the slope before (slpbefore ) the breakpoint, in reference to the
slope before the breakpoint. A negative value of RMC is an alert for
a possible meter anomaly. The more negative is RMC, the higher is
the reduction of water consumption after the considered breakpoint.
Strategy Outline
The idea behind the approach is the following:
Step 1: Estimation of the seasonal component using stl.fit() by
minimizing MAE and removing it;
Step 2: Breakpoints analysis (bp) on y ;
Step 3: Nonparametric trend analysis for analyzing significant
decreases on the segments and selecting the “relevant” break-
ð3Þ
Step 4: RMC is calculated [Eq. (5)] for each breakpoint in bp .
The scheme of the strategy is shown in Fig. 1. In Step 1, the stl.
fit() was chosen because it showed better performance than the
“standard stl” (Cristina et al. 2016). Nevertheless, it is possible to
position. Step 2 uses the function breakpoints() (Zeileis et al. 2002)
with h and breaks estimated by Eqs. (3) and (4), respectively. In the
case of Step 3, a function that includes mk.test() (Pohlert 2020) was
implemented by the authors in R. It considers bp and y as input
and bp as output. Then, to obtain the RMC values using the TS
slopes, another function was build by the authors. The new code
was implemented in R version 4.1.0. The data, models, and code
are available online (Cordeiro et al. 2020). Moreover, a significance
level of α ¼ 5% is considered in the statistical analysis performed.
05021027-3 J. Water Resour. Plann. Manage.
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.
Fig. 1. Scheme of the methodologies proposed using RH2.
Water Consumption Analysis
Dataset
To illustrate the procedure (Fig. 1), eight-time series of billed water
consumption from the same region of the Algarve, Portugal, were
used [available online (Cordeiro et al. 2020)]. The data were pro-
vided by the water company Infraquinta and Loulé Municipality
and are described in the following:
• Two locations: Quinta do Lago Resort, a well-known tourist
destination for holidays; and a small rural village with less than
3,000 inhabitants, mainly retired and older people. Distance be-
tween locations is approximately 26 km;
• Residential Household (RH): four monthly time series of water
consumption, time horizon between January 2011 and December
2016 (nRH ¼ 72), from the village;
• Nonresidential Household (NRH): three hourly time series of
water consumption, time horizon between November 2013 and
December 2018, from Quinta do Lago. However, the water
companies are only concerned with the monthly levels of water
consumption, so the hourly data were aggregated into monthly
by taking the average of each month (nNRH ¼ 62);
• Hotel: monthly time series of water usage with a time horizon
from July 2006 until December 2016 (nHotel ¼ 120);
• Smart meters collected the data of NRH and RH and Hotel by
nonsmart meters.
Fig. 2 shows water consumption variability in households and
hotels. Most of the time series shows a seasonal pattern associated
with the year’s season, increasing in the summer and decreasing
during the winter.
The hotel presents a strong seasonal pattern and a decreasing
trend until 2014 that increases after mid-2014 (due to the water
© ASCE 05021027-4
J. Water Resour. Plann. Manage., 2022, 148(2): 05021027
meter replacement in July 2014). There is also a clear seasonal ef-
fect in RH1, RH2, RH3, and NRH3. Concerning the trend of water
consumption, RH2 and RH3 showed a downward. This pattern is
also observed in smaller periods in the other time series. Moreover,
the magnitude of volumes consumed is significantly different be-
tween the households and the hotel. In households (RH and NRH
type), the water consumption range from 0 to 120 m3 , while at the
hotel, the consumption is higher, varying from 0 to 3,000 m3 .
Time Series Decomposition
The empirical analysis starts by deciding between a robust approach
of STL or not. Outliers were detected by inspecting the weights as-
sociated with the residuals as in Cleveland et al. (1990). Next, for
each time series, the procedure stl.fit (Cristina et al. 2016; Cordeiro
2016) was applied and an object of class stl with components T t , St ,
and I t , as in Eq. (1), is returned. Measurements of error is an alter-
native way of conveying information about the model adjustment,
and for that, the MAE was chosen. This accuracy measure has al-
ready been used in the context of water planning and management,
e.g., Kandiah et al. (2016). The decomposition plots are shown
in Fig. 3.
Decreasing Changes
In this subsection, the objective is to detect periods of changes in
water usage patterns. The detection of breakpoints, which are points
in time k at which statistical patterns change, are estimated on the
seasonally adjusted water consumption, as in Eq. (2). The procedure
tests for structural changes in linear regression models, considering
the maximum number of breaks and the minimum length between
breaks estimated using Eqs. (3) and (4), respectively. As a result,
J. Water Resour. Plann. Manage.
 RH1 RH2
100 20

75 15
m3

m3
50 10

25 5

0 0
2012 2014 2016 2012 2014 2016
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.

RH3 RH4
25 15

20
10
15
m3

m3
10
5
5

0 0
2012 2014 2016 2012 2014 2016

NRH1 NRH2
50 50

40 40

30 30
m3

m3

20 20

10 10

0 0
2014 2015 2016 2017 2018 2019 2014 2015 2016 2017 2018 2019

NRH3 Hotel

120 3000

80 2000
m3

m3

40 1000

0 0
2014 2015 2016 2017 2018 2019 2006 2008 2010 2012 2014 2016

Fig. 2. Times series of water consumption.

breaksRH ¼ 3 and hRH ¼ 0.17, breaksNRH ¼ 3 and hNRH ¼ 0.19, statistically significant (i.e., pvalue < 0.05), the breakpoints are
and breaksHotel ¼ 5 and hHotel ¼ 0.1 were obtained. After ob- considered “relevant” and included in bp , and its decreasing
taining the breakpoints (bp, in Table 1), the MK test for analyzing rate (Sen’s slope) is obtained (see Table 2). Fig. 4 presents the sea-
trends was applied to each segment. If a decreasing segment is sonally adjusted time series, breakpoints (bp), and relevant

© ASCE 05021027-5 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2022, 148(2): 05021027

 RH1 RH2
100
75 15

data

data
50 10
25
5
0
15.0
30
25 trend 12.5

trend
20 10.0
15 7.5
10 5.0
30
seasonal

seasonal
20 3
10
0 0
−10 −3
−20
60
remainder

remainder
40 5
20
0 0
−20
2011 2012 2013 2014 2015 2016 2017 2011 2012 2013 2014 2015 2016 2017
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.

RH3 RH4
25 15
20
10
data

data
15
10 5
5 0
10 7
9 6
trend

trend
8 5
7 4
6
6
seasonal

seasonal
4 2
2
0 0
−2
−4 −2
remainder

10 4 remainder
5 0
0 −4

2011 2012 2013 2014 2015 2016 2017 2011 2012 2013 2014 2015 2016 2017

NRH1 NRH2
50
40 40
30
data

data

30
20 20
10 10
0 0
7
6 12
trend

trend

5
4 10
3
2 8

20 20
seasonal

seasonal

10 10
0
0
−10
remainder

remainder

30 20
20 10
10
0 0
−10 −10
2014 2015 2016 2017 2018 2019 2014 2015 2016 2017 2018 2019

NRH3 Hotel
3000
90
2000
data

data

60
30 1000
0 0
16 1600
trend

trend

14 1400
1200
12 1000
seasonal

seasonal

40 400
20 0
0 −400
remainder

remainder

80 1000
500
40 0
0 −500
−40 −1000
2014 2015 2016 2017 2018 2019 2006 2008 2010 2012 2014 2016

Fig. 3. STL decomposition plots.

© ASCE 05021027-6 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2022, 148(2): 05021027

 Table 1. Breakpoints & decreasing (trend) analysis
Theil–Sen’s slopes
Type Breakpoints (bp) Segment 1 Segment 2 Segment 3 Segment 4 Segment 5
RH1 2012 (June) −0.50 (0.0064) −0.19 (<0.0001) — — —
RH2 2012 (April), 2015 (December) 0.01 (0.8926) −0.02 (0.3947) −0.26 (0.0002) — —
RH3 2014 (May) −0.01 (0.0671) −0.06 (0.0011) — — —
RH4 2015 (November) 0.01 (0.4026) −0.19 (<0.0001) — — —
NRH1 2017 (June) −0.07 (0.0099) 0.12 (0.2629) — — —
NRH2 2016 (December), 2017 (November) 0.07 (0.0252) −0.11 (0.8370) −0.46 (0.0005) — —
NRH3 2016 (December), 2017 (November) 0.01 (0.8015) 6.58 (0.0075) −0.69 (0.0044) — —
Hotel 2008 (February), 2013 (July) −6.14 (0.381) −2.59 (0.0006) −84.56 (0.0003) 5.31 (0.5022) 10.88 (0.1501)
— 2014 (July), 2015 (July) — — — — —
Note: Statistically significant trend slp (pvalue) in boldface.
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.

Table 2. RMC results

Type Date of meter replacement Relevant bp TSa (slpbefore , slpafter ) RMCa
RH1b 2008 (May) 2012 (June) (−0.50, −0.19) 0.62
RH2b 2010 (May) 2015 (December) (−0.02, −0.26) −12.00
RH3b 2003 (November) 2014 (May) (−0.01, −0.06) −5.00
RH4 2012 (January) 2015 (November) (0.01, −0.19) −20.00
NRH1 2018 (June) 2017 (June) (−0.07, 0.12) 2.71
NRH2 2018 (April) 2017 (November) (−0.11, −0.46) −3.18
NRH3 2018 (November) 2017 (November) (6.58, −0.69) −1.10
Hotel 2014 (July) 2008 (February), 2013 (July) (−6.14, −2.59), (−2.59, −84.56) 0.58, − 31.65
— — 2014 (July) (−84.56, 5.31) 1.06
Note: Significant decrease in water consumption in boldface.
a
Round numbers to two decimal places.
b
Outside the time horizon analyzed.

breakpoints (bp , blue dashed line). The dashed grey line corre-
sponds to a breakpoint with nonsignificant adjacent segments
(slpbefore and slpafter ), i.e., pvalue > 0.05, or the water con-
sumption is not decreasing. Table 2 contains the information about
the date of the water meter replacement, bp , TS slopes, and
RMC [Eq. (5)].
The residential household RH1 shows one relevant breakpoint
detected in June 2012 (Table 2; Fig. 4). The water meter readings
after the break are smaller than the first segment, and the magnitude
of the trend (TS slope) is less negative (slpbefore ¼ −0.50 versus
slpafter ¼ −0.19), with RMC ¼ 0.62, as seen in Table 2. Nonethe-
less, the trend continues negative after June 2012 with a significant
slope, justifying an alert and further investigation by the water
utility. Regarding the RH2, the global trend is decreasing (Figs. 2
and 3), and the structural break analysis identifies two breakpoints
(Fig. 4). The decrease of the third segment is statistically significant
(Table 1) and so the second breakpoint (December 2015) is rel-
evant (Table 2). The RMC ¼ −12 reveals that the downward trend
(December 2015–December 2016) is more pronounced in magni-
tude compared to the previous segment (April 2012–December
2015). Since the consumption is decreasing, the water company
should keep RH2 under surveillance. For households RH3 and
RH4, only a relevant breakpoint is detected, May 2014 and
November 2015, respectively. Moreover, comparing the TS slopes
before and after the breakpoint, the magnitude of change is consid-
erably higher in water meter RH4 (RMCRH4 ¼ −20 vs RMCRH3 ¼
−5). RH4 exhibits an abrupt change at the beginning of 2012
(Figs. 2 and 4), and based on the records from the water company,
the meter was replaced in January 2012 (13th month). As already
explained in subsection “Detect structural changes,” the breakpoint
analysis requires a minimal segment size (h), which in this case
is 14 months (corresponding to hRH4 × 72 ¼ 0.2 × 72). So, this
breakpoint was out of reach of being detected. However, if data
before 2011 were available, this issue would no longer exist.
Concerning the nonresidential households from the Quinta do
Lago Resort, NRH1 shows one breakpoint in June 2017. According
to Table 1, the first segment is statistically significant, so the break-
point is considered relevant (Table 2). The relative magnitude of
change is RMC ¼ 2.71, revealing an increase in consumption after
the breakpoint, so there is no motive for alarming the water com-
pany. Besides, this water meter was replaced in June 2018. How-
ever, judging by the graph of the water consumption (Figs. 2 and 4),
it seems a hasty decision by the company. In NRH2, two break-
points were detected (Table 1), but only one is relevant (November
2017) with an estimate RMC ¼ −3.18 (Table 2). According to
this table, this meter was already malfunctioning for half a year
when the substitution took place. In NRH3, the procedure detected
two breakpoints, but only November 2017 is relevant (bp ) with
RMC ¼ −1.10. Note that this decline was detected one year before
the meter replacement (Table 2). If a surveillance system integrated
this approach, the water company would have received an alert one
year before.
In the case of the hotel, from the five segments (four break-
points), only two are statistically significant, corresponding to
bp ¼ fFebruary 2008; July 2013; July 2014g, as seen in Table 2
and Fig. 4. The estimated RMC are 0.58, −31.65, and 1.06, respec-
tively (Table 2). The negative value of RMC for the second break-
point and the decrease in consumption shows evidence that the
meter should have been replaced sooner, perhaps between July
2013 and July 2014.

© ASCE 05021027-7 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2022, 148(2): 05021027

 RH1 RH2
100 20

75 15
m3

m3
50 10

25 5

0 0
2012 2014 2016 2012 2014 2016
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.

RH3 RH4
25 15

20
10
15
m3

m3
10 5

5
0
0
2012 2014 2016 2012 2014 2016

NRH1 NRH2
50 50

40 40
30
30
m3

m3

20
20
10
10
0

−10 0
2014 2015 2016 2017 2018 2019 2014 2015 2016 2017 2018 2019

NRH3 Hotel

3000
100

2000
m3

m3

50

1000

0
0
2014 2015 2016 2017 2018 2019 2006 2008 2010 2012 2014 2016

Fig. 4. Seasonally adjusted time series and breakpoints.

As observed in Table 2, the results indicate that the company the approach in monitoring water consumption by the consumers,
should continue to monitor all households’ meters, except RH1, such as sudden changes and continuous decrease. Besides, the ap-
NRH1, and the hotel since the corresponding RMC values are pos- proach could detect these changes sooner than the actual replace-
itive. To sum up, the previous results illustrate the effectiveness of ment time and in instances in time not foreseen by the water utility.

© ASCE 05021027-8 J. Water Resour. Plann. Manage.

J. Water Resour. Plann. Manage., 2022, 148(2): 05021027

 Discussion
Water utilities need to be aware of water’s responsible use, a con-
cern emphasized by climate change. One form of the companies to
be alert is to steer the effective water consumption and guarantee an
efficient manutention of its equipment. However, when it comes to
the processing of real metering data, several difficulties arise.
It is well known that mechanical water meters metrology be-
comes more and more inaccurate during their operating life-
time due to the “wear and tear” of the measuring components
(Mutikanga et al. 2011). As already mentioned, there are several
studies [e.g., Arregui et al. (2006, 2013) and Pacheco et al. (2020)]
that analyze the influence of parameters in the water meter met-
rology such as fatigue tests, depositions, incorrect water meter siz-
ing, water consumption patterns, among others. Apparent losses
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.
occur due to inaccuracies in metering, illegal use, meter reading
errors, data handling, and billing errors. These type of losses has
a significant impact on the revenue of the water companies. The
cost for the utility of a cubic meter of real losses is much less than
the price of a cubic meter of apparent losses. This issue is of great
importance because while real losses are easily recognized, appar-
ent losses are not perceived well likely because apparent losses
cannot be directly seen and are mostly related to economic losses.
Arregui et al. (2006) present extensive work regarding this issue,
and one key observation is that the water company should study
the client’s water consumption patterns (private domestic or indus-
trial). Additionally, Oviedo-Ocaña et al. (2020) also suggest that
customer meter renovation is an alternative to reduce apparent
losses, mitigating inaccuracies associated with the aging of these
devices.
So far, the usual mechanisms of analyzing the water consump-
tion are performed on a bigger scale, such as considering all the
consumers/water meters of cities, as in Monedero et al. (2016) and
Hester and Larson (2016). To the best of our knowledge, no similar
approach to the one proposed has been developed to examine water
consumption trends for a single water meter. Its purpose is to con-
tribute to a better understanding of water consumption by providing
useful guidance on the replacement of the meters. It will also pro-
vide better analysis and a new perspective on water usage within
water management.
The case study involved houses with different characteristics
(residential and nonresidential, mainly for holidays), and a hotel,
from the Algarve. The Algarve is a popular summer holiday desti-
nation with a Mediterranean climate, meaning hot summers and
mild winters. This seasonal effect is estimated by stl.fit() and re-
moved. The strategy proceeds as follows: breakpoint analysis, non-
parametric trend analysis, and its magnitude, and a change of water
consumption indicator are calculated. The combination of the meth-
odologies was able to detect statistically significant decreasing
trends associated with points in time (i.e., breakpoints) where these
changes occur and in some cases even earlier than the company’s
decision on the water meter replacement. This is an added value of
the strategy. If the water meter is replaced too late, a significant rev-
enue loss (caused by meter under-registration) may occur. On the
other hand, if detected after the actual replacement means that it
was replaced too soon. Therefore, this approach could become a
useful tool to use by the water companies because it provides a bet-
ter understanding of water consumption patterns and makes a con-
scious decision regarding the water meter maintenance (to replace
or not). The proposal for a unitless indicator of the magnitude of
change allows companies to analyze water meters from different
infrastructures, with different water volumes such as a household
or a hotel. The decision to replace the water meter is of the water
company’s responsibility, which must weigh the cost-benefit of
© ASCE
J. Water Resour. Plann. Manage., 2022, 148(2): 05021027
changing one or more meters. This decision analysis could gather
the results already achieved with an econometric analysis of the op-
timum level of apparent losses, as described by Arregui et al. (2018),
or else a financial analysis as the one performed by Pacheco et al.
(2020), to increase the confidence on the timing of the water meter
replacement.
Conclusion and Future Work
Nowadays, water utilities have an essential role in managing water
and wastewater services. One of the problems they face is nonreve-
nue water due to apparent losses. The procedure developed gener-
ates an alert based on decreasing water consumption patterns to
accomplish the water company’s challenge, contributing to water
meter performance assessment. It was developed and implemented
using R language and is supported by reliable statistical methods.
Besides, a new quantitative measure, designated as RMC, is pro-
posed. The objective is to obtain a magnitude of the water usage
change, which can help decide on the meter’s replacement. The
water company must also investigate the decline of the consump-
tion (RMC < 0) since it might be related to a possible anomaly in
the meter. If so, it is the ideal time to replace it. The inclusion of this
new approach in a system that automatically analyzes the data and
updates the information could be the solution sought by water com-
panies to mitigate the impact of apparent water losses.
The proposed approach was successfully applied to eight-
time series of billed water consumption from two locations in the
Algarve, Portugal, representing three different consumption pat-
terns. The data were collected by meters (Residential Households
and the hotel) and smart meters (nonresidential households). A
complement is the visual inspection of the results and their inter-
pretation. This is also a valuable tool worth exploring to understand
and derive more insights into the consumption pattern. Moreover,
the procedure was able to identify points in time that coincide with
the date of replacement of the meters. In some cases, they were
detected earlier than the company’s decision on the water meter
replacement. As so, an alert would have been useful for the com-
pany to minimize revenue loss.
In future work, it is planned a simulation study on water con-
sumption to evaluate the approach’s performance based on quanti-
tative metrics. Furthermore, an association of the decreasing trend
with socio-economic aspects could explain the change in consum-
ers’ behavioral patterns. The latter was not performed because it
was out of the paper scope. Future research on this topic should
be considered to refine and explain the changes in water usage. This
strategy’s extension to identify increasing patterns is straightfor-
ward, translating into a viable approach to manage NRW, alert con-
serving water resources, and encourage responsible water use.
Data Availability Statement
The data, models, and the code used that support this study’s find-
ings are available in a repository online: https://github.com/Clara
Cordeiro/StrategyWaterMeterPerformance.
Reproducible Results
Soraia Pereira (Universidade de Lisboa) downloaded all the ma-
terials, installed, ran the models using the data and functions in
“Workspace_data_functions.RData”, and reproduced the results in
Tables 1 and 2.
05021027-9 J. Water Resour. Plann. Manage.
 Acknowledgments
The authors would like to thank the editors, two referees, and a
reproducibility reviewer for their constructive comments and sug-
gestions, which greatly improved this paper and the information
available in the GitHub repository. The idea behind this work began
at the 140th European Study Group with Industry (ESGI140),
held in Portugal in June 2018, and it was motivated by the chal-
lenge proposed by Infraquinta (https://www.infraquinta.pt/en/) un-
der the theme “Evaluating Water Meters.” The authors are grateful
to the water company Infraquinta and Loulé Municipality (http://
www.cm-loule.pt/pt/Default.aspx) for providing the water con-
sumption data and inside technical guidance. Clara Cordeiro is par-
tially financed by national funds through FCT—Fundação para a
Ciência e Tecnologia under the project UIDB/00006/2020. Ana
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.
Borges work has been supported by national funds through
FCT—Fundação para a Ciência e Tecnologia through project
UIDB/04728/2020. M. Rosário Ramos was partially supported
by National Funding from FCT—Fundação para a Ciência e Tec-
nologia under the project UIDB/04561/2020.
References
Arregui, F., J. Soriano, J. Garcia-Serra, and R. Cobacho. 2013. “Proposal of
a systematic methodology to estimate apparent losses due to water meter
inaccuracies.” Water Sci. Technol. Water Supply 13 (5): 1324–1330.
https://doi.org/10.2166/ws.2013.138.
Arregui, F. J., E. Cabrera, R. Cobacho, and J. Garcia-Serra. 2006. “Reduc-
ing apparent losses caused by meters inaccuracies.” Water Pract. Tech-
nol. 1 (4): wpt2006093. https://doi.org/10.2166/wpt.2006.093.
Arregui, F. J., R. Cobacho, J. Soriano, and R. Jimenez-Redal. 2018. “Cal-
culation proposal for the economic level of apparent losses (ELAL) in
a water supply system.” Water 10 (12): 1809. https://doi.org/10.3390
/w10121809.
Awty-Carroll, K., P. Bunting, A. Hardy, and G. Bell. 2019. “An evaluation
and comparison of four dense time series change detection methods
using simulated data.” Remote Sens. 11 (23): 2779. https://doi.org/10
.3390/rs11232779.
Cleveland, R. B., W. S. Cleveland, J. E. McRae, and I. Terpenning. 1990.
“A STL: A seasonal-trend decomposition procedure based on loess.” J.
Off. Stat. 6 (1): 3–73.
Cordeiro, C. 2016. “stl.fit(): Function developed in cristina et al. (2016).”
Accessed May 21, 2020. https://github.com/ClaraCordeiro/stl.fit.
Cordeiro, C., A. Borges, and M. R. Ramos. 2020. “Script, dataset and
functions used in a strategy to assess water meter performance.” Ac-
cessed December 23, 2020. https://github.com/ClaraCordeiro
/StrategyWaterMeterPerformance.
Criminisi, A., C. Fontanazza, G. Freni, and G. L. Loggia. 2009. “Evaluation
of the apparent losses caused by water meter under-registration in inter-
mittent water supply.” Water Sci. Technol. 60 (9): 2373–2382. https://doi
.org/10.2166/wst.2009.423.
Cristina, S., C. Cordeiro, S. Lavender, P. Costa Goela, J. Icely, and A.
Newton. 2016. “Meris phytoplankton time series products from the SW
Iberian Peninsula (sagres) using seasonal-trend decomposition based on
loess.” Remote Sens. 8 (6): 449. https://doi.org/10.3390/rs8060449.
FEMS. 2021. “Àgua distribuìda/consumida por habitante. Onde se utiliza,
em média, por pessoa, mais e menos água canalizada?” Accessed Feb-
ruary 17, 2020. https://www.pordata.pt/Municipios/%C3%81gua
+distribu%C3%ADda+consumida+por+habitante-484.
Fontanazza, C. M., V. Notaro, V. Puleo, and G. Freni. 2015. “The apparent
losses due to metering errors: A proactive approach to predict losses and
schedule maintenance.” Urban Water J. 12 (3): 229–239. https://doi.org
/10.1080/1573062X.2014.882363.
Fourie, R., A. L. Marnewick, and N. Joseph. 2020. “An empirical analysis
of residential meter degradation in Gauteng Province, South Africa.”
Water SA 46 (4): 645–655.
© ASCE 05021027-10
J. Water Resour. Plann. Manage., 2022, 148(2): 05021027
Gelažanskas, L., and K. A. Gamage. 2015. “Forecasting hot water consump-
tion in residential houses.” Energies 8 (11): 12702–12717. https://doi.org
/10.3390/en81112336.
Gocic, M., and S. Trajkovic. 2013. “Analysis of changes in meteorological
variables using Mann-Kendall and Sen’s slope estimator statistical tests
in Serbia.” Global Planet. Change 100 (Jan): 172–182. https://doi.org
/10.1016/j.gloplacha.2012.10.014.
Helsel, D. R., and R. M. Hirsch. 2002. Vol. 323 of Statistical methods in
water resources. Reston, VA: ASCE.
Hester, C. M., and K. L. Larson. 2016. “Time-series analysis of water de-
mands in three North Carolina cities.” J. Water Resour. Plann. Manage.
142 (8): 05016005. https://doi.org/10.1061/(ASCE)WR.1943-5452
.0000659.
Hyndman, R., G. Athanasopoulos, C. Bergmeir, G. Caceres, L. Chhay, M.
O’Hara-Wild, F. Petropoulos, S. Razbash, E. Wang, and F. Yasmeen.
2021. “Forecast: Forecasting functions for time series and linear mod-
els.” Accessed June 1, 2021. http://pkg.robjhyndman.com/forecast.
Hyndman, R. J., and G. Athanasopoulos. 2018. Forecasting: Principles
and practice. 2nd ed. Melbourne, Australia: OTexts.
Hyndman, R. J., and A. B. Koehler. 2006. “Another look at measures of
forecast accuracy.” Int. J. Forecasting 22 (4): 679–688. https://doi.org
/10.1016/j.ijforecast.2006.03.001.
Kadenge, M. J., V. G. Masanja, and M. J. Mkandawile. 2020. “Optimisa-
tion of water loss management strategies: Multi-criteria decision analy-
sis approaches.” J. Math. Inf. 18 (Feb): 105–119. https://doi.org/10
.22457/jmi.v18a9166.
Kandiah, V. K., E. Z. Berglund, and A. R. Binder. 2016. “Cellular automata
modeling framework for urban water reuse planning and management.”
J. Water Resour. Plann. Manage. 142 (12): 04016054. https://doi.org
/10.1061/(ASCE)WR.1943-5452.0000696.
Kendall, M. 1975. Rank correlation methods. Granville, OH: Charles
Griffin.
Lafare, A. E., D. W. Peach, and A. G. Hughes. 2016. “Use of seasonal trend
decomposition to understand groundwater behaviour in the Permo-
Triassic sandstone aquifer, Eden Valley, UK.” Hydrogeol. J. 24 (1):
141–158. https://doi.org/10.1007/s10040-015-1309-3.
Mann, H. B. 1945. “Nonparametric tests against trend.” Econometrica J.
Econom. Soc. 13 (3): 245–259. https://doi.org/10.2307/1907187.
Miller, C. 2019. “What’s in the box?! Towards explainable machine learn-
ing applied to non-residential building smart meter classification.” En-
ergy Build. 199 (Sep): 523–536. https://doi.org/10.1016/j.enbuild.2019
.07.019.
Miller, C., and F. Meggers. 2017. “Mining electrical meter data to predict
principal building use, performance class, and operations strategy for
hundreds of non-residential buildings.” Energy Build. 156 (Dec):
360–373. https://doi.org/10.1016/j.enbuild.2017.09.056.
Moahloli, A., A. Marnewick, and J. Pretorius. 2019. “Domestic water meter
optimal replacement period to minimize water revenue loss.” Water SA
45 (2): 165–173. https://doi.org/10.4314/wsa.v45i2.02.
Monedero, I., F. Biscarri, J. I. Guerrero, M. Peña, M. Roldán, and C. León.
2016. “Detection of water meter under-registration using statistical al-
gorithms.” J. Water Resour. Plann. Manage. 142 (1): 04015036. https://
doi.org/10.1061/(ASCE)WR.1943-5452.0000562.
Mutikanga, H. E., S. K. Sharma, and K. Vairavamoorthy. 2011. “Assess-
ment of apparent losses in urban water systems.” Water Environ. J.
25 (3): 327–335. https://doi.org/10.1111/j.1747-6593.2010.00225.x.
Ncube, M., and A. Taigbenu. 2019. “Assessment of apparent losses due to
meter inaccuracy—A comparative approach.” Water SA 45 (2): 174–
182. https://doi.org/10.4314/wsa.v45i2.03.
Ohana-Levi, N., S. Munitz, A. Ben-Gal, A. Schwartz, A. Peeters, and Y.
Netzer. 2020. “Multiseasonal grapevine water consumption—Drivers
and forecasting.” Agric. For. Meteorol. 280 (Jan): 107796. https://doi
.org/10.1016/j.agrformet.2019.107796.
Oviedo-Ocaña, E., I. Dominguez, J. Celis, L. Blanco, I. Cotes, S. Ward, and
Z. Kapelan. 2020. “Water-loss management under data scarcity: Case
study in a small municipality in a developing country.” J. Water Resour.
Plann. Manage. 146 (3): 05020001. https://doi.org/10.1061/(ASCE)
WR.1943-5452.0001162.
Pacheco, V. M. F., R. E. Valdés, E. B. Gil, A. N. Manso, and E. Á. Álvarez.
2020. “Techno-economic analysis of residential water meters: A
J. Water Resour. Plann. Manage.
 practical example.” Water Resour. Manage. 34 (Jun): 2471–2484.
https://doi.org/10.1007/s11269-020-02564-x.
Pohlert, T. 2020. “Trend: Non-parametric trend tests and change-point
detection.” Accessed May 21, 2020. https://CRAN.R-project.org
/package=trend.
Quesnel, K. J., and N. K. Ajami. 2017. “Changes in water consump-
tion linked to heavy news media coverage of extreme climatic
events.” Sci. Adv. 3 (10): e1700784. https://doi.org/10.1126/sciadv.170
0784.
R Core Team. 2021. R: A language and environment for statistical com-
puting. Vienna, Austria: R Foundation for Statistical Computing.
Reynaud, A. 2015. “Modelling household water demand in Europe—
Insights from a cross-country econometric analysis of EU-28 countries.”
In Insights from a cross-country econometric analysis of EU, 28.
Luxembourg: Publications Office of the European Union.
Richards, G. L., M. C. Johnson, and S. L. Barfuss. 2010. “Apparent losses
Downloaded from ascelibrary.org by 23.22.233.225 on 11/21/21. Copyright ASCE. For personal use only; all rights reserved.
caused by water meter inaccuracies at ultralow flows.” J. Am. Water
Works Assoc. 102 (5): 123–132. https://doi.org/10.1002/j.1551-8833
.2010.tb10115.x.
© ASCE 05021027-11
J. Water Resour. Plann. Manage., 2022, 148(2): 05021027
Sen, P. K. 1968. “Estimates of the regression coefficient based on Kendall’s
tau.” J. Am. Stat. Assoc. 63 (324): 1379–1389. https://doi.org/10.1080
/01621459.1968.10480934.
Sharma, S., D. A. Swayne, and C. Obimbo. 2016. “Trend analysis and
change point techniques: A survey.” Energy Ecol. Environ. 1 (3):
123–130. https://doi.org/10.1007/s40974-016-0011-1.
Verbesselt, J., R. Hyndman, G. Newnham, and D. Culvenor. 2010.
“Detecting trend and seasonal changes in satellite image time series.”
Remote Sens. Environ. 114 (1): 106–115. https://doi.org/10.1016/j.rse
.2009.08.014.
Willmott, C. J., and K. Matsuura. 2005. “Advantages of the mean absolute
error (MAE) over the root mean square error (RMSE) in assessing aver-
age model performance.” Clim. Res. 30 (1): 79–82. https://doi.org/10
.3354/cr030079.
Zeileis, A., C. Kleiber, W. Krämer, and K. Hornik. 2003. “Testing and
dating of structural changes in practice.” Comput. Stat. Data Anal.
44 (1–2): 109–123. https://doi.org/10.1016/S0167-9473(03)00030-6.
Zeileis, A., F. Leisch, K. Hornik, and C. Kleiber. 2002. “Strucchange: An
R package for testing for structural change in linear regression models.”
J. Stat. Software 7 (2): 1–38. https://doi.org/10.18637/jss.v007.i02.
J. Water Resour. Plann. Manage.