You are on page 1of 17

Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: https://www.tandfonline.com/loi/thsj20

Least squares fitting of the stage–discharge


relationship using smooth models with curvilinear
asymptotes

Youness Mir & François Dubeau

To cite this article: Youness Mir & François Dubeau (2015) Least squares fitting of the
stage–discharge relationship using smooth models with curvilinear asymptotes, Hydrological
Sciences Journal, 60:10, 1797-1812, DOI: 10.1080/02626667.2014.935779

To link to this article: https://doi.org/10.1080/02626667.2014.935779

Published online: 27 Aug 2015.

Submit your article to this journal

Article views: 862

View related articles

View Crossmark data

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
Hydrological Sciences Journal – Journal des Sciences Hydrologiques, 60 (10) 2015 1797
http://dx.doi.org/10.1080/02626667.2014.935779

Least squares fitting of the stage–discharge relationship using smooth


models with curvilinear asymptotes
Youness Mir and François Dubeau
Département de mathématiques, Université de Sherbrooke, Sherbrooke, Quebec, Canada
youness.mir@usherbrooke.ca

Received 14 November 2013; accepted 12 May 2014

Editor D. Koutsoyiannis; Associate editor S. Yue

Abstract In this paper we consider the problem of modelling the stage–discharge relationship by curve fitting
using the least squares method. Our basic idea is to present new models which are more flexible and have the
ability to model phenomena with increasing or unchanging carrying capacity. The new models present a
generalization of some sigmoid smooth models commonly used in practice. They are characterized by a curvi-
linear asymptote and may have several inflection points. The use of these models on six real datasets collected
from the US Geological Survey’s website proves their performance and their ability to model hydrological
phenomena.
Key words hydrology; rating curve; smooth models; curvilinear asymptote; least squares method

Modélisation par moindres carrés de la relation hauteur–débit via des modèles non linéaires avec
asymptotes curvilignes
Résumé Dans cet article nous considérons le problème de la modélisation de la relation hauteur–débit par la
méthode des moindres carrés. Nous proposons l'utilisation de quelques modèles non linéaires d’ajustement qui
permettent une approximation globale des courbes possédant un ou plusieurs point d'inflexions avec la présence
d’une asymptote horizontale, croissante linéaire ou curviligne. Les nouveaux modèles présentés sont des
généralisations de certains modèles non linéaires d’ajustement, avec comportement sigmoïde, les plus couram-
ment utilisés dans plusieurs domaines des sciences appliquées. L’utilisation de ces modèles sur six jeux de
données réels collectés à partir du site web des services d’enquêtes géologiques des États-Unis (US Geological
Survey) a montrée l’utilité et la performance de ces nouveaux modèles sur les jeux de données hydrologiques.
Mots clefs hydrologie ; courbe de tarage ; modèles de lissage ; asymptote curviligne ; méthode des moindres carrés

INTRODUCTION Domeneghetti et al. 2012). Several standard methods


have been used to derive and improve this relation-
In hydrology, establishing a reliable relationship
ship, such as power and polynomial regression meth-
between water depth and flow rate (rating curve) in
ods. Even though the power model is the most used
river and stream channels is of prime importance and
in river hydraulics over recent decades (Callede et al.
hence has attracted the attention of researchers over a
2001, Parodi and Ferraris 2004, Petersen-Øverleir
century. Although this relationship exists in any case
2004, 2005, 2006), several authors have shown that
and can be studied interchangeably (Dymond and
this approach often leads to inaccurate estimates of
Christian 1982, Callede et al. 1997, Braca 2008, Di
flows, mainly where the stage–discharge relationship
Baldassarre and Claps 2011), in hydrological and in
does not cover the full range of flows (Westphal et al.
hydraulic studies the accuracy of the stage and dis-
1999, Schmidt and Yen 2001, Wu and Yang 2008). In
charge values predicted is strongly dependent on the
addition to its limitation to the range of data mea-
reliability of this relationship as well as other factors
sured (Di Baldassarre and Claps 2011), this law does
(Dymond and Christian 1982, Shrestha et al. 2007,
not perfectly describe this relationship, especially

© 2015 IAHS
1798 Youness Mir and François Dubeau

when the raw data have nonlinear behaviour, namely, suggest curves having sigmoid behaviour with curvi-
in the presence of one or more inflection points. This linear asymptotes, we adjust the nonlinear models
change in concavity may be caused by the influence presented in Dubeau and Mir (2011) and Dubeau
of certain factors such as the geology watershed and et al. (2011, 2012) and modify their natural horizon-
the morphology of the channel at the gauging station tal asymptote to accommodate data having curvi-
on high flows of rivers. In a similar situation, a more linear or oblique asymptotes. In addition to their
objective way to model this relationship is to use ability to model data with or without inflection
sums or piecewise combinations of power functions points, the form of the resulting new models is
(Herschy 1995). more flexible and could be an alternative to model
Polynomial models of second or third degree are the stage–discharge relationship over the range of
also commonly used (Herschy 1995, 1999, McGinn data studied and to extrapolate this relationship
and Chubak 2002, Sivapragasam and Muttil 2005, beyond the largest observed data. The objectives of
Braca 2008). Despite their ability to model data with the present study are: (1) to indicate the reliability of
inflection points, polynomial functions are not gen- the modified models and their estimates, (2) to eval-
erally constantly increasing functions. To extrapolate uate their performance on stage–discharge datasets,
a stage–discharge relationship beyond the range of and (3) to compare model fits by means of some
data measured, several numerical methods have been known comparison criteria. The numerical results
used, such as the polynomial regression methods reported demonstrate the efficiency of all the pro-
(Braca 2008), Support Vector Machine (SVM) posed models on the six real datasets considered.
(Sivapragasam and Muttil 2005), or Artificial
Neural Networks (ANN) (Jain and Chalisgaonkar
2000, Deka and Chandramouli 2003, Habid and
DATA
Meselhe 2006, Shrestha et al. 2007). The last
approach was first used to model stage–discharge The study presented in this paper uses datasets
relationships exhibiting the hysteresis phenomenon obtained from six gauging stations located in three
(Tawfik et al. 1997, Petersen-Øverleir 2006), which states of the south and southeast region of the United
is not taken into account in the present study. Owing States of America: Texas, Florida, and Alabama.
to the mathematical relationship existing between Some characteristics of these datasets are summar-
water depth and flow rate (Torsten et al. 2002), to ized in Table 1, namely the station names and abbre-
derive this relationship, many authors suggest the use viations, the river basins, the number of gauging
of smooth models with several parameters able to be measurements (K), and the minimum and maximum
optimized to produce a best fit (Torsten et al. 2002, annual flows. For each dataset the rating table and the
Sivapragasam and Muttil 2005). associated rating curve drawn on logarithmic plotting
In Dubeau et al. (2012) we have made a com- paper are also available. The gauging measurements
parative study of several models with sigmoid beha- fðhk ; Qk ÞgKk¼1 , in metres (m) for gauge height hk and
viour on real data. The problem with this approach is in cubic metres per second (m3=s) for discharge Qk ,
that the carrying capacity of all the mentioned models are daily and were taken between October 2011 and
becomes constant with time, which is mainly unrea- October 2012. The rating curve is a piecewise linear
listic in practice, during flood events for example. In continuous function described by a set of points
other words, from a hydrological perspective, as
fðxi ; yi ÞgNi¼0 summarized in the rating table, with N
more and more water is delivered by a particular
much greater than K, such that the abscissas xi are
meteorological condition, river levels become pro-
gressively higher as discharge monotonically equally spaced, x ¼ xiþ1  xi for i ¼ 0; . . . ; N, and
increases. Several authors have reformulated these a piecewise linear interpolating function yð xÞ such
models to accommodate phenomena with a varying that yðxi Þ ¼ yi for i ¼ 0; . . . ; N. This rating curve is
(Meyer and Ausubel 1999, Meyer et al. 1999, used to predict the value Q ^ k ¼ yðhk Þ to approximate
Shepherd and Stojkov 2007) or increasing (Dubeau Qk (k ¼ 1; . . . ; K). The datasets fðhk ; Qk ÞgKk¼1 and
and Mir 2013, 2014) carrying capacity. their corresponding stage–discharge rating curves
In this paper we investigate the use of some new are plotted in Figs 2(a)–7(a) (stage on x-axis vs dis-
mathematical models in order to derive the stage– charge on y-axis). The six datasets used in this study
discharge relationship of some real datasets by means have been selected randomly from the automated US
of the least squares method. As the collected datasets Geological Survey (USGS) database (the real time
Modeling stage-discharge relationships 1799

Table 1 Stream flow gauging stations.


Station name Abbreviation River basin State Gauging (K) Flow (m3/s) Figure

min max

Paint Rock River near Woodville PRRNW Tennessee Alabama 366 0.3398 336.97 2(a)
Chattooga River above Gaylesville CRAG Coosa Alabama 436 3.3697 148.10 3(a)
Withlacoochee River near Pinetta WRNP Withlacoochee Florida 365 2.1238 148.95 4(a)
Ochlockonee River near Bloxham ORNB Ochlockonee Florida 454 1.4158 139.89 5(a)
Pease River near Vernon PRNV Red Texas 366 0 14.583 6(a)
Clear Creek near Sanger CCNS Trinity Texas 466 0 126.58 7(a)

stream flow map). Some daily data are not approved multiply the resulting model equation by an increasing
yet for publication and may be subject to revision. In curvilinear function given by mðt Þ ¼ pt β þ q, with
addition, according to the USGS website, “the stage– p  0, β > 0, and q > 0, to obtain the modified model
discharge ratings (associated to these datasets) are
developed from a graphical analysis of current-meter F ðt; θÞ ¼ mðt Þf ðt Þ (1)
discharge measurements made over a range of stages
and discharges. These relations change over time as
the channel features that control the relation between This function is well defined for t 2 ðmaxf0; T0 g;
stage and discharge vary.” All this information is þ 1Þ. Note that the case β > 1 gives an increasing
collected from the USGS Water Resources website convex asymptote with no inflection point or with an
and is available at: http://waterwatch.usgs.gov. even number of inflection points. The case 0 < β < 1
gives an increasing concave asymptote with an odd
or even number of inflection points, the parity of the
number of inflection points depends on the values of
MODELS the parameters p and β of the function mðtÞ. The
The basic models f ðt Þ, as well as the acceptable values limiting case β ¼ 1 generates an oblique asymptote
of their parameters able to give an increasing curve with one inflection point (Dubeau and Mir 2013).
with one inflection point are listed in Table 2. The case p = 0 retrieves the natural horizontal
Moreover, this table also contains the lower bound asymptote of the basic model f ðt Þ. In all the cases
T0 of their domain of definition. The asymptote of discussed
 β above,
 the asymptote is given by
all the basic models is given by y ¼ a. In order to y ¼ pt þ q fþ1 , where fþ1 ¼ limt!þ1 f ðtÞ ¼ 1
realize our goal of obtaining a smooth function with a (because we have set a ¼ 1). Note that the modified
curvilinear asymptote, we set the multiplicative con- Gompertz model is a limiting case of the modified
stant a equal to 1 in the basic growth function f ðt Þ and Richards(1) and Richards(2) models as mR ! 1,

Table 2 Parameter values of the basic smooth models.


Model Function Parameter vector ζ T0
 
Michaelis–Menten ω0 þ t c a > 0; k > 0; c > 1; 1 > ω0  0 0
fM ðtÞ ¼ a c
k þ tc
Richards(1) fR ðt Þ ¼ að1  bect Þm a > 0; b > 0; c > 0; m > 1 lnðbÞ
c
Richards(2) fR ðt Þ ¼ að1  bect Þm a > 0; bh0; ci0; m < 0 1
Hossfeld at c a > 0; k > 0; c > 1 0
fH ðt Þ ¼ c
k þ tc
ct
Gompertz fG ðtÞ ¼ aebe a > 0; b > 0; c > 0 1
Logistic a a > 0; b > 0; c > 0 1
fL ðt Þ ¼
1 þ bect
 c
Bridge fB ðt Þ ¼ a 1  ebt a > 0; c > 1; b > 0 0
Exponential b=tc
fE ðt Þ ¼ ae a > 0; b > 0; c  1 0
1800 Youness Mir and François Dubeau

F (t), β > 1

m (t),β > 1
2
F (t), β = 1
m (t), β = 1

m (t), 0 < β < 1 F (t), 0 < β <1


1
f(t)

0
0 50 100 150 200 250 300 350 400 450

Fig. 1 Graphic representation of the basic model f ðt Þ with its asymptote y ¼ 1 and the modified model F ðt Þ with its
asymptote y ¼ pt β þ q for 0 < β < 1 (concave asymptote), β ¼ 1 (oblique asymptote) and β > 1 (convex asymptote) and
their inflection points.

bR ! 0 and, mR bR ! bG respectively, and the mod- Z  ¼ Z ðθ Þ ¼ min Z ðθÞ; θ ¼ arg min Z ðθÞ
θ2Θ θ2Θ
ified logistic model is a particular case of the mod-
ified Richards(2) model with the parameter m equal to (3)
 1 and b is replaced by  b. A special case of the
modified Michaelis–Menten (or Morgan) model The MATLAB constrained least squares function
(with ω0 ¼ 0) is the modified Hossfeld model “lsqnonlin” is used to solve these problems. This
(Zeide 1993). The basic bridges model is the well- function uses the bound constrained trust-region-
known Weibull model (Zeide 1993, Ismail et al. reflective algorithm, which is based on the interior-
2003, Khamis et al. 2005, Darmani Kuhi et al. reflective Newton method (Coleman and Li 1994,
2010). The graphs of f and F with 0 < β < 1, β ¼ 1996). On each dataset, and on each model, we
1 and β > 1 for all these models look like the graphs have tested 10 different starting points θ0 ¼
given in Fig. 1. ðp0 ; β0 ; q0 ; ζ 0 Þ randomly chosen in a large approxi-
mate region of the parameter space using continuous
uniform distributions. For initial values which pro-
duce converging sequences, the sequences converge
MATERIAL AND METHODS
to the same optimal solution. The MATLAB function
Datasets fðhk ; Qk ÞgKk¼1 and the eight modified mod- “fsolve” was used to calculate, when it exists, an
els described in Section 3 were used to investigate approximation T  of the x-coordinate of the inflec-
the stage–discharge relationship by using the least tion point ðT  ; F  Þ of each modified model F, and
squares estimator. The parameter vector θ ¼ the approximate ordinate F  is given by
ðp; β; q; ζ Þ of each modified model F ðt; θÞ is esti- F  ¼ F ðT  Þ. More than one inflection point might
mated by minimizing the least squares criteria exist.
To compare the least squares fitting given by
1X K our modified models and the rating curves pro-
Z ðθÞ ¼ ½F ðhk ; θÞ  Qk 2 (2) vided from the USGS Water Resources website,
2 k¼1
we have indicated in Table 3 the sum of squared
errors (SSE) between the data fðhk ; Qk ÞgKk¼1 and
over the feasible set of the vector of parameters  
θ ¼ ðp; β; q; ζ Þ 2 Θ, where p  0; β > 0; q  0, and the corresponding values ^ k ¼ yðhk Þ K
hk ; Q k¼1
the parameter vector ζ as defined for each basic obtained from the rating curve. The SSE are
model in Table 2. We obtain given by the formula
Modeling stage-discharge relationships 1801

Table 3 River name abbreviations and the sum of squared construct a BCI we proceed as follows. We resample
errors between river data and the rating curve. uniformly with replacement the original dataset
River PRRNW CRAG WRNP ORNB PRNV CCNS fðhk ; Qk ÞgKk¼1 to generate a bootstrap sample
    K
SSE 1101.2 12.601 5.2607 154.18 4.6955 24.035 hl ; Ql l¼1 , in which some pairs ðhk ; Qk Þ may be
repeated several times and other pairs would not
appear. We treat the resulting bootstrap sample
    K
1X K  
^ k  Qk 2
hl ; Ql l¼1 as a new dataset and adjust it to the
SSE ¼ Q (4) modified model F ðt; θÞ to obtain a new vector of
2 k¼1
parameter estimates by minimizing the least squares
criterion (2). We repeat this process a large number of
This value will be compared to the optimal value Z 
times B, for instance we have taken B = 1000. Let us
of (2).    B
Comparison and ranking are made between the denote by θb ¼ pb ; βb ; qb ; ζ b b¼1 the set of the
different model fits and are based on the root mean replicate values of parameters θ ¼ ðp; β; q; ζ Þ. To
square error (RMSE) defined by the formula construct a 95% percentile BCI of a given para-
meter, we sorted its corresponding bootstrap repli-
rffiffiffiffiffiffiffiffiffiffiffiffi
Z cate values in ascending order, then the lower 95%
RMSE ¼ (5) percentile confidence limit is the B:α=2 ¼ 25-th
Kn
ranked element and the upper 95% percentile con-
and the Akaike information criteria (AIC) defined by fidence limit is the B:ð1  α=2Þ ¼ 975-th ranked
(Akaike 1974) element, where α ¼ 1  0:95 ¼ 0:05. To obtain the
relative lower and upper bootstrap confidence limits
AIC ¼ 2K ln ðRMSEÞ þ 2n (6) of this parameter we respectively divide these con-
fidence limits by the optimal value of this parameter
These two criteria are related and take into account given by the adjustment of the modified model
the degree of freedom regarding the number of F ðt; θÞ on the original dataset fðhk ; Qk ÞgKk¼1 . To
observations and the number of estimated para- draw the lower and upper BCB for the graph of
meters. A smaller numerical value of RMSE and/ the optimal modified model F ðt; θ Þ, we consider
or AIC criteria indicate a better fit when comparing each abscissa xi of the rating curve described in
models. The goodness of fit was based on these Section 2 and the set of replicate optimal values
two criteria as well as a visual evaluation of    B
F xi ; θb b¼1 to obtain yli and yui , which are
optimal curves given by models fitted on the
respectively the 25-th and the 975-th ranked element
datasets.
of the replicate optimal values. The lower, respec-
To evaluate the reliability of our model fits, we
tively upper, confidence band is the piecewise linear
have computed a relative bootstrap confidence inter-
continuous function passing through the points
val (BCI) (Efron 1979, 1985) of parameters   N   N
θ ¼ ðp; β; q; ζ Þ, and we have drawn a bootstrap con- xi ; yli i¼0 and xi ; yui i¼0 .
fidence band (BCB) of the graph of each optimal Finally, we have also computed the values of the
modified model F ðt; θ Þ on each dataset. The widths adjusted R squared (R2adj ). This statistics is defined by
of the BCI and the BCB give us more information on the formula (Tarald 1985, Cameron and Windmeijer
the variability of the parameters of the modified 1997)
models. In general, the narrower the BCI and/or the
BCB is, the higher the data density and the less the ðK  1ÞZ 
R2adj ¼ 1  (7)
variability and uncertainty in parameter values. To ðK  nÞSST
measure the linear relationship between each pair of
parameters of each modified model, we have calcu- where Z  is given by (3), SST is the sum of
lated their pairwise correlation. A positive correlation squared differences of each observation and the
means that the values of these parameters increase or overall mean, K is the number of gauging measure-
decrease simultaneously. A negative correlation ments, and n is the number of parameters of the
means that, if the value of one of them increases, model. This criterion is frequently used in the
the value of the other decreases, and conversely. To context of linear regression to measure the linear
1802 Youness Mir and François Dubeau

relationship between two variables and hence consid- observe the same results for the modified
ered as a goodness of fit measure. Since our models Richards(2) and Gompertz models for all the datasets
are nonlinear, it is used here for information. except PRNV because mRð2Þ bRð2Þ  bG . On PRNV,
from Table 4, the Richards(1) and Richards(2) models
give the same results because mRð1Þ bRð1Þ  mRð2Þ bRð2Þ .
NUMERICAL RESULTS AND DISCUSSION On ORNB and CCNS, all the models have no points
of inflection and the adjustment remain convex over
Due to bad least squares modelling of the bridge and
all the range of data. On PRRNW, CRAG, WRNP,
logistic models on the datasets studied here, numer-
and PRNV, all the modified models have two inflec-
ical results for these two models are not included.
tion points. Besides WRNP, on each dataset where
Moreover, for all the datasets, numerical results show
the inflection points occur, they occur in small inter-
that the best estimate ω0 of the Michaelis–Menten
vals and are close together. The presence, the position
model is close to zero, the optimal solutions θ and
and the number of these inflection points in the
the optimal values of the Hossfeld (ω0 ¼ 0) and
stage–discharge relationship may be explained by
Michaelis–Menten (ω0 Þ0) models are close together.
the influence of a multivariate set of variables, as
As these two models gave the same results on all the
the channel’s slope changes for example, which
data studied, we have not added the numerical results
implies a flux change (e.g. linear flux to turbidity).
of the adjustment of the Michaelis–Menten model.
Analysis of the relative confidence limits of each
We have included in Tables 4, 6, 8, 10, 12, and
parameter shows that the confidence interval of the
14 the optimal solutions θ ¼ ðp ; β ; q ; ζ  Þ, the
exponential model is comparatively narrow in rela-
optimal values Z  ¼ Z ðθ Þ, an approximation
tion to the confidence intervals of the other parameter
ðT  ; F  ) of its corresponding inflection points
models. In addition, on WRNP and CCNS datasets,
ðT  ; F  ), the lower and upper relative confidence confidence intervals of the parameter q are very wide.
limits of each parameter, and the pairwise correlation For these two datasets, we can conclude that para-
coefficients among the parameter estimates. The meter q does not seem to be important and we could
values and the corresponding ranking of the compar- set it equal to 0 for the five models. We also observe
ison criteria of the remaining models on each dataset that, when the Richards(2) model and its limiting case
are summarized in Tables 5, 7, 9, 11, 13, and 15. the Gompertz model give the same results, the con-
From these tables, we observe that all the fidence interval of the generalized model is narrow,
remaining modified models fit the six datasets very which means that parameter values of the generalized
well. Except for the WRNP and CCNS datasets model are more precise and present less variability
where the Hossfeld model gives an optimal value around the optimal parameters than the limiting case
Z far greater than that given by the other models, model. Similarly, confidence bands of optimal mod-
the difference between the optimal values Z of all els are narrow for the six datasets and increase mono-
the modified model adjustments on PRRNW, CRAG, tonically with stage height. However, for ORNB,
ORNB, and PRNV are rather small. Furthermore, on PRNV, and, CCNS, this increase in width becomes
these four rivers, all the modified models appear to important from the largest observed data. In fact, for
be an improvement over their corresponding rating reasonably high water levels, we have less informa-

curves obtained from the USGS Water Resources tion on the asymptote ptβ þ q þ1 , and as conse-
website, because the maximum of the optimal values quence more uncertainty around the optimal curve.
Z given by the modified models on each of these In addition, analysis of the correlation matrices
datasets is smaller than the sum of squared error SSE shows that the parameter p is negatively correlated
regrouped in Table 3. On WRNP, the Richards(1) and with parameters β and q. These latter two parameters
exponential models have produced somewhat smaller are positively correlated. On the other hand, it is also
optimal values Z than the SSE produced by the observed that the parameter c is negatively correlated
rating curve. On CCNS, except for the Hossfeld with the parameters β and q, and positively correlated
model, where the optimal value Z is significantly with the parameter p. For the remaining parameters
high, the remaining models gave slightly smaller of the modified models the sign of the pairwise
optimal values compared to the SSE. A visual com- correlation changes from one dataset to another.
parison of the graphs of optimal curves and rating Based on comparison and ranking criteria
curves of datasets corroborates this fact. We also regrouped in Tables 5, 7, 9, 11, 13, and 15, as well
Modeling stage-discharge relationships 1803

Table 4 Paint Rock River near Woodville: Optimal parameters, the relative confidence interval of parameters, approxima-
tion of the inflection points, and the correlation matrix of each modified model.
Parameter Estimate Relative confidence interval Correlation matrix

Richards(1) model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 2.2842e-3 (0.49039, 1.5087)


β 6.8344 (0.96698, 1.0871) −0.74384
q 125.19 (0.70152, 1.7425) −0.062659 0.54203
b 1.1610 (0.82138, 1.0038) −0.021402 −0.024734 0.040292
c 0.38758 (0.51897, 1.6454) 0.17670 −0.51154 −0.74626 −0.43597
m 2.0285 (0.80880, 1.9248) 0.094874 −0.27051 −0.45023 −0.85282 0.80953
2 inflection points (2.0847, 28.921) (2.8021, 46.331)

Richards (2)
model: FR ðt; p; β; q; b; c; mÞ p
β q b c

p 2.9265e-3 (0.65472, 1.3623)


β 6.5976 (0.97368, 1.0368) −0.791996
q 71.955 (0.94932, 1.0997) −0.69915 0.78582
b −5.6048e-3 (0.94490, 1.1549) 0.030986 −0.060951 0.01431
c 0.97779 (0.90068, 1.0293) 0.54962 −0.6077 −0.93301 −0.064021
m −1244.4 (0.89689, 1.0129) −0.015963 0.020287 0.044804 −0.39011 −0.054849
2 inflection points (2.0727, 28.855) (2.7972, 47.416)

Hossfeld model: FH ðt; p; β; q; k; cÞ p β q k

p 2.8650e-3 (0.53360, 1.4288)


β 6.6504 (0.97199, 1.0558) −0.84681
q 87.234 (0.93566, 1.2351) −0.55605 0.78915
k 2.6974 (0.95921, 1.1688) −0.50565 0.72715 0.99197
c 2.7446 (0.89866, 1.0024) 0.39861 −0.54283 −0.88566 −0.92529
2 inflection points (2.1603, 30.893) (2.7284, 45.454)

Gompertz model: FG ðt; p; β; q; b; cÞ p β q b

p 2.9377e-3 (0.66188, 1.3437)


β 6.5954 (0.97514, 1.0373) −0.72161
q 71.927 (0.94175, 1.1035) −0.70192 0.7993
b 6.9648 (0.88719, 1.1932) 0.25812 −0.32849 −0.71697
c 0.97753 (0.88984, 1.0254) 0.50443 −0.58924 −0.92471 0.91841
2 inflection points (2.0717, 28.831) (2.7966, 47.395)

Exponential model: FE ðt; p; β; q; b; cÞ p β q b

p 6.6379e-3 (0.92024, 1.0736)


β 6.7426 (0.99336, 1.0072) −0.58753
q 335.2 (0.97225, 1.0145) −0.33187 0.40437
b 4.1461 (0.99287, 1.0032) −0.3038 0.38251 0.99367
c 0.70993 (0.99528, 1.0060) 0.32592 −0.39908 −0.9783 −0.96793
2 inflection points (2.1821, 29.452) (2.8022, 46.901)

Table 5 Paint Rock River near Woodville: Optimal values, and ranking criteria values for the models.
Model Z Rank Adjusted R2 Rank RMSE Rank AIC Rank
(1)
Richards 615.93 1 0.99706 1 2.0900 2 436.62 2
Richards(2) 663.75 4 0.99682 4 2.1697 5 458.16 5
Hossfeld 645.49 3 0.99692 3 2.1358 3 447.1 3
Gompertz 663.62 5 0.99683 5 2.1656 4 455.08 4
Exponential 615.96 2 0.99706 2 2.0864 1 433.61 1
Rating curve 1101.2 6
1804 Youness Mir and François Dubeau

Table 6 Chattooga River above Gaylesville: Optimal parameters, the relative confidence interval of parameters, approx-
imation of the inflection points, and the correlation matrix of each modified model.
Parameter Estimate Relative confidence interval Correlation matrix

Richards(1) model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 2.0269 (0.66593, 1.5043)


β 2.5340 (0.91454, 1.0880) −0.76989
q 72.225 (0.86725, 1.1182) −0.86044 0.75053
b 3.6834e-2 (0.66844, 1.5602) 0.36328 −0.01634 −0.010517
c 1.3303 (0.96463, 1.0347) 0.61956 −0.80907 −0.89915 −0.37678
m 403.33 (0.10819, 1.4850) −0.12773 0.057141 0.05898 −0.36453 0.064758
2 inflection points (2.5521, 57.092) (3.275, 93.554)

Richards (2)
model: FR ðt; p; β; q; b; c; mÞ p
β q b c

p 1.4759 (0.77745, 1.2531)


β 2.7075 (0.95206, 1.0497) −0.68982
q 78.287 (0.94446, 1.0699) −0.95752 0.83832
b −3.8475e-3 (0.9692, 1.0521) −0.04392 0.021115 0.063678
c 1.3048 (0.97710, 1.0188) 0.84892 −0.81448 −0.95746 −0.093116
m −3838.4 (0.95273, 1.0340) −0.048444 0.082595 0.023879 −0.75531 0.022054
2 inflection points (2.5328, 56.157) (3.2686, 93.233)

Hossfeld model: FH ðt; p; β; q; k; cÞ p β q k

p 13.392 (0.67437, 1.3395)


β 1.5939 (0.90682, 1.1289) −0.88943
q 14.335 (0.29820, 1.8871) −0.98335 0.88008
k 1.9216 (0.97078, 1.0356) −0.87701 0.86956 0.94651
c 4.2971 (0.95592, 1.0395) 0.96956 −0.96047 −0.92913 0.77875
2 inflection points (2.5625, 57.616) (3.0654, 83.021)

Gompertz model: FG ðt; p; β; q; b; cÞ p β q b

p 1.4701 (0.33609, 2.3109)


β 2.7096 (0.83719, 1.2237) −0.86411
q 78.383 (0.77285, 1.1971) −0.74067 0.66941
b 14.753 (0.96112, 1.0433) 0.73347 −0.84706 −0.88559
c 1.3041 (0.94491, 1.0672) 0.92748 −0.93606 −0.98924 0.92805
2 inflection points (2.5329, 56.163) (3.2685, 93.234)

Exponential model: FE ðt; p; β; q; b; cÞ p β q b

p 1.4769e-9 (0.77825, 1.2097)


β 15.924 (0.99037, 1.0117) −0.89234
q 409.53 (0.99402, 1.0037) −0.8279 0.75511
b 5.9110 (0.99914, 1.0005) 0.12376 −0.24785 0.98466
c 1.1707 (0.99876, 1.0016) 0.12462 −0.26253 −0.98681 −0.94674
2 inflection points (2.696, 64.328) (3.6674, 113.01)

Table 7 Chattooga River above Gaylesville: Optimal values and ranking criteria values for the models.
Model Z Rank Adjusted R2 Rank RMSE Rank AIC Rank

Richards(1) 9.6133 3 0.9998 3 0.31481 3 −450.31 4


Richards(2) 9.5654 2 0.9998 2 0.31403 2 −451.31 2
Hossfeld 9.751 4 0.99979 4 0.31624 4 −450.5 3
Gompertz 9.5659 1 0.9998 1 0.31323 1 −454.33 1
Exponential 11.894 5 0.99975 5 0.34928 5 −410.76 5
Rating curve 12.601 6
Modeling stage-discharge relationships 1805

Table 8 Withlacoochee River near Pinetta: Optimal parameters, the relative confidence interval of parameters, approxima-
tion of the inflection points, and the correlation matrix of each modified model.
Parameter Estimate Relative confidence interval Correlation matrix

Richards(1) model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 11.938 (0.87830, 1.0818)


β 1.6757 (0.96987, 1.0817) −0.88932
q 6.2762e-5 (0.85961, 3.4060e+6) −0.75831 0.86992
b 7.334 (0.57421, 1.1095) 0.87099 −0.82188 −0.84479
c 1.7162 (0.64121, 1.0578) 0.88641 −0.90722 −0.96019 0.70476
m 7.0723 (0.65867, 1.5059) 0.28023 −0.35415 −0.40773 −0.10404 0.61283
2 inflection points (3.1187, 62.527) (4.0666, 119.34)

Richards (2)
model: FR ðt; p; β; q; b; c; mÞ p
β q b c

p 9.9591 (0.80647, 1.0042)


β 1.7899 (0.99852, 1.0573) −0.88711
q 2.4291e-5 (0.73452, 3.6376e+5) −0.78952 0.78987
b 4.4241e-3 (0.95836, 1.0529) −0.13700 0.14738 0.15506
c 1.9834 (0.98009, 1.0029) 0.89532 −0.9094 −0.94628 −0.18582
m −21603 (0.94676, 1.0421) −0.094843 0.10547 0.11973 −0.90780 −0.17204
2 inflection points (3.0236, 56.901) (3.7689, 101.40)

Hossfeld model: FH ðt; p; β; q; k; cÞ p β q k

p 7.7283 (0.69523, 1.0434)


β 1.9574 (0.98479, 1.1030) −0.89597
q 1.5168e-4 (0.021927,1.4708 e + 6) −0.87374 0.76208
k 2.4147 (0.99481, 1.0189) −0.76873 0.73904 0.89113
c 7.6848 (0.98091, 1.0242) 0.74575 −0.76005 −0.62627 0.27403
2 inflection points (2.794, 43.55) (3.4195, 80.216)

Gompertz model: FG ðt; p; β; q; b; cÞ p β q b

p 9.9632 (0.71895, 1.0321)


β 1.7897 (0.98872, 1.3148) −0.87515
q 2.0161e-5 (0.6914, 4.166e+6) −0.78937 0.78812
b 95.510 (0.85423, 1.0529) 0.70934 −0.73961 −0.78418
c 1.9831 (0.89993, 1.0141) 0.91936 −0.92391 −0.95814 0.92558
2 inflection points (3.0238, 56.916) (3.7695, 101.45)

Exponential model: FE ðt; p; β; q; b; cÞ p β q b

p 20.348 (0.81893, 1.1731)


β 1.3659 (0.94909, 1.0648) −0.68499
q 10.671 (3.0007e−4, 2.0589) −0.75235 0.73291
b 25.231 (0.98734, 1.0105) −0.37191 0.37285 0.10174
c 3.3823 (0.98532, 1.0134) 0.61540 −0.57672 −0.81974 0.46318
2 inflection points (3.2187, 68.493) (5.2056, 185.82)

Table 9 Withlacoochee River near Pinetta: Optimal values and ranking criteria values for the models.
Model Z Rank Adjusted R2 Rank RMSE Rank AIC Rank

Richards (1)
5.1732 2 0.99991 2 0.24526 2 −488.33 2
Richards(2) 5.8495 5 0.99989 4 0.26080 4 −466.46 4
Hossfeld 15.307 6 0.99972 5 0.42066 5 −298.27 5
Gompertz 5.8492 4 0.9999 3 0.26004 3 −469.5 3
Exponential 4.5974 1 0.99992 1 0.23054 1 −512.37 1
Rating curve 5.2607 3
1806 Youness Mir and François Dubeau

Table 10 Ochlockonee River near Bloxham FL: Optimal parameters, the relative confidence interval of parameters, and the
correlation matrix of each modified model.
Parameter Estimate Relative confidence interval Correlation matrix

Richards(1) model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 2.4677e-3 (0.67514, 2.0288)


β 7.1846 (0.85771, 1.0635) −0.51833
q 575.22 (0.64959, 4.2732) −0.28486 0.52872
b 1.0714 (0.95326, 1.1130) −0.079334 −0.28594 −0.35357
c 0.12497 (0.80504, 1.1076) 0.80552 −0.58828 −0.49793 0.083545
m 1.6756 (0.90985, 1.1508) 0.58862 −0.30146 −0.2279 −0.55631 0.70813
No inflection point

Richards(2) model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 3.8078e-1 (0.47756, 2.1085)


β 3.6324 (0.87602, 1.1287) −0.53243
q 58.719 (0.81811, 1.1403) −0.43144 0.62235
b −4.5925e-4 (0.98369, 1.0352) −0.48006 0.40536 0.61752
c 1.1746 (0.94516, 1.0995) 0.87073 −0.81317 −0.96259 −0.76840
m −18210 (0.99083, 1.0137) −0.53475 0.48360 0.66755 0.62294 −0.77465
No inflection point

Hossfeld model: FH ðt; p; β; q; k; cÞ p β q k

p 3.1219e-1 (0.19503, 2.7012)


β 3.7674 (0.84199, 1.2604) −0.67461
q 69.783 (0.73575, 1.4057) −0.87089 0.74126
k 2.3171 (0.87883, 1.1756) −0.85099 0.91013 0.99527
c 3.1797 (0.93462, 1.0532) 0.67644 −0.71109 −0.87286 −0.90953
No inflection point

Gompertz model: FG ðt; p; β; q; b; cÞ p β q b

p 3.8087e-1 (0.28840, 2.3633)


β 3.6322 (0.86649, 1.2086) −0.80841
q 58.718 (0.76253, 1.2906) −0.72497 0.74465
b 8.3621 (0.91907, 1.0981) 0.67502 −0.61771 −0.79894
c 1.1746 (0.87624, 1.1547) 0.88804 −0.84395 −0.96429 0.91750
No inflection point

Exponential model: FE ðt; p; β; q; b; cÞ p β q b

p 7.7786e-2 (0.87538, 1.1596)


β 5.3402 (0.98061, 1.0186) −0.79005
q 685.75 (0.98748, 1.0191) −0.64695 0.64373
b 5.0442 (0.99731, 1.0036) −0.63462 0.62978 0.99246
c 0.66134 (0.99579, 1.0031) 0.59054 −0.59937 −0.93391 −0.92396
No inflection point

Table 11 Ochlockonee River near Bloxham: Optimal values and ranking criteria values for the models.
Model Z Rank Adjusted R2 Rank RMSE Rank AIC Rank

Richards(1) 139.87 1 0.99792 1 1.1977 1 84.533 1


Richards(2) 150.62 5 0.99776 5 1.2429 5 99.416 5
Hossfeld 150.32 3 0.99778 3 1.2385 3 95.991 3
Gompertz 150.62 4 0.99777 4 1.2397 4 96.386 4
Exponential 142.92 2 0.99789 2 1.2076 2 85.84 2
Rating curve 154.18 6
Modeling stage-discharge relationships 1807

Table 12 Pease River near Vernon, TX: Optimal parameters, the relative confidence interval of parameters, approximation
of the inflection points, and the correlation matrix of each modified model.
Parameter Estimate Relative confidence interval Correlation matrix

Richards(1) model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 3.4017e-13 (0.37248, 1.4573)


β 23.709 (0.96831, 1.0283) −0.61254
q 14.221 (0.96145, 1.1367) −0.24822 0.88252
b 7.2100 (0.81083, 1.0773) 0.093829 0.036336 −0.79874
c 4.7512 (0.88327, 1.0121) 0.012039 −0.56807 −0.85871 0.98266
m 1056.2 (0.90395, 1.0022) 0.03577 −0.069809 −0.37665 0.46866 0.47644
2 inflection points (1.9670, 5.5972) (2.2210, 10.103)

Richardsð2Þ model: FR ðt; p; β; q; b; c; mÞ p


β q b c

p 9.3724e-9 (0.70857, 1.4578)


β 23.211 (0.96707, 1.0264) −0.56359
q 14.369 (0.98922, 1.0036) −0.83851 0.42039
b −0.37301 (0.99965, 1.0031) 0.0053577 −0.29192 0.90951
c 4.5796 (0.99988, 1.0012) 0.020384 −0.33554 −0.91586 −0.99007
m −19931 (0.99962, 1.0031) −0.014434 −0.38984 0.89987 0.97126 −0.99301
2 inflection points (1.9565, 5.5511) (2.1514, 10.051)

Hossfeld model: FH ðt; p; β; q; k; cÞ p β q k

p 3.1267e-1 (0.0033356, 1.0902)


β 4.5511 (0.97269, 2.3254) −0.32257
q 1.4073e-4 (0.12266, 61261) −0.4091 0.80583
k 1.8184 (0.9976, 1.0528) 0.031589 0.64958 0.84465
c 22.595 (0.84704, 1.0220) 0.30906 −0.11959 −0.2198 −0.59728
2 inflection points (1.9046, 5.3443) (2.0524, 10.2434)

Gompertz model: FG ðt; p; β; q; b; cÞ p


β q b

p 7.8718e-10 (0.092279, 5.4440)


β 26.061 (0.7111, 1.1062) −0.29028
q 14.731 (0.9614, 1.0291) −0.22956 0.33651
b 6744.6 (0.76185, 1.0074) −0.066149 0.24053 −0.39568
c 4.5178 (0.96763, 1.0029) 0.041065 −0.18824 −0.51800 0.98237
2 inflection points (1.9586, 5.604) (2.1586, 10.222)

Exponential model: FE ðt; p; β; q; b; cÞ p β q b

p 1.2548e-11 (0.86554, 1.1253)


β 30.796 (0.99508, 1.0044) −0.82515
q 17.683 (0.99351, 1.0051) −0.26015 0.34235
b 168.79 (0.99982, 1.0006) 0.12238 0.18353 −0.70431
c 7.4182 (0.99991, 1.0005) 0.017865 −0.10905 −0.81162 0.94564
2 inflection points (1.9670, 5.795) (2.1659, 10.396)

Table 13 Pease River near Vernon: Optimal values and ranking criteria values for the models.
Model Z Rank Adjusted R2 Rank RMSE Rank AIC Rank

Richardsð1Þ 2.4809 1 0.98977 1 0.19387 3 −441.48 3


Richardsð2Þ 2.4818 3 0.98976 3 0.19391 4 −440.73 4
Hossfeld 2.5585 5 0.98935 5 0.19615 5 −439.57 5
Gompertz 2.4814 2 0.98976 2 0.19317 1 −443.80 1
Exponential 2.4821 4 0.98967 4 0.19322 2 −443.76 2
Rating curve 4.6955 6
1808 Youness Mir and François Dubeau

Table 14 Clear Creek near Sanger, TX: Optimal parameters, the relative confidence interval of parameters, and the
correlation matrix of each modified model.
Parameter Estimate Relative confidence interval Correlation matrix

Richardsð1Þ model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 3.6576 (0.13488, 1.0411)


β 2.0441 (0.98654, 1.4632) −0.95629
q 6.6064e-3 (0.97067, 7.6685e+3) −0.97629 0.97158
b 8.7677e-2 (0.54991, 1.29478) 0.34287 −0.39346 −0.4382
c 1.8122 (0.75941, 1.28409) 0.89032 −0.86793 −0.94948 0.53797
m 121.89 (0.88149, 1.7053) 0.24924 −0.17139 −0.22099 −0.63797 0.23444
No inflection point

Richardsð2Þ model: FR ðt; p; β; q; b; c; mÞ p β q b c

p 3.4904 (0.88872, 1.0925)


β 2.0705 (0.90342, 1.0264) −0.98722
q 3.5808e-5 (0.095178, 7.7910e+5) −0.91005 0.84271
b −6.5360e-3 (0.99490, 1.06735) −0.43614 0.41191 0.47623
c 1.8812 (0.98714, 1.0065) 0.54918 −0.45141 −0.79652 −0.45151
m −19132 (0.98925, 1.0024) 0.16278 −0.20462 0.020856 0.2019 −0.55886
No inflection point

Hossfeld model: FH ðt; p; β; q; k; cÞ p β q k

p 2.7555 (0.17737, 1.2922)


β 2.208 (0.99782, 1.3968) −0.87360
q 1.5485e-6 (4.114e-3,4.6048e+7) −0.92263 0.94733
k 2.7034 (0.61569, 1.0911) −0.73632 0.77506 0.91140
c 8.3522 (0.90894, 1.0148) 0.64374 −0.51304 −0.41993 0.079954
No inflection point

Gompertz model: FG ðt; p; β; q; b; cÞ p β q b

p 3.4902 (0.034389, 1.1435)


β 2.0705 (0.96339, 1.7879) −0.77802
q 1.4079e-5 (6.266e-4, 8.3508e+7) −0.93039 0.91319
b 125.05 (0.43287, 1.0837) 0.73301 −0.73532 −0.86772
c 1.8812 (0.68939, 1.0949) 0.86142 −0.83080 −0.96029 0.96820
No inflection point

Exponential model: FE ðt; p; β; q; b; cÞ p β q b

p 6.7981 (0.91570, 1.0085)


β 1.7201 (0.99685, 1.0232) −0.98973
q 1.0666e-3 (9.205e-2, 3.0884e+3) −0.93458 0.88380
b 47.862 (0.98831, 1.0014) −0.091636 0.15958 −0.16961
c 3.8119 (0.98854, 1.0012) 0.70950 −0.62505 −0.89832 0.53777
No inflection point

Table 15 Clear Creek near Sanger: Optimal values and ranking criteria values for the models.
Model Z Rank Adjusted R2 Rank RMSE Rank AIC Rank

Richardsð1Þ 23.294 4 0.99866 4 0.55917 4 −168.21 4


Richardsð2Þ 23.262 3 0.99867 3 0.55878 3 −168.42 3
Hossfeld 32.818 6 0.99813 5 0.66150 5 −118.11 5
Gompertz 23.261 2 0.99867 2 0.55690 2 −171.46 2
Exponential 20.445 1 0.99883 1 0.52211 1 −191.46 1
Rating curve 24.035 5
Modeling stage-discharge relationships 1809

as the width of the relative confidence intervals of used in this study and its value leads to the same
model parameters summarized in Tables 4, 6, 8, 10, ranking as the optimal value Z  for all the modified
and 12, and on a visual inspection of the optimal models. The numerical and ranking results given by
curve of each modified model, we conclude that for this measurement cannot, in any case, afford us the
PRRNW, WRNP, and CCNS the best fit is obtained conclusion that this statistic could be employed on
by the exponential model, followed respectively in nonlinear problems because it is essentially, and by
the second and third rank by the Richards(1) and definition, a measure of goodness of fit for linear
Hossfeld models for PRRNW, the Richards(1) and regression problems, which is not the case in this
Gompertz models for WRNP, and the Gompertz and paper. For purposes of illustration, we have included
Richards(2) models for the CCNS dataset. On CRAG in Figs 2(b)–7(b) the graphs of the best models with
and PRNV, the best fit is obtained by the Gompertz their confidence bands for the six datasets.
model, followed respectively in the second and third
rank by the Richards(12) and Richards(1) models for
CRAG, and the exponential and Richards(1) models
for the PRNV dataset. On ORNB the best fit is CONCLUSION
obtained by the Richards(1) model, followed by the In this paper we have addressed the problem of
exponential and Hossfeld models. Finally, the establishing the stage–discharge relationship on six
adjusted R-squared criterion has no significant real datasets selected randomly from the USGS data-
impact on comparisons established by the criteria base by using a nonlinear least squares method. For

(a) (b) 400


400
350
350
300
300
Discharge (m3/s)
Discharge (m3/s)

250
250
200
200

150 150

100 100

50 50

0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Gauge height (m) Gauge height (m)

Fig. 2 Paint Rock River near Woodville: (a) Graphic representation of the data points fðhk ; Qk Þgm k¼1 with the rating curve
(dotted line). (b) Graphic representation of the modified exponential model FE ðt; p; β; q; ζ Þ at the optimal solution
ðp ; β ; q ; ζ  Þ (solid line) with the lower (long dashed line) and upper (dot dashed line) bootstrap confidence bands.

(a) (b)
200 200

150 150
Discharge (m3/s)
Discharge (m3/s)

100 100

50 50

0 0
1 1.5 2 2.5 3 3.5 4 4.5 5 1 1.5 2 2.5 3 3.5 4 4.5 5
Gauge height (m) Gauge height (m)

Fig. 3 Chattooga River above Gaylesville: (a) Graphic representation of the data points fðhk ; Qk Þgm k¼1 with the rating curve
(dotted line). (b) Graphic representation of the modified Gompertz model FG ðt; p; β; q; ζ Þ at the optimal solution
ðp ; β ; q ; ζ  Þ (solid line) with the lower (long dashed line) and upper (dot-dashed line) bootstrap confidence bands.
1810 Youness Mir and François Dubeau

(a) (b)
180 180
160 160
140 140
Discharge (m3/s)

Discharge (m3/s)
120 120
100 100
80 80
60 60
40 40
20 20
0 0
1 1.5 2 2.5 3 3.5 4 4.5 5 1 1.5 2 2.5 3 3.5 4 4.5 5
Gauge height (m) Gauge height (m)

Fig. 4 Withlacoochee River near Pinetta: (a) Graphic representation of the data points fðhk ; Qk Þgm k¼1 with the rating curve
(dotted line). (b) Graphic representation of the modified exponential model FE ðt; p; β; q; ζ Þ at the optimal solution
ðp ; β ; q ; ζ  Þ (solid line) with the lower (long dashed line) and upper (dot dashed line) bootstrap confidence bands.

(a) (b)
180 180

160 160

140 140
Discharge (m3/s)
Discharge (m3/s)

120 120

100 100

80 80

60 60

40 40

20 20

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Gauge height (m) Gauge height (m)

Fig. 5 Ochlockonee River near Bloxham: (a) Graphic representation of the data points fðhk ; Qk Þgm k¼1 with the rating curve
(dotted line). (b) Graphic representation of the modified Richards(1) model FR ðt; p; β; q; ζ Þ at the optimal solution
ðp ; β ; q ; ζ  Þ (solid line) with the lower (long dashed line) and upper (dot-dashed line) bootstrap confidence bands.

(a) (b) 20
20

15
15
Discharge (m3/s)
Discharge (m3/s)

10 10

5 5

0
1 1.5 2 2.5 0
1 1.5 2 2.5
Gauge height (m)
Gauge height (m)

Fig. 6 Pease River near Vernon: (a) Graphic representation of the data points fðhk ; Qk Þgm k¼1 with the rating curve (dotted
line). (b) Graphic representation of the modified Richards(1) model FR ðt; p; β; q; ζ Þ at the optimal solution ðp ; β ; q ; ζ  Þ
(solid line) with the lower (long dashed line) and upper (dot dashed line) bootstrap confidence bands.

this purpose we have used new models based on performance and the reliability of such models on
some known sigmoid models and have modified stage–discharge datasets and to classify the new mod-
them to obtain models with curvilinear asymptotes els according to the best fits and variability of their
able to model hydrological datasets. Several compar- parameters. The results of this investigation lead us to
ison criteria were used to compare and evaluate the the conclusion that the parameter values of the
Modeling stage-discharge relationships 1811

(a) (b)
140 140

120 120

100 100
Discharge (m3/s)

Discharge (m3/s)
80 80

60 60

40 40

20 20

0 0
1 2 3 4 5 6 1 2 3 4 5 6
Gauge height (m) Gauge height (m)

Fig. 7 Clear Creek near Sanger: (a) Graphic representation of the data points fðhk ; Qk Þgm k¼1 with the rating curve (dotted
line). (b) Graphic representation of the modified exponential model FE ðt; p; β; q; ζ Þ at the optimal solution ðp ; β ; q ; ζ  Þ
(solid line) with the lower (long dashed line) and upper (dot dashed line) bootstrap confidence bands.

exponential model present less variability and less REFERENCES


degree of uncertainty around the optimal curve for
Akaike, H., 1974. A new look at the statistical model identification.
nearly all datasets studied here, even for high water IEEE Transactions on Automatic Control, 19, 716–723.
levels. This model is followed by the Richards, doi:10.1109/TAC.1974.1100705
Gompertz and Hossfeld models. In addition, accord- Braca, G., 2008. Stage-discharge relationships in open channels:
practices and problems. FORALPS Technical Report, 11.
ing to the optimal values Z  given by the modified Università degli Studi di Trento, Dipartimento di Ingegneria
new models in comparison with the sum of squared Civile e Ambientale, Trento, Italy.
errors between data and their corresponding rating Callede, J., et al., 1997. La variabilité des débits de l’amazone à
Obidos (Amazonas, Brésil). Susla inability of Water Resources
curves collected from the USGS Water Resources under Increasing Uncertainty (Proceedings of the Rabat
website, we have concluded that the exponential Symposium S1), IAHS Publ. 240, 163–172.
and Richards models successfully improve this rela- Callede, J., Kosuth, P., and De Oliveira, E., 2001. Etablissement de la
relation hauteur-débit de l’Amazone à Óbidos: méthode de la
tionship for all the data studied. Obviously, the best dénivelée normale à “géométrie variable”. Hydrological Sciences
model could change depending on the rivers studied. Journal, 46, 451–463. doi:10.1080/0262666010 9492838
As the carrying quantification of the stage–discharge Cameron, A. and Windmeijer, F., 1997. An R-squared measure of
goodness of fit for some common nonlinear regression models.
relationship depends on a set of hydrologic, hydraulic
Journal of Econometrics, 77, 329–342. doi:10.1016/S0304-
and morphological factors, modelling carrying capa- 4076(96)01818-0
city is an open, interesting and important task, but it Coleman, T. and Li, Y., 1994. On the convergence of interior-reflec-
tive Newton methods for nonlinear minimization subject to
is beyond the scope of this paper. In future work, it
bounds. Mathematical Programming, 67, 189–224.
would be interesting to investigate the physical sig- doi:10.1007/BF01582221
nificance of the parameter variations of the modified Coleman, T. and Li, Y., 1996. An interior trust region approach for
new models and validate these models accordingly. nonlinear minimization subject to bounds. SIAM Journal on
Optimization, 6, 418–445. doi:10.1137/0806023
Finally, as a complement to this work, it could be Darmani Kuhi, H., et al., 2010. A review of mathematical functions
interesting to study the behaviour of these models for the analysis of growth in poultry. World’s Poultry Science
with populations where the true values of parameters Journal, 66, 227–240. doi:10.1017/S0043933910000280
Deka, P. and Chandramouli, V., 2003. A fuzzy neural network model
are given, as done in Dubeau and Mir (2013, 2014). for deriving the river stage-discharge relationship.
Hydrological Sciences Journal, 48, 197–209. doi:10.1623/
hysj.48.2.197.44697
Acknowledgements The research work of the second Di Baldassarre, G. and Claps, P., 2011. A hydraulic study on the
author is partially supported by NSERC (Natural applicability of flood rating curves. Hydrology Research, 1,
Sciences and Engineering Research Council of 10–19.
Domeneghetti, A., Castellarin, A., and Brath, A., 2012. Assessing
Canada). The authors wish to thank Professor rating-curve uncertainty and its effects on hydraulic model
Taoufik Bouezmarni for enlightening discussions on calibration. Hydrology and Earth System Sciences, 16, 1191–
this paper. The authors are also grateful to the 1202. doi:10.5194/hess-16-1191-2012
Dubeau, F. and Mir, Y., 2011. Least squares fitting with single
reviewers for their comments and suggestions. inflection point growth curve I - the models. Mathematical
Modelling and Applied Computing, 2, 269–281.
Dubeau, F. and Mir, Y., 2013. Growth models with oblique asymp-
Disclosure statement No potential conflict of inter- tote. Mathematical Modelling and Analysis, 18, 204–218.
est was reported by the author(s). doi:10.3846/13926292.2013.781068
1812 Youness Mir and François Dubeau

Dubeau, F. and Mir, Y., 2014. Exponential growth model: from Parodi, U. and Ferraris, L., 2004. Influence of stage discharge
horizontal to linear asymptote. Communications in Statistics - relationship on the annual maximum discharge statistics.
Simulation and Computation, 43, 2186–2204. doi:10.1080/ Natural Hazards, 31, 603–611. doi:10.1023/B:NHAZ.
03610918.2012.748912. 0000024893.57284.0e
Dubeau, F., et al., 2011. Least squares fitting with single inflection Petersen-Øverleir, A., 2004. Accounting for heteroscedasticity in
point growth curve II - an application. Mathematical Modelling rating curve estimates. Journal of Hydrology, 292, 173–181.
and Applied Computing, 2, 283–301. doi:10.1016/j.jhydrol.2003.12.024
Dubeau, F., et al., 2012. Modelling stage-discharge relationship with Petersen-Øverleir, A., 2005. A hydraulics perspective on the power-
single inflection point nonlinear functions. International law stage-discharge rating curve. Norwegian Water Resources
Journal of Hydrology Science and Technology, 2, 153–167. and Energy Directorate, Oslo, NVE report 05–05.
doi:10.1504/IJHST.2012.047430 Petersen-Øverleir, A., 2006. Modelling stage-discharge relationships
Dymond, J.R. and Christian, R., 1982. Accuracy of discharge deter- affected by hysteresis using the Jones formula and nonlinear
mined from a rating curve. Hydrological Sciences Journal, 27, regression. Hydrological Sciences Journal, 51, 365–388.
493–504. doi:10.1080/02626668209491128 doi:10.1623/hysj.51.3.365
Efron, B., 1979. Bootstrap methods: another look at the jackknife. Schmidt, A.R. and Yen, B.C., 2001. Stage-discharge relationship in
The Annals of Statistics, 7, 1–26. doi:10.1214/aos/1176344552 open channels. New York: Proceedings of the 2001
Efron, B., 1985. Bootstrap confidence intervals for a class of parametric International Symposium on Environmental Hydraulics.
problems. Biometrika, 72, 45–58. doi:10.1093/biomet/72.1.45 Shepherd, J.J. and Stojkov, L., 2007. The logistic population model
Habid, E. and Meselhe, E., 2006. Stage-discharge relations for low- with slowly varying carrying capacity. Anziam Journal, 47,
gradient tidal streams using data-driver models. Journal of 492–506.
Hydraulic Engineering ASCE, 132, 482–492. Shrestha, R., Bárdossy, A., and Nestmann, F., 2007. Analysis and
Herschy, R.W., 1995. Streamflow measurement. London: Chapman propagation of uncertainties due to the stage-discharge relation-
& Hall. ship: a fuzzy set approach. Hydrological Sciences Journal, 52,
Herschy, R.W., 1999. Hydrometry: principles and practices. New 595–610. doi:10.1623/hysj.52.4.595
York, NY: Wiley. Sivapragasam, C. and Muttil, N., 2005. Discharge rating curve
Ismail, Z., Khamis, A., and Jaafar, M., 2003. Fitting nonlinear extension - A new approach. Water Resources Management,
Gompertz curve to tobacco growth data. Pakistan Journal of 19, 505–520. doi:10.1007/s11269-005-6811-2
Agronomy, 2, 223–236. doi:10.3923/ja.2003.223.236 Tarald, O.K., 1985. Cautionary note about R2. American Statistical
Jain, S. and Chalisgaonkar, D., 2000. Setting up stage discharge Association, 39, 279–285.
relations using ANN. Journal of Hydrologic Engineering, 5, Tawfik, M., Ibrahim, A., and Fahmy, H., 1997. Hysteresis sensitive
428–433. doi:10.1061/(ASCE)1084-0699(2000)5:4(428) neural network for modeling rating curves. Journal of
Khamis, A., et al., 2005. Nonlinear growth models for modeling oil Computing in Civil Engineering, 11, 206–211. doi:10.1061/
palm yield growth. Journal of Mathematics and Statistics, 1, (ASCE)0887-3801(1997)11:3(206)
225–233. doi:10.3844/jmssp.2005.225.233 Torsten, D., Gerd, M., and Torsten, S., 2002. Extrapolating stage-
McGinn, R. and Chubak, N., 2002. A rating curve based on lake discharge relationships by numerical modelling. International
levels: evaluating outlet flow for clear creek, riding mountain Conference on Hydraulic Engineering, Warshaw, 1–8.
national park, Manitoba. Prairie Perspectives: Geographical Westphal, J., et al., 1999. Stage-discharge relations on the middle
Essays, 5, 17–29. Mississippi River. Journal of Water Resources Planning and
Meyer, P.S. and Ausubel, J.H., 1999. Carrying capacity: a model with Management, 125, 48–53. doi:10.1061/(ASCE)0733-9496
logistically varying limits. Technological Forecasting and Social (1999)125:1(48)
Change, 61, 209–214. doi:10.1016/S0040-1625(99)00022-0 Wu, R.-B. and Yang, J.-C., 2008. An analytical method of stage-fall-
Meyer, P.S., Yung, J.W., and Ausubel, J.H., 1999. A primer on discharge rating. Hydrological Processes, 22, 2959–2973.
logistic growth and substitution: the mathematics of the doi:10.1002/hyp.6867
Loglet lab software. Technological Forecasting and Social Zeide, B., 1993. Analysis of growth equations. Forest Science, 39,
Change, 61, 247–271. doi:10.1016/S0040-1625(99)00021-9 594–616.

You might also like