Analysis of peaks over threshold within daily data

for flood frequency curve assessment

b
P. CLAPSa & P. VILLANI
a
DIFA Universit della Basilicata Potenza (Italy) [claps@unibas.it]
b
DIC Universit di Parma (Italy) [villani@parma1.eng.unipr.it]

ABSTRACT
We report an attempt to relate basic properties of stochastic processes of basin
intense rainfalls and floods, in order to improve the prediction of flood risk by
moving information from rainfalls to floods. This approach is based on a
stochastic model of the extreme precipitation in which a wet and a dry season
can be distinguished. During the wet period, the process of rainfall occurrences
can be assumed as Poissonian and the distribution of marks can be derived
trough the analysis of the peak over the threshold (POT) process. The basin
operates on the rainfall process as a stochastic filter, constituted by a stochastic
threshold S on precipitation, that reduces the number of runoff events.
Theoretical analysis of the filter mechanism provides the relationships between
stochastic parameters of maximum rainfall and flood. An application to data
from a Southern Italy basin showed that magnitude of S is higher at low
probability level of X and Y than at high probabilities, and the decay is rather
steep. This finding seems to suggest a peculiar two-state behaviour for the basin
antecedent moisture conditions: a wet state, in correspondence of annual
maximum events, and a dry state for base process, both during the wet season
(October-May).
1 INTRODUCTION
Peaks Over Threshold (POT) models have been frequently employed for the
study of flood frequency analysis, even though with much less emphasis than
the popular Annual Maximum Flood (AMF) sampling. Based on the
applications available in literature, the POT approach seems to present under-
employed potential and its reconsideration looks promising for the information
content it can transfer into the flood analysis (see e.g. Lang et al., 1999;
Sirangelo and Iiritano, 1999). In particular, one of the less underemployed
features of the POT analysis is its capability to give information on the peak
distribution and on the stochastic process of peak occurrence from the
characteristics of processes derived from the cut operated by different
thresholds.
The aim of this work is to analyse peak occurrence in daily data of
rainfall and runoff and compare parameters of both POT processes in order to
shed light on the probabilistic aspects of the transformation operated by the
basin. In this analysis we consider a stochastic model simple yet consistent with
the possibility of analysing the basic aspects of the transformation mechanism
under a probabilistic viewpoint. In particular, we show that a Poisson model of
occurrences and a quasi-homogeneous process structure over seasons and over
return periods (no separation probabilistic structure) seem to meet this
requirement.
The problem under study admits a feasible theoretical approach, based
on the theory of derived distributions. According to this approach, it is possible
to derive the peak discharge probability distribution from the analysis of the
joint probability distribution of the rainfall peaks and of the variables describing
basin transformation. Usually, the latter are considered as deterministic
parameters (e.g., Bierkiens and Puente, 1990). However, basin antecedent soil
moisture conditions are decisive in determining the magnitude of the peak
discharge (see e.g. Todini, 1996).
In this paper, we will take explicitly into account the stochastic nature
of some basin parameters and we will use a non-parametric technique to derive
the stochastic structure of a basin parameter related to soil moisture from daily
rainfall and discharge data. In particular, our attention is focused on the
threshold over which rainfall peaks produce discharge peaks. We will show, as
a result, the presence of a two-state behaviour, for low and high probability of
rainfalls and floods. This finding, if confirmed on regional basis, could help
explaining the relative role played by the basin antecedent moisture condition
and the peak rainfall in determining very high flood peaks.
2 STOCHASTIC MODEL OF PEAK OCCURRENCES
2.1 Mechanism of peak occurrences

From the analysis of peak occurrences in basins of the Mediterranean area it

results evident that this stochastic process is not homogeneous in time. A first
basic distinction can be made observing the process patterns in two different
climatic seasons: a dry season during the summer period, and a wet season
otherwise (Figure 1). This approximation looks reasonable and parsimonious as
a starting point (see also Sirangelo, 1994).
As said above, we assume a Poissonian scheme for modelling time
occurrences of flood peaks during the wet season, based on the observation that
the exponential probability distribution of the inter-arrival time of the peaks can
be valid as a first approximation. If necessary, one could also distinguish a finer
time structure during the wet season.

Alento
35

30

25
daily runoff peaks (mm)

20

15

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time in months

Figure 1. Process of occurrence of peaks of mean daily discharge.

2.2 Probability distribution of peaks

As a consequence of the Poissonian hypothesis, if (x) is the mean annual

number of peak occurrences over a threshold of magnitude x and (xo) is the
analogue for a value xo low enough to have a cumulative probability FX(xo) 0,
but high enough to allow the Poissonian hypothesis to be valid, then the
following relationship holds (see e.g. inlar, 1975, p. 94]

(x)/(xo) = 1 FX(x) ( x xo ) (1)

where FX(x) is the cumulative distribution function (CDF) of the random variate
X. In practice, the importance of the choice of xo decreases as much as X
becomes high enough to consider valid the Poissonian hypothesis on
occurrences.

2.3 Statistical relation between rainfall and discharge POT processes

Let us consider:
X = total rainfall depth in 1 day, over the basin (in mm);
Y = total discharge volume in 1 day, from the basin outlet (in mm).
We assume Y to be the result of a deterministic/stochastic filter operated by the
basin over the stochastic process of the daily rainfall. The transformation we
consider is a simple non-linear, threshold model as

Y=0 if XS
(2)
Y = g(XS; ) otherwise

where g(; ) is a general transformation function, that depends on the

parameter vector , and contains information on the basin characteristics. Filter
parameters S and can be both deterministic and stochastic. As a first
approximation, we consider the basin characteristics as deterministic, while
the filter threshold S can take into account the stochastic state of the system
preceding the intense rainfall generating the flood peaks.
Based on these assumptions, discharge peaks will be considered equivalent
to over-threshold rainfalls (Y=XS), producing a relationship between the
annual rate of flood peaks Y and rainfall peaks X of the type of eq. (1), that
gives here the probability of the (external) filter threshold S:

Y = X [1 FX(S)] (3)

Eq. (3) indicates that, for any transformation function g(; ), the mean
annual number of flood peak occurrences can be derived through the knowledge
of the probability distribution of rainfall peaks FX() and of the mean annual
number of rainfall peak occurrences, given a state of the system, i.e. a value of
S. Both Y and X have a probabilistic structure that depends on the threshold
level, as shown in (1). The control operated by the threshold filter S on the
transformation processes will be then inferred by the ratio of the empirical
occurrence parameters Y and X computed through POT analysis. In
particular, using historical records of areal daily rainfall and of mean daily
discharges we will show how it is possible to characterise the probability
distribution of S and the relative role that its stochastic variability plays in
producing high flood peaks.

3 POT ANALYSIS AND CASE STUDY

3.1 Criteria for identification of peaks

On daily discharge records, consecutive and different peak occurrences (in

the statistical sense of independence) must be suitably identified. To this end, a
standard procedure adapted from the Matlab library, namely the findpeaks
function [The Mathworks] was applied to runoff records, leading to the
selection of isolated peaks (see Figure 2). Analysis of interarrival times during
the wet season (Figure 3) showed that additional filtering was not necessary, at
least at the present stage of analysis.

Alento
15
avg. daily discharge (mc/s)

10

0
2900 2950 3000 3050 3100 3150 3200 3250 3300 3350
time in days

Figure 2. temporal process of peak discharge occurrence

 0
Alento
10
1-Cumulative Probability

-1
10

-2
10
0 50 100 150
Interarrival time (days)

Figure 3. Probability distribution of the inter-arrival time of discharge peak occurrences

during the wet season: estimates (circles) and fitted exponential distribution (dots).

3.2 Relation between filter threshold S and probability of the events

To identify the statistical relationship between stochastic filter threshold

S and probability of the peak event we used the following procedure
(graphically represented in Figures 4a and 4b):

(i) Given a probability 1 FX(X) for rainfall peaks, oX/o gives an

occurrence level X and a corresponding threshold X (Fig. 4a).
(ii) Since (3) is valid, we search a value of the discharge peak Y such that
X = Y = (Fig. 4b).
(iii) The difference between X and Y values taken at the same represents
the estimate of SX corresponding to the probability 1 FX(X).
 0
Alento
10

1-F(X)
(a)
-1
10
/0

-2
10

X
-3
10
0 10 20 30 40 50 60
rainfall threshold (mm)

2
Alento
10
rainfall
runoff

(b)
1
10
S(X)

0
10

0 10 20 30 40 50 60
threshold (mm)

Figure 4. Sketch of the procedure of deriving the statistical relationship between the
threshold S and the probability of the rainfall peak.
3.2 Basin description and application of the procedure

Areal rainfall and discharge data from a basin located in southern Italy have
been analysed. The basin belongs to the Alento river, in Campania, has an area
of 285 Km2 and presents only a very small fraction of carbonate formations, so
to allow us to classify it as almost completely impermeable. Based on the basin
area, order of magnitude of discharges can be determined from the figures
below considering that 1 mm of equivalent discharge measures about 3.3 m3/s.
On the basin under study 13 years of daily average discharge and 20 years of
daily rainfall were available. All of the figures presented above are referred to
these hydrological records.
The procedure discussed in the above paragraph was applied to the curves of
(X) (figure 4) obtained for different thresholds on rainfalls and discharges.
Applying the proposed procedure for probability levels ranging from 0.1 to
over 0.99 we obtained the curve represented in Figure 5 , that shows a peculiar
relationship between the probability level F of the rainfall and the stochastic
threshold S of the basin. Note that using the representation in terms of 1-F, in
reverse scale, provides a correct viewpoint in terms of recurrence interval
increasing toward the right part of the scale.

Alento
16

15

14

13

12
S (mm)

11

10

7
0 -1 -2 -3
10 10 10 10
1-F(X)
Figure 5. Statistical relationship between the threshold S and the probability of the
rainfall peak for the Alento river
 In the figure we can distinguish different zones:

i) very low value of F (non-Poissonian zone): we disregarded values obtained

for F<0.5, that are of no interest for the goals of this application, and for
which the Poissonian hypothesis in eq. (1) cannot be considered valid, even
in first approximation. Moreover, F is the cumulative probability of the
whole ensemble of rainfall and runoff peaks (not annual maximum), while
annual maximum rainstorms and floods cover only the higher part of the
CDF;
ii) low value of F (dry state): for low return period of rainfall peaks (F~0.85)
the basin threshold S is relatively constant and high, being 14 mm, that is
36% of mean annual daily flood;
iii) high value of F (wet state): for high return period of rainfall peaks (F>0.99)
the basin threshold S is relatively constant and low, being 8 mm, that is less
than 60% of the value assumed in the dry state;
iv) intermediate value of F (transition phase): during this phase the threshold
S is not a constant and must be considered a stochastic variable.

Based on this result, it seems possible to identify a two-state pattern that

characterises the system transformation in dependence of the probability level
of rainfalls and floods.

4. FINAL DISCUSSION
In order to identify the mechanism leading to flood peak distribution Y
starting from rainfall peak distribution X, we considered a general framework in
which the basin is a deterministic filter that operates through a stochastic
threshold parameter S.
We did not address the theoretical derivation of the distribution of Y that
considers the joint probability distribution of X and S. This distribution is
difficult to derive from usual hydrological information, because the analytical
approach has to face the dependence of S on antecedent moisture condition.
Our alternative approach is based on peak over threshold (POT) analysis of
daily rainfall and discharge: under the simple, but realistic assumption of
Poissonian occurrence of peaks, we showed how it is possible to derive a
relationship between the magnitude of S and the probability level of X and Y.
Previous analyses considered only the annual maximum series (AMS) of
daily rainfall and flood: Ferrari et al. [1993] considered the ratio Y/ X, that
can be considered a surrogate of S, only as a deterministic parameter of the
basin response, constant over a climatic region. Recently, an attempt was made
to explain the observed spatial variability of that ratio taking into account the
permeability characteristic of the basin and a surrogate of the antecedent soil
moisture condition considered constant and related to the Thorntwaite annual
climatic index [Iacobellis et al., 1998]: no estimation of the probability level
associated with this ratio can be made, because of the use of the AMS. The
presented procedure, taking the advantage of using the POT techniques of daily
rainfall and discharge, is the first attempt to explicitly estimate the relative role
played by deterministic and stochastic parameter of the basin, in transforming
heavy rain into flood peak discharge.
An application to a case study in a Southern Italy basin has shown a peculiar
probabilistic structure of S: the magnitude of S is higher at low probability level
of X and Y than at high probabilities, and the decay is quite steep. This findings
seems to suggest a peculiar two-state behaviour for the basin antecedent
moisture conditions: a wet state, in correspondence of annual maximum events,
and a dry state for base process. It is important to consider that both states were
found during the wet season (October-May). Other authors found a two-state
basin behaviour, but in two different seasons (e.g. Grayson et al., 1997).
Such a mechanism could explain some extraordinary flood events
experienced in the past in the same region of the case study and not justified by
extraordinary rainfall peaks: inundations could have been produced by a
combination of high rainfall peak and extraordinary wet condition of the basin,
represented by a very low value of S. Further investigations are needed to
evaluate the possibility that POT give information useful for the separation
process.
Identification of the probability structure of S can be made in terms of: (i)
magnitude characterising wet and dry state; (ii) magnitude of difference
between the two states (or jump), and (iii) location, in probability, of the break-
points between the two states. The ensemble of information resulting on S have
suggested the possibility of a regionalization of the proposed procedure, whose
goal is the consideration of this stochastic parameter as a function of basin
hydro-geological characteristics: to this aim, further investigation are needed, in
order to extend the different environments in which the procedure can be
applied and to test and, eventually, relax, the base hypotheses of Poissonian
processes.

6. REFERENCES
Bierkens, M.F. P. and C. E. Puente, Analytically Derived Runoff Models Based
on Rainfall Point Processes, Water Resour. Res., 1990, vol. 26, no. 11, pp.
2653-2665.
inlar, E., Introduction to Stochastic Processes, Prentice-Hall, Inc., Englewood
Cliffs, N.J. (USA), 1975, pp. 402.
Iacobellis, V., Claps, P., & Fiorentino, M. (1998) Sulla dipendenza dal clima dei
parametri della distribuzione di probabilit delle piene, Proc. XIV Conf di
 Idraulica e Costruzioni Idrauliche, vol II, 213-224, CUECM, Italy, (in
Italian).
Ferrari, E., S. Gabriele and P. Villani, Combined regional frequency analysis of
extreme rainfalls and floods, 6th IAHS/IAMAP Symp., Yokohama (Japan),
1993.
Grayson R.B., A.W. Western and F.H.S. Chiew, Preferred states in spatial soil
moisture patterns: local and non-local controls, Water Resour. Res., 1997,
vol. 33, no. 12, pp. 2897-2908.
Lang, M., T.B.M.J. Ouarda and B. Bobe, Toward operational guidelines for
over-threshold modeling, J. Hydrol., 225, 1999, pp. 103-117.
Sirangelo, B., Un modello Poissoniano non omogeneo per lanalisi delle
precipitazioni cumulate su lunghi intervalli temporali, Idrotecnica, 3, 1994,
pp. 125-158.
Sirangelo, B. and G. Iiritano, I processi puntuali nello studio delle occorrenza
delle precipitazioni intense, LAcqua, 6, 1999, pp. 35-60.
The MathWorks, Natick, MA (http://www.mathworks.com).
Todini, E., The ARNO rainfall-runoff model, J. Hydrol., 175, 1996, pp. 339-
382.