FLOOD05893

• County: Statewide
• Basin: Statewide
• Title: Forecasting Heavy Rainfall from Rain Cell Motion using Radar Data
• Date: 6/2/1992
• Prepared For: Journal of Hydrology
• Prepared By: Elsevier Science Publishers
• Floodplain - Doc Type: Educational/Technical/Reference Information

382

P.J. BR~MAUD AND V.B. POtNTIN

FORECASTING HEAVY RAINFALL FROM RAIN CELL MOTION

383

stratiform rain area, a succession of convective bands, the development of a
frontal band and isolated storms, The radar data for these four events have
different spatial scales (240m, 360m, 432m and 800m respectively for the
pixel size) and different time intervals between successive radar pictures
(4min, 5min, 8min and 12min), and one event has been used to forecast
the rain at three different time intervals (5, 10 and 15min), Each forecast-
ing method is judged by using the values, computed for each prediction
prediction time step, of six "quality criteria: three 'concordance' criteria
(critical success index or CSI, Rousseau index or RI and cross-correlation
coefficient or CC) and three hydrological criteria, These last criteria,
introduced by Einfalt et al. (1990), quantify the underestimation and the
overestimation of the observed rain, and are defined by

I .-
- L (dH,,, - dH,,,,)
n_ 1...1

dHS+

I '.
- L (dH,,, - dH'b')
n+ i...l

value, in contrast to the critical success index CSI and the Rousseau index RI,
which are calculated from a contingency table of the type 'rain/no rain', i,e,
independent of the rain quantities, Besides, CSI and RI are, by definition, less
sensitive than CC to the number of 'rainy pixels' inside the picture. However,
the RI criterion is calculated on the same threshold rainy pixels as for the
hydrological criteria, to emphasize the important hydrological events,
For each event and for each forecasting method, the values of each criterion
are evaluated at each time step and are reordered in increasing order (or
decreasing order for dHS+) in such a way as to draw the resulting curves from
left to right with improving performance (efficiency curves), For each method,
the best performance values are on the righthand side of the curves and the
worst are on the lefthand side, This implies that the method whose efficiency
curve is globally above the efficiency curves of the other methods (or globally
below for the criterion dHS+) can be considered the best method for this
criterion. The use of these efficiency curves avoids the troughs and peaks of
the time evolution curves which are due to morphology and intensity changes
of the radar echoes from one time step to the next.
These efficiency curves of the CC, RI, dHS_ and dHS. criteria, evaluated for
a threshold value of 5 mm h-I (except for CC, which is calculated from all the
pixels of both the observed picture and the forecast picture), appear in Figs,
6(a)-6(d) respectively, for the 'Covennes 1988' event. In these figures, the
PERS, EXTRA and PARAPLUIE methods are depicted with dotted line,
dashed line and solid line respectively, The 'Covennes 1988' event is the most
important hydrological event studied (rainfall rate greater than 125mmh-1
for more than 24 min), The successive pictures of this event, recorded at
12min intervals for the first 3h and at 8min intervals for the next 3h, show
the formation and the propagation of an intense frontal band (Fig, I),
As its efficiency curve is globally abo~e the other curves in Figs, 6(a)-6(c), the
PARAPLUIE method gives the best performance according to the CC, RI or
dHS_ criteria, but all the methods generally overestimate the precipitation in
the same manner (Fig, 6(d)), The differences between the performance of the
EXTRA and PARAPLUlE methods occur because EXTRA detects the
motion of the entire frontal band (5ms-1 towards the east) whereas
PARAPLUIE detects the motion of the rain cells which are embedded within
the band (II,S m S-I towards the north-northeast), As these rain cells provide
the largest part of the precipitation observed at ground, the worse forecast of
their location by the EXTRA method makes it underestimate the real quantity
of the observed rain.
The same efficiency curves of the criteria, evaluated for a threshold value
of 3 mmh-I, are shown in Figs, 7(a)-7(d) for the 'Paris 1989' event, which is
the only event for which the SCOUT 11,0 method can be compared with the

dHS_

dHS = (n_' dHS_) + (n+' dHS+)
n_ + n+

where dHobs and dHror are respectively the equivalent intensity (R) calculated
on nine pixel areas for the observed picture and for the forecast picture which
is deduced, by each method, from the two previous observed pictures.
Similarly, n_ and n+ are equal to the number of pixels where
dHror - dHobs :E; 0 and to the number of pixels where dRfor - dRobs > 0,
respectively, All these criteria are calculated only on the observed rainy pixels
whose equivalent intensity exceeds a fixed threshold value (3 or 5mmh-1
according to the hydrological characteristics of the event) to show how each
method is able to forecast the real quantity of heavy precipitation, Therefore,
the criterion dHS_ evaluates the underestimation of the observed rain.
whereas the criterion dHS+ characterizes the overestimation. and the criterion
dHS weighs the first two against each other, According to these hydrological
criteria, the best method is that which gives the best performance according
to dHS_ and dHS. whatever the dHS value is, or that which gives the best
performance according to dHS_ (or dHS.) and dHS if dHS_ and dHS+
show different tendencies.
As these hydrological criteria do not take into accol.!nt the false alarms. i.e.
forecast rain which isinot observed. it is important. even for hydrological
purposes, to consider the results according to concordance criteria. Of these
last criteria, the cross~correlation coefficient CC is the most suited for
hydrology, because its value depends on the concordance of each reflectivity