1

VERY HIGH RESOLUTION PRECIPITATION FORECASTING ON LOW COST HIGHPERFORMANCE COMPUTER SYSTEMS IN SUPPORT OF HYDROLOGICAL MODELING

Daniel Soderman1, Francesco Meneguzzo2*, Bernardo Gozzini2, Daniele Grifoni2, Gianni Messeri3,Matteo Rossi3, Simone Montagnani3, Massimiliano Pasqui2, Andrea Orlandi2, Alberto Ortolani3, Ezio

Todini4, Giovanni Menduni5, and Vincenzo Levizzani6

1Foreca Ltd, Pursimiehenkatu 29-31B, 00150 Helsinki, Finland2Institute of Biometeorology - National Research Council (IBIMET-CNR), Firenze, Italy

3Laboratory for Meteorology and Environmental Modeling (LaMMA-Regione Toscana), Firenze, Italy4Department of Earth and Geo-Environmental Sciences, University of Bologna, Bologna, Italy

5Arno River Basin Authority, Firenze, Italy6 Institute of Atmospheric Sciences and Climate - National Research Council (ISAC-CNR), Bologna,

Italy

1. INTRODUCTION

Early flash flood warnings are needed toimprove the effectiveness of civil protectionefforts to mitigate damages and save lives. Thisis especially true for small to medium size (sub-)basins where the response times to rainfall inputare of the order few hours or less.

This is the case of the Arno River Basin(Italy), and of virtually any basin, when uppersections are considered.

As part of the effort to producecontinuous accurate hydrological flood forecastsalong the Arno river and its tributaries, the ArnoRiver Basin Authority sponsors the verificationand improvement of the operational RegionalAtmospheric Modeling System (RAMS) as aneffective QPF tool to feed hydrological (flood)models in the frame of the ARTU project(http://www.arno.autoritadibacino.it).

Furthermore, the EURAINSAT project(an EC research project co-funded by theEnergy, Environment and SustainableDevelopment Programme within the topic“Development of generic Earth observationtechnologies”, Contract number EVG1-2000-00030) sponsors the development of the RAMSmodel with the twofold purpose of providingreliable input to rapid update satelliteprecipitation estimation algorithms and usingsuch estimates to effectively improve mesoscaleQPF forecasts by means of operationalprecipitation assimilation(http://www.isac.cnr.it/~eurainsat).

Finally, the MUSIC project (research

project also co-funded by the EC under theabove Development Programme within the topic“Development of generic Earth observationtechnologies”, Contract number EVK1-CT-2000-00058) sponsors the development of advancedhydro-meteorological forecasting tools, whereaccurate QPF plays a decisive role.

The need to run the RAMS model overdifferent domains at varying horizontalresolutions led to the use of low cost highperformance and effectively scaling computersystems.

2. THE RAMS MODEL

The Regional Atmospheric ModelingSystem, RAMS, version 4.4, is used for theoperational mesoscale weather forecasts at theRegional Meteorological Service of Tuscany,Italy (Meneguzzo et al., 2002; see also:http://www.lamma.rete.toscana.it/rams-web/e_index.html), and was used for all the casestudies described in the following.

RAMS was originally developed in theearly 70's essentially as a research tool; now it iswidely used both for research and operationalforecast purposes in many operational centersaround the world. Since the early ‘90s a largenumber of improvements have been introducedboth as regards the physics (mainly moist andland surface processes) and the computationalpoint design (new numerical schemes andparallel computing). A general description of themodel can be found in Pielke et al. (1992), whilea technical description can be found inMRC/*Aster (2000).

RAMS today represents the state-of-the-art inatmospheric numerical modeling beingcontinuously improved on the basis of multi-disciplinary work both at Colorado State

J4.8

____________________________________* Corresponding author address: FrancescoMeneguzzo, Institute of Biometeorology,National research Council, Via Caproni 8,50145 Firenze, Italy; e-mail:meneguzzo@lamma.rete.toscana.it

2

University and at several other research andoperational centers worldwide. Beyondnumerical modelers, meteorologists, computerexperts, remote sensing experts, geographers,agronomists, forest experts and geologistsparticipate to the development of the various ofcomponents of RAMS.

2.1 Initialization; objective analysisRAMS needs a set of atmospheric data

analyzed over the domain grid and at itsboundaries both at the initial time of thesimulation and at future times. Several datatypes can be combined and processed by anisentropic analysis package (ISAN). Theisentropic coordinates have several advantagesas regards to other coordinate systems: sincethe synoptic scale atmospheric flow is to a firstapproximation adiabatic, the objective analysisperformed over an isentropic surface will “pack”in frontal areas, thus providing a higherresolution description of the discontinuities.

Besides, since the isentropic surfacesare steeper close to the fronts, shorterwavelength features are transformed into longerwavelength ones and can be analyzed moreaccurately. ISAN allows a separate data analysisover the different nested grids (two-way nestingis implemented), thus leading to a better use ofdata from dense networks. The atmosphericfields distributed over regular grids which aregenerally provided by larger scale models, arefirst analyzed and interpolated over the RAMSgrid using a polar-stereographic domain whichminimizes distortions. Following, the data arevertically interpolated in the isentropic andterrain-following co-ordinate systems. After this,a large set of a variety of data types can beassimilated.

The Barnes (1973) objective analysisscheme is used for wind, pressure and relativehumidity. The local surface observation dataover land and water can be assimilated andpropagated along the vertical up to a variablealtitude. A four-dimensional data assimilationscheme has been implemented in RAMS, wherethe model fields can be nudged towards theobservations during the simulations; a furthernudging term is thus added to the prognosticequations.

2.2 Basic EquationsThe basic equations of RAMS are three

dimensional, non-hydrostatic, compressible andtime-split in an horizontal rotated polar-stereographic transformation and vertical terrainfollowing height. The non-hydrostatic formulationand the large number of terms retained in theRAMS equations allow to attain in principle anyspatial horizontal resolution, which has made

applications from planetary climate to tunnelwind possible.

2.3 MicrophysicsThe representation of cloud and precipitation

microphysics in RAMS includes the treatment ofeach water species (cloud water, rain, pristineice, snow, aggregates, graupel, hail) as ageneralized Gamma distribution (Pielke et al.,1992; Walko et al., 1995). The scheme allowshail to contain liquid water and contains thedescription of the homogeneous andheterogeneous ice nucleation, and the ice sizechange by means of vapor deposition andsublimation.

A very efficient solution technique is availablefor the stochastic collection equation and a newtechnique for the prediction of sedimentation orprecipitation of hydrometeors, allowing thedefinition of the fall velocity on the basis of thegamma size distribution.

2.4 Representation of Cumulus ConvectionThe cumulus convection parameterization in

RAMS (Tremback, 1990), which is still to beapplied to coarse grids (over 15 km horizontalspatial resolution) is relatively simple. Thisreflects on one side the greater effort spenttowards the explicit representation of convectiveprocesses at very high resolution, and on theother the consideration that simple cumulusconvection schemes allow more straightforwardassimilation of non standard data such as cloudand precipitation observations (diabaticinitialization).

The scheme implemented in RAMS is ageneralized form (Molinari, 1995; Molinari et al.,1995) of the Kuo (1974) equilibrium scheme.

The equilibrium schemes assume that theconvection is fed by the consumption of theconditional instability, with a speed equal to itsproduction by the processes occurring at the gridmesh size scale.

Convection can be generated by a largenumber of different causes, the main of which isrepresented by the low level air massconvergence originated by gravity waves, whereits interaction with the boundary layer sometimescan trigger the convection.

The cumulus convection schemes deal withthe convection dynamics as a sub-grid process,i.e. not explicitly solved by the equations ofmotion.

This imposes a lower limit to the modelresolution enhancement, since at highresolutions a convective process might spanmore grid nodes, thus undermining the sub-gridhypothesis.

3

2.5 Representation of Land SurfaceProcesses

The surface heterogeneities connectedto the vegetation cover and the land use areassimilated and represented in great detail inRAMS by means of the LEAF-2 (LandEcosystem Atmosphere Feedback) model.

LEAF-2 (Walko et al., 2000) representsthe vertical exchange of water and heat inseveral soil layers, including the effects offreezing and melting, the temporary water andsnow cover, the vegetation and the canopy air.The surface domain meshes are further sub-divided into patches, each identified by aseparate vegetation cover and land use, soiltype and initial soil moisture (Fig. 1).

The balance equations for soil energyand moisture, surface water, vegetation andcanopy air, and exchange with the freeatmosphere, are solved separately for eachpatch. A hydrological model based on the Darcylaw for the lateral downslope water transportexchanges the moisture in the sub-surfacesaturated layers and the surface runoff.LEAF-2 assimilates standard land use datasetsto define the prevailing land cover in each gridmesh and possibly the patches, thenparameterizes the vegetation effects by meansof biophysical quantities.

Fig. 1. LEAF-2 vertical levels and patches

3. COMPUTATIONAL SYSTEM ANDCOMPUTING PERFORMANCES

Numerical weather prediction modelsare amongst the computationally mostdemanding systems in existence so far. Using acomplete set of Navier – Stokes equations alongwith complexes numerical schemes for soil –vegetation – air interactions and microphysicalprecipitation description enlarge the NWPcomputational demands. However, since fewyears, parallel computational systems based onthe Linux operating system are becomingpopular and today represent real “highperformance – low cost” systems. Such parallelarchitectures also represent an alternative to

massive parallel systems which arecharacterized by very high costs. Thefundamental scheme is called Beowulf – like andit comprises a number of PCs, called alsoNodes, connected through a switch togetherwithin a local area network (see scheme inFig. 2).

Fig. 2. Summary scheme of the Linux BeowulfCluster, called “Bellerophon”, used in this study

A special node, called “master” is theDynamical Name Server and provides all thenetwork specifics. A collection of libraries called“middleware” handles communicating andtransferring information among nodes (Fig. 3).

Fig. 3. Conceptual scheme of parallelcomputing by the use of a Linux Beowulf Cluster,where RAMS 4.4 is the Numerical Weather Model,MPICH stands for "Message Passing InterfaceChamaleont", i.e. the actual parallel environment

Thus the number of operations needed,during an atmospheric simulation, is distributedamong available cluster nodes reducing the totalamount of overall time needed. A large numberof problems arise, essentially due to thecommunication aspects between nodes.Network traffic, calibration and synchronicityamong available nodes can reduce parallelcomputational efficiency, but essentially dueboth to the increased CPU computational powerand to the availability of efficient and robustnetwork drivers it is possible to achieve a goodperformance. Every simulation step is executedin a specific amount of seconds which is called

4

“Wall Clock Time” and depends on thecomputational system used.

Such time is the total time spent by thesystem to handle both computational andInput/Output data transfer, represented forexample by the ingestion of boundary conditionsand/or output data file production. Suchactivities, even though much longer, are lessfrequent than the pure computational ones, andrepresent a minor fraction of overall simulationtime. As regards the specific RAMS behavior allthe Input/Output transfer and computational jobdistribution is performed by the special nodecalled “master” which doesn’t play any actualcomputational role and which uses only afraction of the Wall Clock Time. Table 1summarizes the typical values for a medium sizeRAMS simulation.

Table 1. Summary table of time statistics both forMaster process Time and Wall Clock Time

Typical performance values arepresented in three graphs, including the WallClock Time series (Fig. 4), Master – processWall Clock Time distribution (Fig. 5) and Node –process distribution an(Fig. 6).

In Fig. 4 a sequence of time steps ispresented, where several regimes should berecognized due to different computationalactivities. In this case the I/O time step durationis essentially linked to the hourly model outputfrequency, and, during the first period, the largeramount of time needed is due to the other LinuxSystem activities not connected with thesimulation.

Analysis of the time step distributionreveals the small spread around the averagevalues for the Master time step and the WallClock Time step as well (Figs. 5 and 6).

Fig. 4. Wall clock time series for a typical simulation,the total amount of time needed during a simulation

time step could be divided into three different classes:computational (Comp), Input/Output data transfer

(I/O), Linux operating system activities (Sys)

Fig. 5. Master time step distribution according to theirduration in seconds

Fig. 6. Wall Clock Time distribution for each simulationstep according to their duration in seconds

4. CASE STUDIES

4.1 Data and MethodsThe RAMS model is used for mesoscale

high-resolution atmospheric prediction in theMediterranean area. The NCEP/NCARreanalyses (Kalnay et al., 1996) provide itsinitialization for the historical floods whichaffected the Arno river (geography in Fig. 7).

5

(a)

(b)Fig. 7. Location of the Arno basin (dark green

boundaries) in Europe (a) and orography of the samearea (b)

The validation of RAMS forecasts,limited to precipitation (QPF), is also performedover relevant sub-basins of the Arno river(Fig. 8).

Fig. 8. Relevant sub-basins of the Arno River basin,where specific model validation was performed.

Degrees are Latitude (positive north) and Longitude(positive east)

The ECMWF operational analysesprovided the initialization for the MAP IOP2 casestudy chosen to demonstrate the performance ofsatellite precipitation assimilation.

4.2 ARTU and MUSIC Projects: PredictingFloods Affecting the Arno River

4.2.1 The hydrological model TOPKAPIRainfall-runoff models constitute the

core of most flood forecasting systems, andwhilst it is true that several new models havebeen developed in the last decades, it has beendemonstrated that, for the purposes of floodforecasting, the most important effect they mustrepresent is the dynamics of the overall soilfilling and depletion mechanisms and theresulting variation in the size of the saturatedbasin area; in fact, when large portions of thebasin reach saturation the rainfall that falls onthe catchment has a direct effect on the floodmagnitude without being attenuated by the soil'sabsorption capacity. These concepts havespawned models which are very widely usedtoday and have been extensively described inthe literature, such as the model developed byZhao R.J., (1977), by Moore and Clarke (1981),the TOPMODEL (Beven and Kirkby, 1979) theARNO (Todini, 1996).

More recently a new rainfall-runoffmodelling approach, the TOPKAPI(TOPographic Kinematic APproximation andIntegration) has been introduced (Ciarapica andTodini, 2002; Todini and Ciarapica, 2002),starting from a kinematic wave pointrepresentation of the movement of water in thesoil and on the surface (Todini,1995). Byintegrating the point differential equations overthe finite dimension of a pixel a non-linearreservoir model equation is obtained. Thecatchment behaviour can thus be reproduced bythree cascades of non-linear reservoirsrepresenting respectively, the soil, the surface,the drainage network.

The advantage of using distributedmodels, as TOPKAPI lies in the possibility ofincorporating all the Digital Elevation Model(DEM) information and using appropriate soiland surface equations. Moreover, this type ofphysically based distributed model is definitelysuperior, due to the extremely reduced need forcalibration and to the semi-distributedconceptual models, such as Sacramento(Burnash, 1973) or SSAR (Rockwood, 1961;Rockwood and Nelson, 1966; Rockwood et al,1972) or ARNO (Todini, 1996), which have beenwidely used in the past for developing real timeflood forecasting systems.

TOPKAPI does not require the classicalcalibration: parameters such as porosity andhydraulic conductivity can be established on thebasis of the soil types and land use, whileroughness for the overland flow can be relatedto the land use and the roughness in the

6

drainage network can be related to the order ofthe channel in the drainage system.

Finally, as opposed to Shetran or MikeShe (Abbott et al, 1986a,b) which require longcomputational time, TOPKAPI runs very fast,since its equations are solved analytically andthe speed of computation makes it compatiblewith operational real time flood forecasting. Forinstance, in the model of the Arno river, wherethree cascades of approximately 8000 non-linearreservoirs are solved at each time step,therefore, the computation of discharges for onefull year at hourly time steps requires only 8minutes on a standard PC.

Several applications of TOPKAPI havebeen completed chiefly for the development ofreal time flood forecasting systems: the followingtwo applications highlight the interestingproperties of the model in the case of a relativelysmall catchment with data available from adense rain gauge tele-metering network, and inthe case of a very large one using free of chargedata available on the web.

Figure 9. The new real time flood forecastingsystem for the Arno

The application of TOPKAPI to the Arnoriver aimed at forecasting floods at Florence (theNave di Rosano gauging station is just upstreamFlorence) using 1 by 1 km pixels. Fig. 9 showsthe front end of the operational real time floodforecasting system developed on the Arno river.

No calibration was applied and themodel parameter values were derived fromliterature. The drainage network was derivedfrom the Digital Elevation Model availablethrough HYDRO1K of USGS (Fig. 10); soilparameters were estimated from tables providedby the USDA on the basis of soil types (Fig. 11)and land use (Fig. 12) maps; overland andchannel roughness values were derived fromChow (1959) and from Barnes (1967) accordingto land use and channel order.

Figure 10. The 1x1 km Digital Elevation Model of theArno basin

Figure 11. The soil types map for the Arno River

Figure 12. The land use map for the Arno basin (fromCorine)

7

0

500

1000

1500

2000

2500

92/1

0/13

/3/0

92/1

0/18

/3/0

92/1

0/23

/3/0

92/1

0/28

/3/0

92/1

1/2/

3/0

92/1

1/7/

3/0

92/1

1/12

/3/0

92/1

1/17

/3/0

92/1

1/22

/3/0

92/1

1/27

/3/0

92/1

2/2/

3/0

92/1

2/7/

3/0

92/1

2/12

/3/0

Nave Rosano_ObsNave Rosano_Cal

Figure 13. Calibration results at Nave di Rosano,immediately upstream Florence

Although numerical measures of fit tests areavailable, as one can understand from Fig. 13,calibration of the TOPKAPI model wasconsidered successful and appropriate for real-time flood forecasting on the Arno basin.

4.2.2 100-year flood: November 1966At 00 UTC on November 2, 1966 a high-

pressure belt extended over a great part of theAtlantic Ocean, the British Isles, Scandinavia,and east to Russia (peak pressure of 1044 hPa).

Over central-southern Europe a lowpressure field dominated, with a relativelyorganized structure over the Gulf of Biscay. Themain cyclonic westerlies over the Atlantic weretemporarily confined to far northern latitudes. Inthe mid-troposphere a wide through extendedfrom Scandinavia to the Iberian Peninsula,associated with very cold air masses, a distinctcold advection over the Mediterranean, and ahigh-troposphere jet followed the tightestisobars. Notable is the great wave amplitude,about 40° latitude.

On November 3 the situation evolved togreater vorticity in the through, also as aconsequence of the vorticity advection by the jetstream from very northern latitudes, and thedeepening of the surface vortex. The throughaxis started rotating anti-clockwise, triggering astrong elevated heat wave over centralMediterranean and a dynamic ridge overBalkans and eastern Europe. Ascending verticalvelocity also developed west of Italy.

At 00 UTC on November 4, the northAtlantic very cold air masses reached westernMediterranean, feeding the (deepening) cyclonicvortex over central-western Mediterranean withextreme baroclinic instability. The heat flux overeastern Italy and further east reinforced, leadingto surface jets and a sudden geopotentialincrease. A blocking situation rapidly developed(1032 hPa peak pressure over Balkans).Extreme ascending vertical velocity was foundover western and northern Italy.

Starting from early afternoon onNovember 4, the surface low weakened after thebaroclinic instability flux diminished. Thermaladvection rapidly weakened too, and anelevated weak ridge grew over westernMediterranean. Vertical velocity changed signover western Italy during November 4.Such evolution is synthetically described inFig. 14.

(a)

(b)

(c)Figure 14. 500-hPa geopotential heights at 00 UTC onNovember 3 (a), November 4 (b) and November 5 (c),

1966

The precipitation which fell during theevent was absolutely extreme, reaching morethan 300 mm in Tuscany (central-western Italy)and more than 500 mm in north-eastern Italy.The rain volumes were also extraordinary.

Over the Arno basin, the highestprecipitation fell during November 3, afternoon,

8

and cumulated precipitation averaged to about120 mm in the period 12 UTC November 3 to12 UTC November 4 (Fig. 15).

Arno basin / total rainfall12 UTC on 1996/11/03 to 12 UTC on 1996/11/04

020406080

100120140160180200

13 15 17 19 21 23 1 3 5 7 9 11

hour

mm

0

5

10

15

20

25

mm

/h

Figure 15. Average precipitation over the Arno riverbasin (hourly and cumulated)

Four simulations were executed, eachone with three to four nested grids (two-waysnesting), and 26, 36 and 50 vertical levels. Therelevant features are explained in Table 2.below.

NAME START DURATION HORIZONTALRESOLUTIONVERTICAL

LEVELSG 1 G 2 G 3 G 4

LIV36Ini12 3 NOV / 12UTC 36 hours 80 16 3.2 36

LIV36Ini12G4 3 NOV / 12UTC 36 hours 80 16 3.2 1.6 36

LIV50Ini12 3 NOV / 12UTC 36 hours 80 16 3.2 50

LIV26Ini12 3 NOV / 12UTC 36 hours 80 16 3.2 26

Table 2. Summary of features of the four RAMSsimulations for the November 1966 flood

The three/four RAMS' grids are shown inFig. 16. The vertical structures of the domainsare shown in Fig. 17.

Figure 16. Horizontal domains of the RAMSsimulations of the November 1966 flood

Figure 17. Vertical domains of the RAMS simulationsof the November 1966 flood

The comparison of observed andpredicted precipitation on the finest resolutiongrids for every simulation, in the period from12 UTC on November 3rd to 12 UTC onNovember 4th are shown in Fig. 18 (patterns),Fig. 19 (basins averages), Fig. 20 (frequency in

9

intensity classes), Fig. 21 (Heidke Skill Score,e.g. Wilks, 1995).

(a)

(b)

(c)

(d)

(e)Figure 18. Validation of RAMS QPFs: comparison of

observed vs forecast rainfall distributions. (a)Observations; (b) Liv26Ini12; (c) Liv36Ini12; (d)

Liv50Ini12; (e) Liv36Ini12G4 (see Table 2)

Figure 19. Validation of RAMS QPFs: comparison ofobserved vs forecast average rainfall over Sieve and

Casrntino sub-basins and whole Arno watershed

Figure 20. Validation of RAMS QPFs over the wholeArno watershed: comparison of observed vs forecastdistribution of rainfall, partitioned into intensity classes

10

Figure 21. Validation of RAMS QPFs: Heidke SkillScore (HSS) of RAMS' QPFs

The validation vs the rain gauges provides someguidance to produce accurate QPFs:• the event was a large scale one, spanning

the Atlantic, central/western Europe,Mediterranean and northern Africa, and theprecipitation systems covered hundreds ofkm;

• the resolution increase (3.2 km to 1.6 km)produces little effect, if any;

• the increase of vertical levels (andresolution) from 26 to 36 and, to a lesserextent, to 50 (which avoids some under-estimations), produces a very significant andsometimes decisive QPF improvement;

• under-estimation is apparent, anyway mostof missing rainfall occurred in the earlysimulation period (model spin-up),suggesting that initial conditions andinitialization time are relevant, and maybethat frequent runs are suitable for accurateQPF.

The quantitative precipitation forecastsprovided by the simulation Liv36Ini12G4 (seeTable 2) over the fourth grid were used toproduce flood (discharge) forecasts at theFlorence section of the Arno river (Fig. 22) bymeans of the TOPKAPI model (Sect. 4.2.1.; alsoin Todini and Ciarapica, 2002).

Figure 22. Florence section along the Arno river

Fig. 23 shows the flood forecasts at the Florencesection, issued at different times, which assumethat rainfall forecasts are not available (i.e.assuming future zero rainfall), compared with theflood forecast based on the rainfall observed atall times. The strong dependence of the floodforecasts on the initial time is apparent, and theflood warning could be provided only about threehours in advance.Fig. 24 shows the flood forecasts at the Florencesection, issued at different times, whichassimilate the rainfall forecasts provided byRAMS, compared with the flood forecast basedon the rainfall observed at all times. Even theearliest flood forecast, issued at 18 UTC onNovember 3, could lead to an accurate floodwarning, i.e. about nine hours in advance, andsix hours earlier with regards to the forecastissued without using QPF.

0

500

1000

1500

2000

2500

3000

3500

4000

4500

66/11/0305:00

66/11/0311:00

66/11/0317:00

66/11/0323:00

66/11/0405:00

66/11/0411:00

66/11/0417:00

66/11/0423:00

66/11/0505:00

66/11/0511:00

66/11/0517:00

66/11/0523:00

66/11/0605:00

66/11/0611:00

66/11/0617:00

66/11/0623:00

Known rainfallh = 18:00h = 21:00h = 00:00

Figure 23. Flood forecasts (in m3/s) issued at differenttimes (18 UTC and 21 UTC on November 3 and

00 UTC on November 4) without the knowledge onfuture rainfall, compared to the flood wave

reconstructed at Florence by means of TOPKAPI andthe measured rainfall

0

500

1000

1500

2000

2500

3000

3500

4000

4500

66/11/0305:00

66/11/0311:00

66/11/0317:00

66/11/0323:00

66/11/0405:00

66/11/0411:00

66/11/0417:00

66/11/0423:00

66/11/0505:00

66/11/0511:00

66/11/0517:00

66/11/0523:00

66/11/0605:00

66/11/0611:00

66/11/0617:00

66/11/0623:00

Known rainfallh = 18:00h = 21:00h = 00:00

Figure 24. Flood forecasts (in m3/s) issued at differenttimes (18 UTC and 21 UTC on November 3 and

00 UTC on November 4) using the QPF provided byRAMS, compared to the flood wave reconstructed atFlorence by means of TOPKAPI and the measured

rainfall

11

It can be concluded that quantitativeforecasts provided by RAMS could lead to adecisive improvement (i.e. anticipation) of theflood warning, when coupled to an accuratehydrological model. Thus, the QPF revealsaccurate at least in the useful range for floodprediction.

4.2.3 30-year flood: October 1992A very elongated south-westerly flux

from southern north-Atlantic occurred overMediterranean and Italy on October 30th, when abaroclinically and orographically induced lowdeveloped downwind Iberian Peninsula. Cold airpenetrated far western Mediterranean during the36 hours period and the flux became morecyclonic (warm advection). Widespread verticalvelocity developed over western Mediterraneanand central-northern Italy.

Such evolution is synthetically describedin Fig. 25.

(a)

(b)Fig. 25. 500-hPa geopotential heights at 00 UTC on

October 30 (a) and October 31 (c), 1992

The precipitation which fell during theevent was very extreme, reaching more than300 mm in Tuscany (central-western Italy) andmore than 500 mm in north-eastern Italy. Highprecipitation rates are shown over most of Arnoriver basin.

Five simulations were executed, testingthe impact of initialization (00 UTC on October30th and 12 UTC on October 29th), spatial

horizontal resolution (finest resolution 3.2km and1.6 km) and assimilation of sea surfacetemperature (climatological and observed, atresolution 1 deg Lat-Lon). Details are given inTable 3. below.

Name G1 G2 G3 G4 V.L sst INILIV36I003G 80 16 3.2 36 CL 0LIV36I004G 80 16 3.2 1.6 36 CL 0

LIV36SST3G 80 16 3.2 36 OB 0LIV36SST4G 80 16 3.2 1.6 36 OB 0

LIV36SST4G_29 80 16 3.2 1.6 36 OB -24Table 3. Summary of features of the five RAMS

simulations for the October 1992 flood (Gn =horizontal resolution of n-th grid, V.L = vertical levels,

sst = sea surface temperature, CL = climatological sst,OB = observed sst, INI = initialization time: 00 =

00 UTC October 30th, -24 = 00 UTC October 29th)

The comparison of observed andpredicted precipitation on the finest resolutiongrids for every simulation, in the period from00 UTC on October 30th to 12 UTC on October31st are shown in Fig. 26 (patterns), Fig. 27(basins averages), and Fig. 28 (frequency inintensity classes).

(a)

(b)

(c)

12

(d)

(e)

(f)Fig. 26. Validation of RAMS QPFs: comparison ofobserved vs forecast rainfall distributions.between

00 UTC on October 30th 1992 and 12 UTC on October31st 1992. (a) Observations; (b) Liv36I003G; (c)

Liv36I004G; (d) Liv36SST3G; (e) Liv36SST4G; (f)LIV36SST4G_29 (see Table 3)

Fig. 27. Validation of RAMS QPFs: comparison ofobserved vs forecast average rainfall over the

Casentino and Valdichiana sub-basins and the wholeArno watershed

13

Frequency DistributionCumulated rainfall from 00UTC-30/10/1992 to 12UTC-31/10/1992 on

Arno river basin raingauges

0

5

10

15

20

25

135

Cumulated rainfall classes (mm)

Num

ber o

f rai

ngau

ges

ObservationLiv36SST3GLiv36SST4GLiv36I004GLiv36I003GLiv36SST4G_29

Fig. 28. Validation of RAMS QPFs: comparison ofobserved vs forecast distribution of rainfall, partitioned

into intensity classes

The validation vs the rain gauges providessome guidance to produce accurate QPFs:• The event was again a large scale one; no

explosive cyclogenesis occurred, contrary tothe 1996 event. Precipitation systems werewidespread, just like the vertical velocitypattern.

• The resolution increase from 3.2 km to 1.6km) produces significant and positiveeffects.

• The use of “observed” sea surfacetemperature instead of climatological ones,at the same spatial resolution (1 deg)produces relevant and positive effects,mostly at the finest resolution (1.2 km).

• The initialization "during" the event (00 UTCon October 30th) produced a significantunderestimation, with most of missingrainfall occuring in the early simulationperiod (model spin-up).

• The simulation initialized at 00 UTC onOctober 29th, 24 before the other simulationsand before the beginning of the rainstorm,produced generally very accurate forecastsin the first 18 hours of the study period (i.e.between 24 and 42 hours after itsinitialization), resulting sometimes in betterforecasts over the whole period, but showingafterwards a relevant decay in accuracy.

• Initialization time thus shows to be critical forthe success of the numerical prediction, atleast during large scale rainstorm, whichsuggests the need to pursue either animprovement of initial conditions (e.g. bymeans of diabatical initialization, seePar. 4.2) or more frequent predictions (e.g.every 6-12 hours), or both.

4.3 EURAINSAT Project: Precipitationassimilation

4.3.1 The rapid update satellite precipitationestimation algorithm

The method adopted to estimate theinstantaneous rain rate from satellite has beenproposed by Turk et al. (2000) and it is based onthe idea of blending low-Earth orbiting (LEO)microwave (SSM/I, TRMM) and geostationary(GEO) infrared data (GOES-East/West, GMS,and Meteosat) together, so to exploit theinherent advantages of each sensor.

Microwave-based imagers are suited toquantitative measurements of precipitation dueto the physical connection between upwellingmicrowave radiation and the underlying cloudprecipitation structure. On the other handgeostationary weather satellites imagingsystems provide the rapid temporal update cycleneeded to capture the growth and decay ofprecipitating clouds systems on a scale ofseveral kilometers.

The blending of LEO and GEOmeasurements allows an hourly (or less) globalrain rate analysis which avoids the spatial andtemporal coverage gaps characteristic of swath-limited LEO data. This kind of analysis is neededfor assimilation into numerical weather predictionmodels, especially for quantitative precipitationforecasting. In the rapid update (RU) algorithmtime- and space coincident microwave and IRdata are saved each time a SSM/I or TRMMpass intersects with any of the four operationalgeostationary satellites. The SSM/I rain rate iscomputed via the operational NOAA-NESDISscheme (Ferraro, 1997) at the A scan samplingspacing (25 km) of the instrument scanoperation. Once every few hours an updatecycle starts and accumulate the most recent 24hours of past coincident data.

Separate histograms of the IRtemperatures and the associated microwave-based rain rates are built in a 15° grid between60°S-60°N latitude. The grid boxes are 5° apart,providing a spatial overlap, so to assure that thestatistical relationships transition smoothly fromadjacent regions. To utilize only the most recentrain evolutionary history, the histograms areaccumulated until the percent coverage of agiven box reach a threshold (lets say 75%). Thelarger the percentage threshold, the longer oneneeds to look back in time, drawing on anincreasingly earlier stage of the cloud lifecycle.

Depending on how recently a givengeographical region was imaged by a MWsensor, the data used in some of the histogramboxes may be only a few hours old, whereasother regions may require more look-back time

14

to reach the coverage threshold. If a region hasnot reached the coverage threshold by the look-back time limit of 24 hours, the RU rain rateestimate is rendered (temporarily) unavailablefor this region. Once the statistical relationshipsTB vs rain rate are established for each 15°region, the calculation of the instantaneous rainrate for each geostationary pass isstraightforwardly accomplished by simplyscanning the available global lookup tables.

Figure 29. shows an example of theapplication of the method. The upper paneldepicts the rain rate derived by means of theFerraro (1997) algorithm for an SSM/I pass overItaly starting on 15.33 UTC on September 20th,1999. The lower panel is the rain rate derivedfrom METEOSAT-7 IR data, slot 34, coveringdata acquired from 16:00 to 16:30 UTC.

The example clearly shows that themethod, due to the recent microwave pass,established a TB-rain rate relationship suitable tofollow the evolution of the cell. Indeed theduration of the MW cloud characterization is thecritical point of the method, and stronglydepends on the specific precipitation system.

Fig. 29. a) The rain rate derived from SSM/I data andb) from the IR TB METEOSAT-7 data by means of RU

4.3.2 Precipitation assimilation in model RAMSIn RAMS the non-resolved convective

precipitation can be computed by means of theKuo parameterisation scheme (Kuo, 1974;Molinari, 1985). From the model atmosphericvariables at the base and along the column thecumulus convection module computes the watervapor convergence and then the precipitatingfraction of such water. Then the modulecomputes the convective tendencies oftemperature and moisture by means of a 1-Dcloud model.

In the inverted Kuo scheme theconvective rainfall becomes an input, whichwater vapor and convective tendencies arecomputed from.

The impact of the observed rainfallvalues in the model equations is through therates of humidity and temperature change, dueto condensation and latent heat release,according to the observed rainfall rate, in eachgrid column where the observed convectiverainfall rate is non-zero

The assimilation is performed at severaltime-steps during the earlier (typically six) hoursof the model simulation, by weighting theobserved rainfall rate with a time dependentnudging function (which can be made dependenton observation errors too).

The observed rainfall partitioning in non-resolved convective precipitation and resolvedone is a tricky problem. Ancillary data (e.g.lightning) are generally difficult to be interpretedand to be managed in an operational context. Inthe work it is assumed that the resolved raingiven by the model computation is reliableenough to be subtracted from the satellite totalrainfall, in order to obtain the non-resolvedconvective quantity (negative resulting valuesbeing moved to null ones). Satellite estimationerrors can be taken into account in this processby weighting the subtraction terms.

A very simple scheme of the procedureis represented in Fig. 30.

| |

| |

t0+6h|

|

Control Run

Assimilation Runof satellite rainfallestimationperiod of

assimilation

t0 t0+24h

Fig. 30. Simple scheme of the rainfall assimilationprocedure

4.3.3 Case study: MAP IOP2 (19-21 September1999). Synoptic meteorology and verification ofthe satellite precipitation estimation

The meteorology of the case study wascharacterized by a deep cyclone on the Atlanticwest-south-west of British Isles, with a very

15

elongated baroclinic front approaching westernMediterranean and France on September 19th1999, anticipated by a southerly flow,intensifying on September 20th from interiornorthern Africa and producing a convergence tonorthern Italy and Alps from , where it convergedfrom the seas surrounding the peninsula (south-west and south-east). The pre-frontal and frontalprecipitation was widespread. The focusing andchanneling of the flow, together the advection ofthe strongly unstable elevated mixed layer fromSahara, produced locally intense rainstormsover the main ridges of the Italian Appenninesand Alps and the respective foothills, especiallyover the western portions of such mountains.

The patterns of geopotential at 500 hPaand Sea Level Pressure at 00 UTC onSeptember 19th and 20th are shown in Fig. 31.

(a)

(b)

(c)

(d)Fig. 31. Synoptic meteorology of the case study. (a)and (b): 500 hPa geopotential height and sea level

pressure, respectively, at 00 UTC on September 19th;(c) and (d): 500 hPa geopotential height and sea levelpressure, respectively, at 00 UTC on September 20th

On September 21st the northernexpansion of a dynamical anticyclone fromnorthern Africa produced the eastward motion ofthe Atlantic cyclone and the weakening of thepositive vorticity over central and westernMediterranean, until on September 22nd theanticyclonic curvature of the flow, the increasingsea level pressure and mid-troposphericgeopotential marked the rapid increase ofstability (not shown).

The patterns of geopotential at 500 hPaand Sea Level Pressure at 00 UTC onSeptember 21st and 22nd are shown in

Fig. 32.

(a)

(b)

16

(c)

(d)Fig. 32. Synoptic meteorology of the case study. (a)and (b): 500 hPa geopotential height and sea level

pressure, respectively, at 00 UTC on September 21st;(c) and (d): 500 hPa geopotential height and sea levelpressure, respectively, at 00 UTC on September 22nd

In order to validate the algorithm ofrainfall estimation from satellite data (samplepicture in Fig. 33), it was performed acomparison with data from a rain gaugesnetwork. A comparison with data from a networkof rain gauges was performed by consideringnine hydrographic basins in the area covered bythe rain gauge network. The average rainfallvalues, computed from the data of all the raingauges corresponding to each one of the abovehydrographic basins, were compared with thecorresponding average of the satellite retrievedrainfall values. An excellent agreement wasfound for four out of the nine analyzed basins.

Fig. 33. Sample picture of the satellite estimatedrainfall

In such basins both the temporal phaseof the event and the total cumulated rainfall arewell retrieved by the algorithm. For one of thefour basins an excellent accordance is shown inthe temporal phase, but a significantdiscrepancy in the cumulated rainfall.

For the three other basins, we have aninsufficient time alignment and strongdiscrepancies in the cumulated rainfall.

In Fig. 34 (a) and (b) it is reported thetime evolution of the observed and satelliteestimated cumulated rainfall for the ValdarnoInferiore basin, where an excellent agreement isshown, in the period from 00 UTC on September20th to 00 UTC on September 22nd.

Valdarno Inferiore

0

10

20

30

40

50

60

70

1 4 7 10 13 16 19 22 2 5 8 11 14 17 20 23 3 6 9 12 15 18 21

hours

Cum

ulat

ed R

ainf

alls

(mm

)

RaingaugesRainfall Estimation

(a)Valdarno Inferiore

02468

101214161820

1 4 7 10 13 16 19 22 2 5 8 11 14 17 20 23 3 6 9 12 15 18 21

hours

Hou

rly R

ainf

alls

(mm

)

RaingaugesRainfall Estimation

(b)Fig. 34. Validation of the satellite rainfall estimationfor the Valdarno Inferiore basin in the period from

00 UTC on September 20th to 00 UTC on September22nd. (a): time evolution of cumulated rainfall; (b): time

evolution of hourly rainfall

17

The performance of the satelliteestimation algorithm appears to be excellent, inthis case study, for the coastal areas whereorography is particularly smooth, while it showslarge deficiencies (not shown) for steeporography areas, mainly due to known problemsaffecting the microwave rainfall retrievals overthe land surface and the lack of orographicalcorrections in both the microwave and infraredmodules.

Model RAMS was run for 48 hours since00 UT on September 20th over a large domaincovering central-western Mediterranean andEurope at 20 km spatial horizontal resolution.Fig. 35 shows the domain with a sample hourlyprecipitation field superimposed.

Fig. 35. Domain of model RAMS and sample hourlyprecipitation field after 12 hours of simulation

Two simulations were executed: acontrol run without assimilation of the satelliteestimated rainfall, and a dynamical satelliteassimilation run in which the assimilation isperformed.

Fig. 36 shows the basins for which therainfall forecasts produced by the twosimulations executed with RAMS model arevalidated.

Fig. 37 shows the results of thevalidation for four basins.

Sieve

Fig. 36. Hydrographic basins used for the validationof the rainfall forecasts produced by the RAMS model

18

Fig. 37. Validation of RAMS QPFs: comparison ofobserved vs forecast average rainfall over the

Valdarno Inferiore, Serchio, ValdiChiana and ValdarnoMedio sub-basins

The main results of satellite precipitationassimilation into RAMS can be summarized inthe following points:

• The convective rainfall assimilationtechnique correctly converges.

• The impact of convective rainfall assimilationon simulations started during convectionevents is relevant.

• Assimilation improves spin-up phases andrainfall forecasts up to several hours.

• Positive effects of assimilation are generallyfound both on the Kuo - parametrized andthe resolved parts of the rainfall forecastedby the model.

• The improvements are apparent severalhours after the end of the assimilationperiod.

Other experiments are anyway necessary toevaluate the impact of the precipitation

assimilation in different synoptic and mesoscalemeteorological conditions .

5. DISCUSSION AND CONCLUSIONS

This work describes the recentimprovements in a regional meteorologicalprediction system based on model RAMS,mainly aimed at quantitative precipitationforecasting.

The sensitivity of the prediction quality toseveral key components is analyzed, comprisinghorizontal and vertical resolution, initializationtime, quality of the assimilated sea surfacetemperature and precipitation assimilation.

Special emphasis is given to the lowcost high-performance computer system, whichallows to pursue the improvements whilepreserving operational simulation times, and tothe coupling to a hydrological forecasting systemin view of early flood warnings. Theimprovement of such warnings also providesome more information on the quality of theatmospheric model forecasts.

Some guidance is derived as regardsthe production of accurate quantitativeprecipitation forecasts, which constitute a soundbasis for future work and developments.

6. REFERENCES

Abbott, M.B. et al., 1986a: An introduction to theEuropean Hydrological System - SystèmeHydrologique Européen, "SHE", 1: History andphilosophy of a physically-based, distributedmodelling system. J. Hydrol., 87, 45-59.

Abbott, M.B. et al., 1986b: An introduction to theEuropean Hydrological System - SystèmeHydrologique Européen, "SHE", 2: Structure ofphysically-based, distributed modelling system.J. Hydrol., 87, 61-77.

Barnes H.H., 1967: Roughness Characteristicsof Natural Channels. Geological Survey Water-Supply Paper 1849. US Gov. Printing Office,Washington.

Beven, K.J. and Kirkby, M.J., 1979: A physicallybased, variable contributing area model of basinhydrology. Hydrol. Scien., 24, 1-3.

Burnash, R.J.C., Ferral, R.L. & Mc Guire,R.A.,1973: A General Streamflow SimulationSystem - Conceptual Modelling for Digital

19

Computers, Report by the Joint Federal StateRiver Forecasts Center, Sacramento.

Chow Ven Te, 1959: Open Channel Hydraulics.McGraw-Hill Book Company, Inc., KogakushaCompany, Ltd., Tokyo.

Ciarapica L., Todini E., 2002: TOPKAPI: amodel for the representation of the rainfall-runoffprocess at different scales. HydrologicalProcesses, 16(2), 207-229.

Ferraro, R. R., 1997: Special sensor microwaveimager derived global rainfall estimates forclimatological applications, J. Geophys. Res.,102, D14, 16715-16735.

Kalnay, E. and Coauthors, 1996: TheNCEP/NCAR Reanalysis 40-year Project. Bull.Amer. Meteor. Soc., 77, 437-471.

Kuo, H. L., 1974: Further studies of the influenceof cumulus convection on large-scale flow. J.Atmos. Sci., 31, pp 1232-1240.

Meneguzzo, F., Menduni, G., Maracchi, G.,Zipoli, G., Gozzini, B., Grifoni, D., Messeri, G.,Pasqui, M., Rossi, M., and C.J. Tremback, 2002:Explicit forecasting of precipitation: sensitivity ofmodel RAMS to surface features, microphysics,convection, resolution. In: MediterraneanStorms. 3rd Plinius Conference 2001. Ed. by: R.Deidda, A. Mugnai, F. Siccardi. GNDCI Publ. N.2560, ISBN 88-8080-031-0, 79-84.

Mission Research Corporation (MRC)/*ASTERDivision, 2000: RAMS. The RegionalAtmospheric Modeling System. TechnicalDescription. [Available at http://www.aster.com].

Molinari, J., 1985: A general form of Kuo’scumulus parameterisation, Mon. Wea. Rev.,113, pp 1411-1416.

Moore, R.J. and Clarke, R.T., 1981: Theprobability-distributed principle and runoffproduction at point and basin scales. Hydrol.Sci. J., 30(2), 273-297.

Pielke, R. A. & Coauthors, 1992: Acomprehensive meteorological modelingsystem-RAMS. Meteor. Atmos. Phys., 49, 69–91.

Rockwood, D.M., 1961: Columbia StreamflowRouting by Computer. Trans. ASCE, Paper n.3119.

Rockwood, D.M. and Nelson, M.L., 1966:Computer application to Streamflow Synthesisand Reservoir regulation. Proc. IV Int.Conference on Irrigation and Drainage.

Rockwood, D.M et al., 1972: User Manual forCOSSARR Model, US Army EngineeringDivision North Pacific. Portland Oregon.

Todini, E., 1995: New trends in modelling soilprocesses from hillslope to GCM scales, in TheRole of Water and the Hydrological Cycle. In:Oliver, H. R. and Oliver, S. A. Global ChangeNATO ASI Series, Series I: GlobalEnvironmental Change, 31, 317-347.

Todini E., 1996: The ARNO Rainfall-Runoffmodel, J. Hydrol., 175, 339-382.

Todini, E. and Ciarapica L., 2002: The TOPKAPImodel. In Mathematical Models of LargeWatershed Hydrology. V. P. Singh, D. K. Frevertand S. P. Meyer (eds.), Water ResourcesPublications, Littleton, Colorado, Chapter 12,471-506.

Tremback, C. J., 1990: Numerical simulation ofa mesoscale convective complex: modeldevelopment and numerical results. Ph.D.dissertation, Atmos. Sci. Paper No. 465,Colorado State University, Dept. of AtmosphericScience, Fort Collins, CO 80523.

Turk, F. J., G. D. Rohaly, J. Hawkins E. A.Smith, F. S. Marzano, A. Mugnai, and V.Levizzani, 2000: Meteorological applications ofprecipitation estimation from combined SSM/ITRMM and infrared geostationary satellite data,In Microwave Radiometry and Remote Sensingof the Earth’s Surface and Atmosphere, P.Pampaloni and S. Paloscia Eds., VSP Int. Sci.Publ., 353-363.Walko, R. L., W. R. Cotton, M. P. Meyers, and J.Y. Harrington, 1995: New RAMS cloudmicrophysics parameterization. Part I: Thesingle-moment scheme. Atmos. Res., 38, 29–62.

Walko, R. L., & Coauthors, 2000: Coupledatmosphere–biophysics–hydrology models forenvironmental modeling. J. Appl. Meteor., 39,931–944.

Wilks, D.S., 1995: Statistical methods in theatmospheric sciences: an introduction.Academic Press, 464 pp.Zhao, R.J., 1977. Flood forecasting method forhumid regions of China. East China College ofHydraulic Engineering, Nanjing, China.