Ezio Todini

Introduction

The problem of flooding is as old as time. However, while natural flooding of large areas did not create situations more dangerous than others in a prehistoric world, the expansion of human activity and the aggregation of people in larger and larger urbanised areas has made more and more preventing damages caused by floods as well as control and management of flood waters a problem of vital necessity.

From the end of the eighteen century onwards, with the advent of the industrial age, there have been two courses of action: hydraulic works on the territory, such as land reclamation works, which in many cases upset a land's equilibrium based on overflow, and the channelling of watercourses, especially in mountain and foothill areas, with the result that the problem of flooding is brought downstream even to areas that were originally protected. In addition, recent years have seen booming population and indiscriminate urbanisation create extremely dangerous situations, with floodplain areas that are inhabited, or with houses built at the foot of dikes, where the safety tends to vanish during prolonged periods of flooding.

The problem of flooding is a global problem that has been increasing at a worrisome pace in recent years. A brief analysis of the following tables, which respectively show trends in flood-related damage (Figure 1, in billions of US dollars), number of people stricken (Figure 2, in millions), and number of casualties (Figure 3, in thousands), provides an idea of how serious the situation is.

Fig. 1 - Flood-related damage world-wide, in billions of US dollars

Fig. 2 - Number of people stricken by floods world-wide, in millions

Fig. 3 - Flood-related casualties world-wide, in thousands

Damage is expected to rise inexorably in the years to come, partly due to grater risks posed by larger urbanised areas, and partly due to climatic changes taking place. The considerable amount of carbon dioxide in the atmosphere, which is produced mainly by the combustion of hydrocarbons, exacerbates the greenhouse effect increasing the Earth's capacity to retain the long-wave radiation that it produces as a black body subjected to solar radiation. This radiation cannot be released into space, thereby upsetting the balance of energy trapped by the Earth and increasing the amount of available heat. Without entering into the long-standing dispute on the generalised effects of this increased availability of energy and heat, it must be recognised that increased available energy leads to greater evaporation (and therefore a greater quantity of water vapour available for rain) as well as to greater atmospheric vorticity with increased space-time variability of rain frequency. In other words, the greenhouse effect will result, at a specific location, in longer and more frequent rainy and/or drought periods, and rainfall and discharge extremes will be accentuated, and what is shown in figures 1, 2, 3 tends to confirm this hypothesis.

To deal with floods, structural and non-structural measures have to be implemented in order to reduce the risk of overflowing river banks and dikes. Structural measures cannot eliminate completely the risk given the impossibility of building larger and larger structures to cope with extremely low probability events. Therefore, an important role is left to non-structural measures to be compared, evaluated and actuated in real time, which implies the need of accurate flood forecasts with a sufficient time in advance to allow for their implementation. This is why decision makers have become increasingly interested not only in flood forecasting, but also in decision support systems that present in a synthetic and graphical form the alternative choices and the evaluation of the expected damages or benefits arising from their implementation. To be of any practical value, these expert systems must allow for the simulation of alternative management policies under the uncertain evolution of natural events.

In the field of flood risk mitigation and control the present state of the art allows for the development of reliable rainfall-runoff models as well as one or two dimensional hydraulic models, but in general forecasts are strongly affected by the knowledge of the uncertainty associated with rainfall measurements and particularly with future rainfall extrapolations. In addition, between Meteorologists and Atmosphere Physicists on one side and Hydrologists, Soil Physicists and Engineers on the other side a common language has not yet developed, each looking at the forecasting problem with ideas and postulates largely reflecting his own traditional approaches, with very little interest on satisfying the needs of the others, or at least trying to find together what can and what cannot be done. The same difficulties also exist between Scientists and Decision Makers, which most of the time prevents the definition and the timely implementation of operationally useful actions.

An operational real time flood forecasting system can be a complex system according to the actual needs of forecasting lead time and to the size and complexity of the system to be monitored and controlled. In order to analyse the actual requirements of a real time operational flood forecasting system one must consider all the following components:

- a precipitation forecasting model (deterministic and/or stochastic);

- a catchment model (deterministic and/or stochastic);

- a flood routing model;

- a flood plain model;

- a Geographical Information System (GIS);

- a geo-referenced Data Bank;

- an Expert System shell.

The data acquisition system and the precipitation forecasting models will be described in the following sections prior to the presentation of a research project aimed at studying the feasibility and the level of aggregation needed for the implementation of the envisaged real-time flood decision support system

Precipitation Data acquisition systems

At present, there are essentially three basic systems for providing precipitation measurements, which can be used for real time flood forecasting.

The most commonly and widely used rain sensors for developing operational real- time flood forecasting systems are the conventional ground based tele-metering raingauges, generally linked to a central station by means of telephone lines or by radio links (VHF or UHF) or, less frequently, by means of Meteor-burst equipment or via satellite through Data Collection Platforms (DCPs). There are several reasons in favour of the conventional equipment based upon raingauges. Firstly National Services have a long tradition in using raingauges, which means that long historical records are generally available for calibrating the rainfall-runoff models; secondly, in real time flood forecasting there is also need of other ground based hydro-meteorological measurements, such as for instance water levels in rivers and air temperatures close to the soil, which sensors may be integrated into the overall data acquisition system so that the cost of the additional rain sensors becomes marginal. Finally, in developing countries, training of local personnel and maintenance result more technically and economically accessible with ground based equipment rather than with other sources such as weather radars or satellites. The density of raingauge networks depend on several factors (WMO, 1981) and must be determined specifically for a single case depending upon the orography and the spatial correlation of observations. Techniques such as Factor Analysis or Kriging are generally used to provide an economically viable but sufficiently dense measurement network. The spatial description of the rain field based upon rainguges may not be accurate at very small scales (i.e. 100•100 m2), but tends to be sufficiently accurate, for flood forecasting purposes, on larger scales of the order of 10.000•10.000 m².

Another precipitation measurement system is the weather radar system, which importance has grown in the last decade, particularly after the introduction of the dual polarisation systems and the Doppler radars. There are now over 100 countries operating more than 600 weather radars and development programmes have been established in several countries as well (Rosa Dias, 1994). The European Union sponsored COST72 (1985) and COST73 (Collier C.G., 1990) for establishing a weather radar network in the participating countries. In the USA NEXRAD (1984) is a programme for establishing a network of 175 S-band Doppler weather radars and in the UK the FRONTIERS programme combines radar and METEOSAT images to produce very short precipitation forecasts (Browning K.A., 1979; Browning and Collier, 1982). There are two major benefits in using radars: the first one is a finer spatial description of the precipitation field and the second one lies in the possibility of observing approaching storms sometimes before arriving over the catchment of interest. A major disadvantage lies in the need for recalibration of parameters used for converting reflectivity to rain, which generally also requires the installation of a conventional ground based raingauge network.

The third potentially useful measurement system is based upon the analysis of clouds shown by the geo-stationary satellite images (Milford and Dugdale, 1989). This approach has been successfully used for the development of the Nile Flood Early Warning System (Grijsen et al., 1992). Unfortunately, the methodology developed by the Dept. of Meteorology of the University of Reading is adequate for long time intervals and very large catchments (the Blue Nile catchment area is on the order of 500.000 km²) in the Tropical region, while there have been no convincing applications of the technique for smaller catchments in the sub-tropical or in the temperate zones.

When planning an operational flood forecasting system, there is the need of choosing among these precipitation measurement systems which have extremely different characteristics from the point of view of the information content. This requires an analysis of the size and of the nature of the problem which may start from giving an answer to the following question: "When is the improved spatial description of precipitation provided by radar an essential requirement?". The answer is: "For very small size catchments, and in particular for small mountain or urban catchments". For these catchments in fact, given the very small sampling time required for runoff forecasting (5' - 15') one radar image may be viewed as a snapshot of the spatial distribution of rain while, in order to obtain the same result an extremely dense raingauge network would be required. On the other hand, when the flood forecast is needed for small (> 100 - 200 km²) or medium size (1000 - 20.000 km²) catchments the measurements provided by the conventional raingauge networks tend to be more than adequate.

This statement may appear too strong, but a number of considerations should be pointed out. First of all the problem here considered is real-time flood forecasting; different considerations would apply if dealing with water balance problems for agricultural practices or with problems of soil erosion and diffusion of non-point pollution sources.

A wide number of operational flood forecasting systems are based upon the conventional rain-gauge networks and are properly operating, as one can see from the Table 1, showing extremely high correspondence between the river flows measured and computed using raingauge precipitation measurements as input (Todini, 1995a). These performances depend on the fact that the rainfall-runoff process can be fairly well modelled at scales of the order of magnitude of 200-300 km² and at these scales the spatial integral of rain can be adequately represented through a reasonably dense raingauge network. One reason for this improved performance of the raingauges is due to the increased length of the time sampling interval (for instance 1 hour) required by the larger spatial scale, which tends to increase the spatial correlation of rain measurements at the different gauges as a consequence of the storms movement, thus reducing the overall variability.

Table 1 - Catchment size and quality of rainfall-runoff model performances, expressed in terms of explained variance, using conventional raingauge networks.

Recently several distributed rainfall-runoff models have been developed, such as for instance Système Hydrologique Européen - SHE (Abbot et al., 1986a,b) or the original distributed version of TOPMODEL (Beven and Kirkby, 1979), that raised the interest of meteorologists given the possibility of using the distributed type models in conjunction with radar data for flood forecasting. Unfortunately, these models are essentially useful for the simulation of the "internal" characteristics of the catchments (local water balances, soil erosion etc.) or for detailed analysis, but are not particularly geared for running in real time. A lumped version of the TOPMODEL , the TOPSIMPL (Obled, 1995, personal comm.), has in fact been recently developed to be used as a "lumped conceptual model" for real-time flood forecasting problems.

A second question to be dealt with relates to the time sampling interval needed for real time flood forecasting, which ranges from the 5' - 15' needed in urban or steep mountain catchments, to 1h for a wide variety of catchments from 200 to 1.000 km2 (or sub-units into which larger catchments may be sub-divided), to 3h for very large catchments (20.000-30.000 km2 or more). It is not advisable to operate on larger time steps because of the high non-linearities in the soil response to precipitation. The time sampling interval must in fact be substantially smaller than the time required to fill the soil and reach saturation and it is a function of both rain intensity and soil moisture capacity.

This aspect brings about one of the major communication problem between Meteorologists and Hydrologists. Traditionally, Meteorologists use to sample precipitation at daily intervals or at most every 6 hours. Unfortunately the sampling interval of 6 hours is rarely adequate for real time flood forecasting; this is not difficult to be understood considering that a total of 120 mm of rain over a six hours period could result from 120 mm in only one hour or from 20 mm of rain every hour during the six hours interval, thus causing a totally different soil response. It is true that the overall behaviour of a large catchment appears as a smoothed quasi-linear response function with time to peak and time of concentration much larger than six hours, which originated the idea of the Unit Hydrograph approach (Sherman, 1932). Unfortunately linear or non-linear response function models tend to model only surface runoff regardless to the initial status of the soil moisture, which has been demonstrated to play a major role in flood formation (Beven and Kirkby, 1979; Todini, 1995b). Although multiple unit hydrograph piecewise linearised models such as the CLS (Natale and Todini, 1976a,b; Todini and Wallis, 1977) or the SCLS (Wang et al., 1987) have tried to overcome the problem, nevertheless all the recent approaches have shown that a considerable gain can be obtained with models that tend to compute the dynamical variation of the saturated areas, which can be obtained by soil moisture accounting models such as for instance the TOPMODEL and the ARNO model (Todini, 1995a). The quality of results, in terms of greater stability of the efficiency measures in split sample tests, is given by the possibility of verifying the behaviour of the model not only during flood events, but also on long continuous records, keeping track of all the small and medium flood events as well as of the overall water balance. These continuous conceptual type models are based on simplified physical processes and must therefore be run with the appropriate time intervals to represent the rain-soil moisture saturation process, which implies time intervals smaller than six hours.

Two final remarks on the validity of the different data acquisition systems to be used in real time:

- more than the use of a unique rainfall data acquisition system, one should consider the possibility of using a combination of the different systems comprising ground based raingauges, weather radars and possibly, in the nearby future, satellites. This, a part from providing complementary information, will also improve the reliability of the overall flood forecasting system with the redundancy required to overcome failure problems, frequently appearing during storm events;

- it must be acknowledged that improvements in real time flood forecasting are not to be searched in the quality of the rainfall measurements (which can already be improved to the required degree of approximation) but rather in the possibility of accurately forecasting rainfall for the next 6 to 12 hours to allow for the analysis and the timely implementation of the warning or of the flood control measures.

THE RAINFALL FORECASTING PROBLEM

Rainfall forecasts aimed at generating effective and timely real time flood forecasts must be provided at the scale of the sub-units into which the catchment is being divided (~ 200-300 km2) sampled in time at the appropriate time interval (generally 1 hour) and extended for a lead time of at least 12 hours. This is obviously what Hydrologists would require for transforming precipitation forecasts into operational flood forecasts. In addition, Hydrologists pressed by Decision Makers would also like to obtain a measure of the forecasting uncertainty in order to incorporate it into the decision making process. Unfortunately at the present stage, these requests have not yet been entirely satisfied with the appropriate degree of reliability. This is one of the fields where research should concentrate in the nearby future.

There are several ways of approaching the precipitation forecasting problem, which may be roughly categorised into stochastic and deterministic. Traditionally, Hydrologists are more familiar with the stochastic models, while Meteorologists tend to determine precipitation either on the basis of images extrapolation (radar and satellite) or by means of atmosphere dynamics models. A detailed description of available stochastic models is given in Foufoula-Georgiou and Georgakakos (1991). More recently Modified Turning Band (MTB) approaches have been introduced which are able of reproducing observed rain field patterns (Mellor, 1995; Mellor and O'Connell, 1995; Mellor and Metcalfe,1995) and Bayesian coditional probability models have been studied for incorporating in the forecasts the most recent rain patterns (Todini and Di Bacco, 1995). In a recent paper Collier (1994) describes objective methods of forecasting precipitation using weather radar data, essentially based upon several image treatment techniques and advocating the use of expert systems. On the other hand Collier does not seem to be to much attracted by the use of numerical dynamical models to derive wind fields for improving rain forecasts. The stochastic models, are generally able of providing forecasts extended to the 12 hours required lead time, but the variability of the precipitation traces tends to be quite large, and consequently the forecasts are of little practical value, while the techniques based upon the analysis of radar images can hardly provide reliable quantitative precipitation forecasts beyond one or two hours, which can be extended up to 6 hours as in the case of the FRONTIERS by analysing a larger ensemble of available data.

In several countries weather forecasts based upon Mesoscale Atmospheric Models are available which generally extend the forecast horizon to 72 hours. Unfortunately the precipitation forecasts provided by these models are not usable within the frame of rainfall-runoff models, given the size of the Mesoscale model meshes which, in the best cases are of 50•50 km2, and consequently the average precipitation over the mesh is totally unrealistic at the sub-catchment scale. More recently a new category of atmospheric models based upon primitive equations was introduced, the so called Limited Area Models or LAMs (Lazic and Telenta, 1990). LAMs are linked to the Mesoscale models in terms of boundary conditions and allow for an improved orographical representation as well as for finer grid modelling. These models, generally based upon hydrostatic assumptions in the vertical dimension, have a theoretical minimum of 10•10 km2 meshes, while for further reducing the mesh sizes, non-hydrostatic models are still under development. LAMs are able of extending in time precipitation forecasts for the required 24 hours, but the quality of forecast in terms of amount of precipitation is still to be assessed. Recent non conclusive experiments (Todini, 1995c) showed that the timing and the precipitation pattern is well described, while the precipitation amounts can be missed, at a specific location, by orders of magnitude. Everybody agrees on the need of post-processors to be applied to the precipitation output of a LAM, but a better way would be an improved local calibration of he model and the introduction of the possibility of using the precipitation fields provided by the available measurement systems (raingauges and radars) to correct, during the data assimilation phase, the state variables of the model.

After briefly enumerating the available techniques presently available for producing quantitative rainfall forecast aimed at providing reliable real time flood forecasts it must be acknowledged that the problem is still in a development phase, with a number of options which seem more attractive than others and where research is still needed to produce operationally useful results. Bearing in mind that the final objective is planning and decision under the uncertainty of the future natural events, it appears that additional research is also required to assess the ways in which the precipitation forecast uncertainty could be accounted for to better estimate the uncertainty associated with the flood forecasts.

THE AFORISM PROJECT

From what was presented in the previous sections it clearly appears that although good quality precipitation measurements are available together with rainfall-runoff models which can provide reliable flood estimates, the major requirements of a decision maker are, on the one hand, improved quantitative precipitation forecasts and, on the other hand, the integration of all the sub-systems into a versatile and comprehensive decision support system. In 1990 a research project was funded by the Commission of the European Communities, within the frame of the funding programme EPOCH with the aim of evaluating the possibility of such integration. The project, which title was "A Comprehensive Forecasting System for Flood Risk Mitigation and Control" (AFORISM) (Todini, 1995c) aimed at verifying different approaches in rainfall modelling as well as in rainfall-runoff modelling and their integration in a decision support system both for planning and for real time flood forecasting and management. In order to do so a feasibility study, based upon the Reno river, aiming at integrating all the innovative technologies in an operational decision support tool for flood forecasting and flood impact analysis was carried out by the following research groups:

1) University of Bologna (I);

2) Regional Meteorological Service of Emilia Romagna (I);

3) University College Cork (EI);

4) University of Newcastle Upon Tyne (UK);

5) National Technical University of Athens (GR);

6) Instituto Superior de Agronomia (P);

7) Institut National Polytechnique de Grenoble (F);

8) Ecole Polytechnique Federale de Lausanne (CH).

The main objectives of AFORISM have been:

(i) To study the level of aggregation and the required interfaces to set up a comprehensive flood forecasting scheme which uses radar data, telemetering rain-gauge network data as well as the ECWMF model results, to perform a dynamic stochastic forecast of future rainfall traces, by means of stochastic precipitation models as well as of a Limited Area Model.

(ii) To compare a number of different rainfall-runoff models ranging from the extremely simplified event models, through the continuous lumped semi-distributed models to the complex distributed differential models, in view of their inclusion in the forecasting and management system of objective (i) and the possibility of improving the representation of catchment behaviour.

(iii) To use the Quantitative Precipitation forecasts in conjunction with rainfall runoff models and 1D-2D flood routing models to perform real time flood forecasting in flood prone areas

(iv) To assess the impact of alternative management scenarios, by means an Expert System, combining the results of the flood plain model with socio-economical information stored in a Geographical Information System (GIS).

(v) To disseminate the results to all EU agencies mainly involved in flood risk mitigation and control.

As previously mentioned, in order to assess the capability of integration of all the components a common experiment was launched on the Reno river (~4000 km2) which topographical computerised maps were made available by the Regione Emilia Romagna. Three years (1990-1992) of continuous records for precipitation (48 raingauges), temperature (15 thermometers) and water levels (22 level gauges) were provided to the project by the Servizio Nazionale Idrografico e Mareografico. Most of the data were only available in graphical form, which required long time consuming digitalisation. After digitalisation the data were checked and validated by comparison with the original graphs; in addition the recorded data were sampled at one hour sampling interval and all the missing data were reconstructed by means of a Kalman Filter approach (Todini, 1993).

Fig. 4 - The Reno river in Italy

At the Emilia Romagna Meteorological Service, a study was conducted which main task was the improvement of the rainfall estimations combining the available doppler radar precipitation data with those provided by the raingauge network in the frame of an objective analysis scheme.

The MTB model was fitted to five storm events which were observed on the Reno catchment around the time of the only major flood in the records available for this study. The fitting was done in such a way that the depth, time of arrival, and duration of these storms is reproduced, on average, by realisations of the model. This means the production of 'alternative' storms which, although never having happened in practice, may with equal probability have happened instead of the five observed events. It was therefore possible to produce an ensemble of flood-producing rainfall scenarios to be used to investigate the effects of incomplete sampling of rainfall fields in the domain of the Reno catchment. The physically-based distributed catchment modelling system SHE was calibrated on the Upper Reno catchment. The available digital elevation model was used to derive grid square elevations, and river widths and depths were derived from limited information on channel sizes. The soil types, land coverage and vegetation types have been discretized into meaningful classes, and these have been distributed accordingly over the grid elements also. The grid size itself (originally available on 200 metre resolution) was reduced to 800 metre resolution, and covering an area of about 40 by 60 kilometres, comprises 2196 computational elements. Monte-Carlo experiments were then performed with the MTB and SHE models whereby a large number of synthetic rainstorm events were sampled according to hypothetical rain gauge networks consisting of 3, 6, 12, 18, and 24 rain gauges distributed approximately evenly about the catchment, and the rainfall field was also sampled on a grid of the same spatial resolution as the catchment model. All the storms, in all the sample configurations, were then input to the SHE model and the resulting runoffs assimilated. These results have been analysed in order to determine the nature and magnitudes of the errors (i.e. differences in runoff arising from the same storm but with different sampling arrangements).

Different rainfall-runoff models to be used for real time flood forecasting were compared and a general consensus was reached on the need of simple models (thus avoiding the complex distributed physically based ones). In addition, due to the difficulty of estimating the initial soil moisture conditions, continuous type models were chosen, also given the availability of continuously recorded real time data. The ARNO model and the TOPMODEL, where analysed as possible candidates and were calibrated using the 1990 rainfall and runoff data. The final choice was for the ARNO model for two reasons. The first one relates to the more uniform performance of the model in verification periods, and the second one relates to the fact that the ARNO model is already imbedded into an automatic real-time flood forecasting system, known as the European Flood Forecasting Operational Real Time System (EFFORTS), which was developed within the frame of a R&D project funded by the CEC and has been extensively used in practical applications, nowadays being operational on several rivers (Fuchun, Po, Arno, Tiber, Adda, Oglio, Danube) (ET&P, 1992). Calibration of the ARNO rainfall-runoff model was successfully obtained by using the historical record (1990-1991), while in order to calibrate the flood routing model for the Reno river, a number of problems had to be solved. In particular the rating curves provided by the Hydrographic Office in Bologna needed verification: this was done by establishing a measurement campaign on the upstream cross sections, while using an original technique for estimating the downstream ones. The cross sections available for the river Reno were used for calibrating the flood routing Parabolic and Backwater (PAB) model selected (Todini and Bossi, 1986). The results of calibration, although reasonably good showed the inadequacy of the available cross sections. The Reno River Authority was informed and the operational decision was taken of performing, in due time a new measurement campaign. As previously stated the results were good enough for the purpose of the project and this calibration was then used within the EFFORTS package.

In addition, the setting up of the real-time flood forecasting system required the following operations to be performed:

- Digitisation of the basin map and measurement point location;

- Structuring of the necessary system definition files;

- Calibration of the real time automatic data reconstruction models for precipitation, temperature and level stations;

- Calibration of the Kalman Filter based rainfall extrapolation models;

- Calibration of the Kalman Filter based real time updating noise models.

The entire set of operations was successfully performed bringing about a system which can already be used for operational flood forecasting in real time.

In order to analyse the effects of possible real time decisions a 2-D model was then calibrated and used in combination with dike failure hypotheses, in order to simulate the inundation event of November 26th, 1990. The model, which is based upon the Control Volume Finite Elements approach (Di Giammarco et al., 1995), was originally developed at Bologna University. Within the frame of AFORISM, the package was then modified in order to allow for breaklines (such as dikes, roads etc.), as well as for other hydraulic specific elements which are currently encountered in practical applications in collaboration with the University of Lausanne. The model was calibrated by comparing the actual flood map with the model results and the final impact results were demonstrated in an animated succession by means of a GIS (Di Giammarco et al., 1994) interacting with available socio-economical data.

A joint experiment was then set up in collaboration with the Emilia-Romagna Meteorological Service group. Nine quantitative precipitation forecasts up to 24 hours were generated on the basis of an operational Limited Area Precipitation model with 10 x 10 km2 grid (Todini, 1995c). The forecasts were aggregated at the scale of the subcatchments used in the flood forecasting model and compared, for the first 12 hours, with the precipitation calculated by using the available rain gauge system. After applying a correction factor, the precipitation estimates were used as input to the rainfall-runoff model. The results showed interesting possibilities, although scaling problems between the precipitation forecasted by the model and that measured at the rain gauges, still exist. Further investigation should therefore be devoted to this approach, which seems extremely promising, by developing an improved post-processor for the precipitation estimates.

Finally, the results of the three years research were presented on June 1994 at an International Workshop organised in Bologna under the auspices of the World Meteorological Organisation and of the Commission of the European Communities.

CONCLUSIONS AND FUTURE LINES OF RESEARCH

AFORISM was the first phase, the feasibility study, of a complex comprehensive tool to be developed in order to support both planning and forecasting. From the experiments that were carried out during the project a number of elements have emerged which should guide in the operational implementation of the system as well as in developing future lines of research.

The first point relates to the definition of the precipitation measurement system, which, although it does not really constitute the bottleneck of the overall system if it is conveniently designed, nevertheless it should benefit of the redundancy provided by different data acquisition systems.

The second point relates to the rainfall runoff models to be used for real time flood forecasting which often raises lengthy discussions among Hydrologists. It was agreed that it is mostly convenient to use continuous type semi-distributed models for real-time flood forecasting, because of their flexibility, their extended possibility of calibration, their capability of reproducing complex systems combined with very small computer time requirements.

The third point relates to the need for reliable quantitative precipitation forecasts over areas of 200 - 300 km2 in size extended in time up to 12 hours and sampled at time intervals of 1 hour, an essential point that may greatly improve the efficiency and the performances of the overall system. Additional work has to be performed in this domain in order to improve the precipitation forecasts using the Limited Area Model on the basis all the available sources of precipitation measurements such as the conventional raingauge networks, the weather radars, and, possibly with the satellite measurements and images.

The fourth point reflects the need for a complex integrated Decision Support system which should include a Data Base, a Geographic Information System an Expert System shell and a number of models, such as for instance the 1D-2D hydraulic models that would allow the Decision Maker to evaluate the possible outcomes and the effects (positive or negative) of his planned interventions.

The fifth point, which emerged in the discussions with the Authorities, is the need for the establishment of unified and co-ordinated Manual of Procedures without which the use of flood forecasting systems as well as of the Decision Support Systems would be nullified.

Taking into consideration all these points, there are two concluding remarks that should be emphasised. The flood forecasting and management problem is a complex problem which involves several disciplines as well as several different Authorities and the problems are far more complex than the "simple" precipitation or flood forecasts. The final comment, which is also a hope, is that most of the research in this field should be addressed in providing the data and the forecasts that "are needed" and not the ones that "can be provided", which unfortunately in many cases seems the present situation, merely reflecting marketing strategies more than the will of actually solving problems.

REFERENCES

Abbott M.B., Bathurst J.C., Cunge J.A., O'Connell P.E. and Rasmussen J., 1986a - An introduction to the European Hydrological System - Système Hydrologique Européen, "SHE", 1: History and philosophy of a physically-based, distributed modelling system, Journ. of Hydrol., 87: 45-59.

Abbott M.B., Bathurst J.C., Cunge J.A., O'Connell P.E. and Rasmussen J., 1986b - An introduction to the European Hydrological System - Système Hydrologique Européen, "SHE", 2: Structure of physically.based, distributed modelling system, Journ. of Hydrol., 87: 61-77.

Beven K.J.and Kirkby M.J., 1979 - A Physically-based variable contributing area model of basin hydrology, Hydrological Sciences Bull., 24(1), pp. 43-69.

Browning K.A., 1979 - The FRONTIERS plan: a strategy for using radar and satellite imagery for short range precipitation forecasting. Met. Mag., 108,161-184.

Browning K.A. and Collier C.G., 1982 - An integrated radar - satellite nowcasting system in the UK, in Browning K.A. ed. "Nowcasting", Academic Press, London, pp 47 - 61

Collier C.G., 1990 - COST 73: The development of a weather radar network in Western Europe, in Collier C.G. and Chapuis M. eds. Weather Radar Networking Seminar on COST Project 73, Kluwer Academic Publishers, Netherlands, EUR 12414 EN - FR.

Collier C.G., 1994 - Objective Methods of forecasting precipitation using weather radar data, in Almeda-Teixeira M.E., Fantechi R., Moore R., Silva V.M. eds. "Advances in radar hydrology". European Commission Report EUR 14334 EN.

OST Project 72, 1985 European Commission Report EUR 10353 EN - FR.

Di Giammarco P., Todini E., Consuegra D., Joerin F., Vitalini F., 1994 -Combining a 2-D flood plain model with GIS for flood delineation and damage assessment, in P. Molinaro and L. Natale eds. "Modelling of flood propagation over initially dry areas" Proc. of the Specialty Conf. ENEL-DSR-CRIS, Milan, 171-185.

Di Giammarco P., Todini E., Lamberti P., 1995 - A conservative finite elements approach to overland flow: the control volume finite element formulation. - accepted Journ. of Hydrol., (in press).

ET&P, 1992, The FUCHUN River Project - A Computer Based Real Time System - Final Report CEC Research Contract n. CI13-0004-I.

Foufoula-Georgiou E. and Georgakakos K.P., 1991 - Hydrologic Advances in Space-Time Precipitation Modeling and Forecasting. In D.S. Bowles and P.E.O'Connell (eds.), Recent Advances in the Modeling of Hydrologic Systems, 47-65.

Grijsen J.G., Snoeker X.C., Vermeulen C.J.M., Mohamed El Amin Moh. Nur, Yasir Abbas Mohamed, 1992 - An Information System for Flood Early Warning, in Saul A.J. ed. "Floods and Flood Management", Kluwer Academic Publishers, pp 263-289.

Lazic L.,Telenta B, 1990 - Documentation of the UBC/NMC eta model. WMO/TMRP Techn. Report, 40.

Mellor D., 1995 - The Modified Turning Bands (MTB) Model for space-time rainfall. 1: Methodology. Accepted Journ. of Hydrol., in press.

Mellor D. and O'Connell P.E., 1995 - The Modified Turning Bands (MTB) Model for space-time rainfall. 2: Estimation of raincell parameters. Accepted Journ. of Hydrol., in press.

Mellor D.and Metcalfe A.V., 1995 - The Modified Turning Bands (MTB) Model for space-time rainfall. 3: Estimation of the storm/rainband profile and a discussion of future model prospects. Accepted Journ. of Hydrol., in press.

Milford J.R. and Dugdale G., 1989 - Estimation of rainfall using geostationary satellite data. Applications of Remote Sensing in Agricultural Sciences, University of Nottingham, Butterwoth, London.

Natale L.,Todini E., 1976a - A stable estimation for linear models - 1. Theoretical development and Monte-Carlo experiments - Water Resources Research Vol. 12 n. 4, pp. 667-671.

Natale L.,Todini E., 1976b - A stable estimator for linear models - 2. Real World Hydrologic applications- Water Resources Research - Vol. 12 n. 4, pp. 672-675.

NEXRAD, 1984 - Next Generation Weather Radar, Programmatic Environmental Impact Statement R400-PE201, NEXRAD Joint System Program Office, US Dept. of Commerce.

Rosas Dias Manuel P., 1994 - Radar Measurements of Precipitation for Hydrological Purposes, in Almeda-Teixeira M.E., Fantechi R., Moore R., Silva V.M. eds. "Advances in radar hydrology". European Commission Report EUR 14334 EN.

Sherman L.K., 1932 - Streamflow from rainfall by the unit graph method, Engineering News Record, N. 108, pp. 501-505.

Todini E., Wallis J.R., 1977 - Using CLS for daily or longer period rainfall - runoff modelling, in Ciriani T.A., Maione U., Wallis J.R. eds. "Mathematical models for surface water hydrology", John Wiley & Sons, London.

Todini E. and Bossi A., 1986 - PAB (Parabolic and Backwater) an unconditionally stable flood routing scheme particularly suited for real time forecasting and control, Journal of Hydraulic Research, Vol. 24, n. 5, pp. 405-424.

Todini E., 1993 - Un filtro di Kalman per la ricostruzione automatica dei dati mancanti Presented at the STACHEM'93 Conference in Venice (in Italian).

Todini E., 1995a - The ARNO rainfall-runoff model, accepted Jour. of Hydrol., in press.

Todini E., 1995b - New Trends in Modelling Soil Processes from Hillslope to GCM Scales, in Oliver H.R. and Oliver S.A. eds. "The Role of Water and the Hydrological Cycle in Global Change", NATO ASI Series I/31, Springer-Verlag.

Todini E., 1995c - AFORISM - A Comprehensive Forecasting System for Flood Risk Mitigation and Control, Final Report of Contract EPOC-CT90-0023, submitted to the Comm. of European Comm., in press.

Todini E., Di Bacco M., 1995 - A Bayesian approach to rainfall modelling, Proc. Statistical and Bayesian methods in Hydrological Sciences, An International Conference in honour of Jaques Bernier -UNESCO, in press.

Wang J.M., Zhang R.F., Xu G.W., 1987, The synthesized constrained linear system (SCLS). Journ. Hydraulic Eng.of China, n. 7 (in Chinese).

WMO, 1981 Guide to Hydrological Practices, WMO publications, no. 168, 4th edit., Geneva.