Towards understanding the differences between deterministic and probabilistic flood hazard estimation methods

Transcription

1 16 Risk in Water Resources Management (Proceedings of Symposium H03 held during IUGG2011 in Melbourne, Australia, July 2011) (IAHS Publ. 347, 2011). Towards understanding the differences between deterministic and probabilistic flood hazard estimation methods MAGDALENA ROGGER 1, ALBERTO VIGLIONE 1, RALF MERZ 1, ROBERT KIRNBAUER 1, HERBERT PIRKL 2 & GÜNTER BLÖSCHL 1 1 Vienna University of Technology, Institute of Hydraulic Engineering and Water Resources Management, Austria rogger@hydro.tuwien.ac.at 2 Technical Office for Geology, Vienna, Austria Abstract This paper presents the first results of a study on understanding discrepancies between flood estimates from local-scale, process-based deterministic methods and regional-scale probabilistic methods. Runoff processes in 10 pilot catchments in Tyrol were modelled with a continuous distributed rainfall runoff model using detailed catchment information to assist in selecting the model parameters. Hydrogeologic information from field trips was found to be extremely useful for setting the spatial patterns of the storage capacities in the model. Parameters not identifiable from the field trips were obtained from a multi-step calibration to runoff. It is expected that the model is able to extrapolate more accurately to extreme events than a model based on calibration alone. Key words flood hazard; design floods; hydrogeology; runoff modelling INTRODUCTION Estimating the flood hazard is usually based on a probabilistic approach of fitting a statistical distribution to a sample of observed flood peaks, or on a deterministic approach of calibrating an event based on a rainfall runoff model to runoff data (Beven, 2005). While these procedures are reasonably accurate for the conditions the sample represents, extrapolation to conditions beyond the sample, such as extreme events and changed conditions, are often not very reliable (Peel & Blöschl, 2011). In many practical applications, such as the determination of design values for flood control measures, this may cause significant problems, since the probabilistic and deterministic methods may yield flood peak estimates that differ widely for the same catchment. Choice of method in a specific case generally depends on data availability and on the specific tasks. In case a long flood record is available, flood frequency analysis (the probabilistic approach) is usually applied and has been established as the standard approach (Klemeš, 1993). It has the advantage that it allows specification of the return period of the flood peaks. However, the available flood records are often too short to reliably extrapolate to large return periods, in particular in small basins (Merz et al., 2008a,b). Deterministic approaches, on the other hand, are often applied in ungauged catchments where no flood data are available. They are usually based on catchment characteristics, which have the benefit that local conditions are accounted for. However, as detailed information such as precipitation data, soil properties, geology, land use, is often unavailable in ungauged catchments, it is difficult to choose a suitable model type and to properly calibrate the model parameters. This paper presents first results of a study that is aimed at identifying the reasons for discrepancies between probabilistic and deterministic methods and ultimately combining the advantages of the two. The study is set in 10 pilot catchments in Tyrol in the Austrian Alps. METHODOLOGY To understand discrepancies between the probabilistic and deterministic methods, the derived flood frequency approach (Sivapalan et al., 2005; Apel et al., 2006) was adopted here, where a rainfall runoff model is driven by a long, synthetic series of rainfall. The spatially distributed continuous rainfall runoff model of Blöschl et al. (2008) was used. The main advantage of this type of model is that a range of information on the catchments, such as orthophotos, land-use Copyright 2011 IAHS Press

2 Understanding the differences between flood hazard estimation methods 17 information, runoff coefficient maps and especially hydrogeologic runoff process maps can be included. In this manner runoff processes on the local scale can be captured in detail. The overall modelling strategy follows the spirit of the multi-pillar approach of Gutknecht et al. (2006), who proposed a combination of different methods for the estimation of extreme flood hazards, including flood frequency statistics and the design storm method (Fig. 1). In this paper, one part of the overall strategy, namely the continuous rainfall runoff modelling, is presented. Particular focus is on hydrogeologic runoff process maps, which were used to represent the storage capacities of the catchments in detail. Fig. 1 Different approaches for extreme flood hazard estimation. RAINFALL RUNOFF MODEL AND PARAMETER SELECTION The rainfall runoff model is a conceptual, continuous water balance model on a pixel basis (Blöschl et al., 2008). It consists of a snow routine, a soil moisture routine and a flow routing routine. The snow routine represents snow accumulation and melt by the degree-day concept. Runoff generation and changes in the soil moisture state of the catchment are described by the soil moisture routine using a nonlinear function that takes into account maximum soil moisture and evapotranspiration. Runoff routing on the hillslope is represented by one upper and two lower soil reservoirs, while runoff routing in the stream network is expressed by cascades of linear reservoirs. Runoff processes on the local scale were simulated in 10 alpine catchments in Tyrol, western Austria (Table 1). The catchments range in size from 4 km 2 to 98 km 2. For the model, a high temporal resolution of 15 min and a spatial resolution of m per pixel were chosen in order to capture the flood peaks well. The parameters of the model were Table 1 Study catchments in Tyrol. Stream Catchment size (km 2 ) Max. altitude (m a.s.l.) Min. altitude (m a.s.l.) Mean annual precipitation (mm/year) Wattenbach Weerbach Stampfangerbach Teischnitzbach Trisanna Navisbach Walchentaler Bach Debantbach Hornbach Längentalbach Mean annual discharge (m 3 /s)

3 18 Magdalena Rogger et al. chosen based on an approach proposed by Reszler et al. (2006) for identifying parameters of spatially distributed rainfall runoff models. The main goal of the parameterization was to describe runoff processes as realistically as possible. Therefore the choice of parameters was based not on an automated procedure, but on a manual process analysis for each catchment. The approach consists of four steps: Step 1: Step 2: Choice of a priori parameters based on field information. Manual adjustment of parameters by comparing simulated and observed discharges on a seasonal scale. Step 3: Manual fine adjustment of parameters by comparing simulated and observed discharges on the event scale. Step 4: Consistency check of calibrated parameters with field information, and possibly another iteration to Step 1. The a priori choice of parameters was based on all available information of the catchments. Specifically, this included orthophotos with 1 m resolution, detailed land-use maps, runoff coefficient maps from field trips and hydrogeologic runoff process maps from field trips. For determining the parameters of the upper soil storages, which represent surface runoff and interflow, different classes of hydrological response units (HRUs) were defined, such as sealed areas, rocks, forest, meadow, alpine vegetation, etc., which were assumed to behave similarly regarding surface runoff generation. Hence, each class consists of the same set of parameters for the upper soil storages. The choice of parameters as well as the choice of the different HRU classes was based on runoff coefficient maps of the catchments. The runoff coefficient maps were prepared by using the rule-based method of Markart et al. (2004) and information from the field trips. The rule-based method is based on four indicators of surface runoff generation: vegetation (e.g. spruce, pine, shrubs, grassland, ), soils (rich in skeleton, fine soils, cohesive, ), land use (pasture with intensive grazing,...), and indicator plants for water logging (dry to wet). Markart et al. (2004) obtained the rules from a very large number of irrigation experiments in the Alps. In the study catchments, most forested areas are associated with low runoff coefficients and significant storage. Large storage coefficients were therefore used for these areas in the model. On the other hand, bare rocks have a high runoff coefficient, so small storage coefficients were selected, resulting in a fast runoff response. The choice of parameters for groundwater storage was based on hydrogeologic runoff process maps of the individual catchments. These maps were compiled through expert judgement of an experienced hydrogeologist based on extensive field trips as well as on all available supporting information. The geological analysis included an assessment and interpretation of orthophotos and digital terrain models regarding river networks, wells, geomorphologic aspects and hydrogeological characteristics, and a detailed evaluation of hydrogeological maps and maps of unconsolidated sediments. During field trips to each catchment, the first drafts of the maps were refined and additional information was collected. For example, spot discharge measurements at selected points of the river network were made. Based on the field trips the maps were finalised. In Fig. 2 the hydrogeologic runoff process map of the Stampfangerbach catchment is shown. The map illustrates which areas contribute to surface runoff during flood events and highlights the storage potentials of the catchments. In each map, runoff processes are divided into different classes: predominant runoff in deep groundwater; predominant runoff in shallow groundwater; predominant interflow; surface runoff on rocks; runoff on glaciated areas; predominant runoff on saturated areas. If needed, additional classes, such as karst formations or runoff in rocks and fissures, were defined. The hydrogeological information is an extremely valuable input to hydrological models. It is rarely available, but it helps to better estimate the storage capacities of the catchments, which improves the ability of the model to simulate the nonlinear catchment response in the transition from small to large events; thus the time during a rainfall event at which the storage capacity of one catchment is exceeded can be described more precisely. The cross-section in Fig. 2 shows what kind of extra information can be gained by using the hydrogeologic runoff process maps for the choice of parameters in the model. The map gives

4 Understanding the differences between flood hazard estimation methods 19 qualitative information about the depth at which the runoff processes occur and helps identify dominant processes in each part of the catchment. Areas with runoff on rocks, for instance, are defined in such a way that mainly surface runoff occurs in the model. Areas with deep runoff processes on the other hand, have a high percolation rate into the subsurface and a low tendency to contribute to surface runoff. In karstic areas, subsurface flow into neighbouring catchments may occur. The model parameters for the groundwater component were discussed with the hydrogeologist to understand the uncertainty ranges. Fig. 2 Hydrogeologic runoff process map of the Stampfangerbach catchment with cross section. Fig. 3 Hornbach catchment comparison of simulated and observed discharge from April to October RESULTS OF RUNOFF SIMULATIONS For all catchments, at least 20 years of discharge data were available and used for calibrating the model along with the field information, as well as for validating the simulation results by split sample tests. Particular attention was paid to representing extreme flood events well, so individual events were examined graphically during the modelling procedure. Figure 3 shows the simulation results of the Hornbach catchment for the year 2005, which is part of the calibration period. Figure 4 shows the August 2005 flood event in the Hornbach catchment in detail, as well as the same event in the Weerbach catchment. The model captures both the seasonal water balance and the dynamics of the flood events well. Most notably, the receding limb of the flood hydrograph is simulated well, which shows that the storage properties

5 20 Magdalena Rogger et al. of the catchments are represented accurately in the model. This is because hydrogeological information was available and used in setting the parameters of the model. Figure 4 also shows that the rising limbs of the observed and simulated hydrographs correspond well, which suggests that the point in time when the storage capacity of the catchments is exceeded is represented correctly. (a) (b) Fig. 4 Simulated and observed hydrographs of the 2005 flood event in the (a) Hornbach and (b) Weerbach catchments. (a) (b) Fig. 5 Flood frequency statistics of the observed versus simulated flood peaks in the (a) Hornbach and (b) Weerbach catchments. In order to make sure that the model represents the flood statistics well, the flood frequency statistics of the simulations are compared to those of the observations in Fig. 5. For most of the range of the flood frequency curve, the simulations and observations are similar. However, the largest events are different. In the Hornbach catchment the simulations are slightly larger than the observations, while in the Weerbach catchment they are smaller. A detailed analysis of these events indicates that the differences are mainly due to uncertainties in the measured rainfall data. CONCLUSIONS AND OUTLOOK The analyses so far have indicated that the hydrogeologic information from field trips was extremely useful for setting the spatial patterns of the storage capacities in the model. This becomes evident when comparing the receding limbs of the simulated hydrographs with the

6 Understanding the differences between flood hazard estimation methods 21 observed hydrographs. Representing the subsurface storage well is very important, as it improves the ability of the model to simulate the nonlinear catchment response in the transition from small to large events. The next steps in this study involve the generation of long-term precipitation series. To this end, the stochastic precipitation model of Sivapalan et al. (2005) will be calibrated to high-quality rainfall records in Tyrol with 15-min temporal resolution. The precipitation model will be used to generate long time series to be used in Monte Carlo simulations, along with the runoff model presented in this paper. It is expected that this approach will be able to better represent the extrapolation of the flood frequencies to extreme events than fitting a probability distribution to flood peak data. This will help shed light on the current discrepancies between probabilistic and deterministic methods of estimating flood hazards. Acknowledgements This study was performed as a part of the Flash Floods Tyrol (HOWATI) project funded by the Hydrographic Service Tyrol, the Austrian Forest Engineering Service in Torrent and Avalanche Control, Section Tyrol and the AdaptAlp project. We would like to thank all project partners for their excellent collaboration. REFERENCES Apel, H., Thieken, A. H., Merz, B. & Blöschl, G. (2006) A probabilistic modelling system for assessing flood risks. Natural Hazards 38, Beven, K. J. (2005) Rainfall runoff modelling. Introduction. In: Encyclopedia of Hydrological Sciences 122, John Wiley, UK. Blöschl, G., Reszler C. & Komma J. (2008) A spatially distributed flash flood forecasting model. Environ. Modelling & Software 23, Gutknecht, D., Blöschl, G., Reszler, Ch. & Heindl, H. (2006) Ein Mehr-Standbeine Ansatz zur Ermittlung von Bemessungshochwässern kleiner Auftretenswahrscheinlichkeit (A multi-pillar approach to the estimation of low probability design floods). Österreichische Wasser- und Abfallwirtschaft 58, Klemeš, V. (1993) Probability of extreme hydrometeorological events a different approach. In: Extreme Hydrological Events Precipitation, Floods and Droughts (ed. by Z. W. Kundzewicz et al.) (Proc. Yokohama Symp., July 1993), IAHS Publ IAHS Press, Wallingford, UK. Markart, G., Kohl, B., Sotier, B., Schauer, T., Bunza, G. & Stern, R. (2004) Provisorische Geländeanleitung zur Abschätzung des Oberflächenabflussbeiwertes auf alpinen Boden-/Vegetationseinheiten bei konvektiven Starkregen, Bundesministerium für Land- und Forstwirtschaft, Umwelt und Wasserwirtschaft BFW-Dokumentation 3/2004, Merz, R. & Blöschl, G. (2008a) Flood frequency hydrology 1. Temporal, spatial, and causal expansion of information. Water Resour. Res. 44(8), W Merz, R. & Blöschl, G. (2008b) Flood frequency hydrology 2. Combining data evidence. Water Resour. Res. 44(8), W Peel, M. C. & Blöschl, G. (2011) Hydrologic modelling in a changing world. Progr. Phys. Geogr. (in press). Reszler, Ch., Komma, J., Blöschl, G. & Gutknecht, D. (2006) Ein Ansatz zur Identifikation flächendetaillierter Abflussmodelle für die Hochwasservorhersage. Hydrologie und Wasserbewirtschaftung 50, Sivapalan, M., Blöschl, G., Merz, R. & Gutknecht, D. (2005) Linking flood frequency to long-term water balance incorporating effects of seasonality. Water Resour. Res. 41, W06012.