H12C-01 INVITED 10:20h
The use of (Uncertain) Landscape Discretization Schemes in Process-Oriented Catchment Models of Different Complexity
Process-oriented catchment models require spatially distributed data sets for defining the model structure and parameterizing different modules. However, as the classical input time series of hydrological variables, also theses spatial data sets are uncertain and the using them in models involve many assumptions by the modeller. For instance, the use of point data sets require an adequate regionalization (problems of station distribution and representativeness of locations etc.), and the use of spatial patterns is often difficult due to the limited availability of suitable data and due to scale issues. Breaking up the catchment in different functional units (`hydrotopes') is one way to incorporated hydrological process understanding into a catchment model. In this paper, two different models of different complexity are compared. On the hand, the process-oriented catchment model TAC-D (tracer aided catchment model, distributed) was used, which is a fully distributed raster models with different modules to simulate all hydrological processes continuously on an hourly base. On the other hand, an event-based, semi-distributed model was applied that is based on the Geomorphological Instantaneous Unit Hydrolograph (GIUH) approach. The studies were performed for catchments in the Black Forest Mountains, Germany, and in the Kitzbueheler Alps, Austrian. For both models spatial delineations of functional units of different complexity were used to explore the uncertainty introduced by assuming just one delineation. In addition, the model prediction uncertainty introduced by the spatial delineation is compared to the uncertainty using distinct precipitation input data sets (ground station data vs. radar data). Beside the simulated runoff also the distribution of runoff components and the predictions of hydro-chemical parameters are investigated.
H12C-02 10:40h
Identifying a Switch in the Catchment State through a Hierarchical Mixture of Experts model
Despite the abundance of models to describe the rainfall-runoff process, there is not a single model that will perform reliably over the range of possible catchment types and conditions. Evidence exists of the catchment responding differently under different antecedent conditions, so using a model with a rigid structure can lead to significant bias in the modelled hydrograph. Consider for many conceptual models, could it be said that the saturation excess overland flow mechanism is valid for all seasons, for dry and wet antecedent conditions? Could it be possible that the true model is much more complex and fluctuates between two or more model states? An alternative approach to selecting a single model is to combine the results from several hydrological models. Methods based on Bayesian statistical techniques provide a means to compare and combine competing models whilst allowing for model uncertainty. A framework that allows different modelling configurations to apply in multiple hydrologic states is presented. Each model configuration is adopted at a given time with a probability that depends on the current hydrologic "state" of the catchment. This framework is known as a Hierarchical Mixture of Experts (HME). The HME model framework is applied to a range of Australian catchments of varying attributes using an established conceptual model. Results from this application are compared to the alternative where the hydrologic state is assumed to be stationary. The comparison is performed under a Bayesian framework, enabling assessment of the accuracy of each alternative irrespective of the complexity the model contains. Due to the probabilistic basis under which the model exists, Bayesian techniques form an ideal basis to identify the distribution of the parameters.
H12C-03 10:55h
Utility of different data types for calibrating flood inundation models within a GLUE framework
In this paper we explore the value of different types of data in constraining the predictions of a simple two-dimensional hydraulic model, LISFLOOD-FP, applied to the January 1995 flooding on the River Meuse, The Netherlands. For a 35 km reach of the Meuse between Borgharen and Maaseik a data set has been assembled consisting of Synthetic Aperture Radar and air photo images of inundation extent, downstream stage and discharge hydrographs, two stage hydrographs internal to the model domain and 84 point observations of maximum free surface elevation. The data set thus contains examples of all the types of data that can potentially be used to calibrate flood inundation models. 500 realisations of the model have been conducted with different friction parameterisations and the performance of each realisation has been evaluated against each observed data set. Implementation of the Generalised Likelihood Uncertainty Estimation (GLUE) methodology is then used to determine the value of each data set in constraining the model predictions and to determine the reduction in parameter uncertainty resulting from the updating of generalised likelihoods based on multiple data sources.
H12C-04 11:10h
Uncertainty in Flood Inundation Predictions
Accuracy in the prediction of flood inundation for both risk mapping and real time forecasting requires accuracy in the representation of flow dynamics, input upstream and lateral fluxes, channel and flood plain topography, flood plain infrastructure, and all the different forms of energy losses in the flow at flood stages. Most predictive models of flood inundation are limited in their accuracy in all of these requirements so that some uncertainty in respect of the inundation predictions is inevitable. In this presentation an ensemble prediction methodology is presented that allows the evaluation of prediction quantiles for patterns of flood inundation. Model evaluation and weighting is evaluated within an extended GLUE framework.
H12C-05 11:25h
AN EVALUATION OF MODEL STRUCTURE UNCERTAINTY EFFECTS FOR HYDROLOGICAL SIMULATION
The complexities of sustainable water management have led to increasing use of integrated hydrological models for hydrological prediction and forecasting. While the inherent uncertainty in hydrological simulation is widely recognised, finding methods that address all the sources of uncertainty remains a complex and challenging problem. One of the most challenging aspects of this problem is how to consider model structure uncertainties. Relatively few studies have directly addressed the effect of model structure on model performance and uncertainty predictions. In many cases the model structure error is used to account for residual errors once the other sources of uncertainty have been quantified. To address the issue of model structure uncertainty a general integrated modelling framework is described for considering the effect of different model structures on model predictions. A methodology is proposed for evaluating alternative model structures. This methodology is then applied to a US NWS study catchment, the Blue river basin as part of the NWS Distributed Model Intercomparison Project. The relative performance of different acceptable model structures is evaluated as a representation of structural uncertainty and compared to the uncertainty estimates arising from measurement uncertainty, parametric uncertainty and the rainfall input. The results show that model performance is strongly dependent on model structure and the uncertainty associated with model structure is similar in magnitude to the other sources. The same methodology was applied to evaluate multimodel ensembles. It was found that the ensemble average of 10 acceptable models performs better than any single model in a split sample test. Regression methods were then used to identify which model structures provide significant contributions to accurate hydrological simulation.
http://www.nws.noaa.gov/oh/hrl/dmip/
H12C-06 11:40h
Sensitivity Analysis to Identify `Soft Data' for the Evaluation of a River Water Quality Model
A sensitivity analysis is performed to identify the parameters of a river water quality model that have the most influence on the model outputs. Results of a sensitivity analysis provide guidelines about how parameter uncertainty will affect the model output, but always relate to the specific circumstances under which the model was build and calibrated. If the model has to be applied on a river with different characteristics, again an extended dataset is needed to identify the important parameters of the model and the associated uncertainty levels. If uncertainty and characteristics of the river basin can be linked in advance, this could open perspectives for model applications in ungauged basins. The aim of this research is to examine this link by testing the sensitivity of a river water quality model to the a priori assumption of parameter values. In non-linear models, the propagation of uncertainty in a particular parameter depends on several factors, such as the values of the other model parameters and the specific conditions. The values of parameters refer in most cases to specific circumstances. For example, a river with high algae blooms during summer periods will have its parameters of the algae growth model adapted to the growing species when calibrated. The presented analysis can reveal important information about the uncertainty propagation for situations were no or poor measurements are available. Indeed, if general clusters can be found of cases in which some parameters are more sensitive than others, then this information can be used as 'soft data' to identify when certain parameters become more important than others. Once the important parameters are known, optimal experimental design techniques can be used to determine the optimal measurement strategy that allows a better identification of these parameters before calibrating the model. Here, a water quality model of the river Dender implemented in the ESWAT simulator, is used as an application of the described methodology on a real case study. Latin Hypercube Sampling around nominal values is applied, with ranking of the parameter uncertainty after a multi-linear regression. This sampling is repeated with different nominal values for the parameters within realistic ranges. By grouping the model simulations on the basis of water quality variables exceeding certain levels during some periods of the year, clusters are formed. Finally, the parameter values of those clusters are evaluated in relation to external circumstances.
H12C-07 11:55h
Assessing model and data uncertainty by combing a tracer experiment with process modeling
While field studies of hillslopes often result in complex hydrological descriptions reflecting the spatial and temporal variability of water source and flow path, most modeling efforts find it challenging to incorporate this information directly into model structures. This is frequently due to the disparity between the scale of field observations and model sub-units and natural hillslope heterogeneity. Thus, parameters are usually estimated through a calibration-validation exercise with measured runoff even if the runoff information content limits the identifiability of model parameters. We approach this problem on a well-studied hillslope (WS10, H.J. Andrews, Oregon, USA) by combining the merits of a simple physically-based model with applied line source tracers (Amino G acid and bromide) and detailed soil property measurements (i.e., based on a database containing values of hydraulic conductivity, porosity, and water retention characteristics determined from $>$400 soil cores collected at the WS10 hillslope). We parameterized a relatively simple model structure that represents the dominant runoff generation processes in forested hillslopes by performing a multi-criteria calibration (runoff, tracer, and groundwater levels) over the range of measured soil properties. We were able to demonstrate that acceptable simulations with an improved parameter uncertainty can be found within the parameter range determined by the measured field data. We attributed the success our study to the limited model complexity used to structure our model that was largely field data driven, but constrained by both tracer and runoff.