Infoscience

Thesis

Improving Alpine Flood Prediction through Hydrological Process Characterization and Uncertainty Analysis

Among the many challenges of Alpine flood prediction is describing complex, meteo-hydrological processes in a simplified, robust manner that can be easily integrated into operational forecasting. In this dissertation, improved methods to characterize these processes are developed and integrated into the hydrological modeling component of an operational flood forecasting system used in the Swiss Alps. Detailed studies are conducted to improve hydrological model inputs, processes and outputs. Improvements, detailed in four chapters of this thesis, address the overarching goal of this work – the reduction of flood forecasting uncertainty. The accuracy of flood predictions in Alpine areas is contingent upon adequate interpolation of meteorological forcings, which has significant impacts on discharge volumes and flood peaks. This thesis demonstrates an improvement in the interpolation of temperature and precipitation inputs using a robust variogram which considers anisotropy and using a geostatistical interpolation method to distribute inputs in space and time. Results show that using elevation as the external drift factor better describes orographically-induced precipitation and temperature gradients. Also, the consideration of anisotropy is integral in detailing spatial patterns of precipitation induced by storm advection. Hydrological flood forecasting in mountainous areas also requires accurate partitioning between rain and snowfall to properly estimate the extent of runoff contributing areas. Partitioning is improved in this work by using a new method to integrate Limited Area Model output. Unlike standard hydrological procedures in inferring snowfall limit estimates based on dry, ground temperature measurements, Limited Area Model output considers the vertical, humid, atmospheric structure in its snowfall limit calculations. In effect, this method provides good estimates of runoff contributing areas in the spring as evidenced by validation on discharge measurements and satellite images of snow coverage. Accurately describing snowmelt processes on a sub-daily scale is also of critical importance in Alpine flood forecasting. However, the complex topography of the study region has limited observations available for validation. This thesis presents the development of a new physically-based snowmelt method applicable to regions with limited data. This method uses only daily minimum and maximum temperatures to mimic the effects of radiation. A comparative analysis of snowmelt methods is validated with snow lysimeter data and with a unique, distributed meteorological dataset collected by a wireless weather station network. Results demonstrate that the new method is competitive with more complex snowmelt methods as shown by accurately reproducing diurnal snowmelt cycles. Conveying limits of certainty on flood prediction outputs to users is critical because of epistemic and aleatory errors inherent to environmental modeling. Due to the presence of these errors, the GLUE methodology and multi-criteria performance ideas have been adapted with a fit-for-purpose uncertainty estimation technique in the final part of this thesis. With this method, hydrological model parameters are constrained based on hydrograph behavior, with a particular focus on flood peak response. A key component of the technique is a visualization tool which shows acceptable ensembles of discharge with respect to individual and combined criteria. By integrating the aforementioned input and process improvements into the hydrological model, calibration achieves model outputs that capture observed river discharge. Also, the uncertainty associated with hydrological modeling output error is reduced. Findings of this thesis are applicable to operational flood forecasting in general and have proven utility in improving hydrological model predictions in mountainous regions. Due to the novelty of the developments in terms of new methods or the use of tools and data sources previously unexploited in flood forecasting, further testing of the improvements is recommended. Future research in quantifying the chain of uncertainty produced by combining probabilistic forecast inputs with the hydrological output ensembles is also critical when the improved flood forecasting model becomes fully operational.

Related material