Parameter regionalization for hydrological models using CFG

PI: Univ.Prof. Dr. Karsten Schulz Co-PI:Dr. Mathew Herrnegger Hydrological models are important tools to help determining effective water resources management practices. For example, they can be applied to estimate runoff, optimize flood protection measures, or to decide the most effective strategies for electric power production. Hydrological models are also used in climate change impact assessments, for example, when future drinking water or irrigation water amounts have to be estimated. Hydrological and especially runoff generation processes are very complex and are dependent on a number of factors. Next to the precipitation amounts and intensities, the generation of runoff is largely controlled by the physio-geographic conditions of the landscape/catchment. Most important are the topography of the area, the land use and land cover (including impermeable surfaces and vegetation), the soil (depth and texture) and the prevailing soil moisture conditions. Hydrological models need to incorporate the most dominant hydrological processes as well as the physio-geographical characteristics of the catchment into a mathematical formulation; the catchment characteristics then enter the model in the form of so-called process or model parameters. One of the major current challenges in hydrology is that these catchment characteristics cannot be observed or determined catchment-wide in a way that is required as input by the models. In this project, we will develop a methodology that will be able to estimate so-called transfer-functions directly from runoff data measured at several gauges within the catchment. These transfer-functions mathematically relate available landscape characteristics (e.g. digital elevation modelsopography from laser scans; land use land cover from satellite images) to the necessary model parameters at each pixel/location in the catchment. The new method will not only estimate the mathematical structure of these transfer functions, it will also derive its coefficients. Thereby using novel methods from Applied Informatics and in particular from the area of Context Free Grammar. The new method to derive the necessary transfer functions will be developed and implemented for three hydrological models differing in complexity (among which the spatially distributed conceptual rainfall-runoff model COSERO, used by the VERBUND trading GmbH, Austria for its operations) and tested for a number of different catchments in Austria and in Europe. In this way, we anticipate developing a methodology, which will allow for highly improved runoff prediction in catchments by incorporating an improved estimation process for model parameters via transfer-functions that are based on the available physical characteristics of a catchment. .

The prediction of flood events and the assessment of climate change on the water balance and on natural hazards are carried out with the help of mathematical models. These models describe mathematically how runoff develops from rain and snow, or how climate change causes glaciers and snow to melt, or impacts the formation of groundwater/drinking water, for example. The models need a variety of information for each point in the landscape, such as the water conductivity and porosity of the soils or geologic bedrock, the root distribution and other characteristics of the vegetation, or the topography of the area. These quantities are not measurable everywhere, but are needed in the models and therefore have to be estimated spatially, as so-called model parameters, representing local conditions. In this project, we have developed a new methodology that is able to estimate these model parameters for any point in the catchment from data of Geographic Information System (GIS) and satellite imagery. The method is able to determine so-called transfer functions from runoff and spatially distributed catchment information, which allows to calculate the necessary model parameters for each point in the catchment. The method uses a so-called autoencoder (a neural network) and can both derive the mathematical formula for the transfer functions and simultaneously determine the necessary coefficients in the formulas. We could show that this method is very well transferable to other areas and that with additional data both the estimation of the formulas of the transfer functions and hydrological predictions can be performed with significantly lower uncertainties.