Regionalisation of flow-duration curves on global scale – assessment of methods

Water is essential for human well-being and healthy ecosystems, but population growth and changes in climate and land-use are putting increased stress on water resources in many regions. To ensure water security, knowledge about the spatiotemporal distribution of these resources is of great importance. However, estimates of global water resources are constrained by limitations in availability and quality of data. This thesis explores the quality of both observational and modelled data, gives an overview of models used for large-scale hydrological modelling, and explores the possibilities to deal with the scarcity of data by prediction of flow-duration curves.The evaluation of the quality of observational data for large-scale hydrological modelling was based on both hydrographic data, and model forcing and evaluation data for basins worldwide. The results showed that a GIS polygon dataset outperformed all gridded hydrographic products analysed in terms of representation of basin areas. Through a screening methodology based on the long-term water-balance equation it was shown that as many as 8–43% of the basins analysed displayed inconsistencies between forcing (precipitation and potential evaporation) and evaluation (discharge) data depending on how datasets were combined. These data could prove disinformative in hydrological model inference and analysis.The quality of key hydrological variables from a numerical weather prediction model was assessed by benchmarking against observational datasets and by analysis of the internal land-surface water budgets of several different model setups. Long-term imbalances were found between precipitation and evaporation on the global scale and between precipitation, evaporation and runoff on both cell and basin scales. These imbalances were mainly attributed to the data assimilation system in which soil moisture is used as a nudge factor to improve weather forecasts.Regionalisation, i.e. transfer of information from data-rich areas to data-sparse areas, is a necessity in hydrology because of a lack of observed data in many areas. In this thesis, the possibility to predict flow-duration curves in ungauged basins was explored by testing several different methodologies including machine learning. The results were mixed, with some well predicted curves, but many predicted curves exhibited large biases and several methods resulted in unrealistic curves.