| Abstrakt: |
Valid simulation results from global hydrological models (GHMs), such as WaterGAP3, are essential to detecting hotspots or studying patterns in climate change impacts. However, the lack of worldwide monitoring data makes it challenging to adapt GHMs' parameters to enable such valid simulations globally. Therefore, regionalization is necessary to estimate parameters in ungauged basins. This study presents new regionalization methods for WaterGAP3 and aims to provide insights into selecting a suitable regionalization method and evaluating its impact on the simulation. Our results suggest that machine learning-based methods may be too flexible for regionalizing WaterGAP3 due to a significant performance loss between training and testing. In contrast, the most basic regionalization method (using the concept of spatial proximity) outperforms most of the developed regionalization methods and a pre-defined benchmark-to-beat in an ensemble of split-sample tests. The method selection, whether spatial proximity-based or regression-based, has a greater impact on the regionalization than the specific details on how the method is applied. In particular, the descriptor selection plays a subsidiary role when at least a subset of selected descriptors contains relevant information. Additionally, our research has shown that regionalization causes spatially varying uncertainty for ungauged regions. For example, India and Indonesia are particularly affected by higher uncertainty. The impact of regionalization in ungauged areas propagates through the water system, e.g., one water balance component changed by approximately 2400 km3 yr-1 on a global scale, which is in the range of inter-model differences. The magnitude of the impact of regionalization depends on the variability in regionalized values and the region's sensitivity for the analysed component. [ABSTRACT FROM AUTHOR] |
