Investigating uncertainty in global hydrology modelling
Smith, Katie A.(2016)
As projections of future climate raise concerns over water availability and extreme hydrological events, global hydrology models are increasingly being employed to better understand our global water resources and how they may be affected by climate change. Being a relatively recent development in hydrological science, global hydrology modelling has not yet undergone the same level of assessment and evaluation as catchment scale hydrology modelling. Until now, global hydrology models have presented just one deterministic model output for use in scientific research. Recently, multi-model ensembles have compared these outputs for different global models, but this has been done prematurely as the uncertainties within individual models have yet to be understood.\ud \ud This study demonstrates a rigorous uncertainty investigation of the 123 parameters within the Mac-PDM global hydrology model over 21 global river catchments. Mac-PDM was selected for its relative simplicity amongst global hydrology models, and its suitability for application using high performance computer clusters. A new version of the model, Mac-PDM.14 is provided, with updated soil and vegetation classifications. This model is then subjected to a 100,000 parameter realisation Generalised Likelihood Uncertainty Estimation (GLUE) experiment, requiring 40 days of high performance computing time, and outputting over 2Tb of data. The top performing model parameterisation from this experiment provides an annual average error of 47% when compared to observed records, a 45% improvement over the previous version of the model, Mac-PDM.09. Given the computational expense of such an experiment, smaller sample sizes of parameter realisations are explored. Whilst the top performing parameterisation in a sample size as small as 1,000 can perform almost as well as that from 100,000 parameterisations, the number of good parameterisations is fewer, and the range of model uncertainty may therefore be significantly underestimated.\ud \ud Mac-PDM.14 is shown to have a lower mean absolute relative error than all models involved in both the Water and Global Change (WATCH) project and the Inter-Sectoral Impacts Model Intercomparison Project (ISI-MIP). Parameter uncertainty is compared to model uncertainty, and the uncertainty range between the models within the WATCH and ISI-MIP projects is comparable to the parameter uncertainty within Mac-PDM.14. Catchment specific calibrations of the global hydrology model are explored, and it is demonstrated that the model performance is improved by 22 to 92%, for the Niger and the Yangtze respectively, with catchment specific parameter values over a global calibration. Approximate Bayesian Rejection is applied to explore the catchment specific parameter values that result in good parameter performance. Few trends can be identified from this analysis, which suggests that Mac-PDM may be over-parameterised. Catchment specific calibrations in both high latitude and arid to semi-arid regions show significant improvement over global calibration, which indicate a deficiency in model structure; the addition of a glacier component to Mac-PDM is recommended. Model calibrations are validated using the ISI-MIP forcing dataset, and the best model performance gives an error of 44%. This is a betterment on the performance with the WATCH forcing data used in calibration, and so implies that models not need to be recalibrated every time new forcing datasets are employed.\ud \ud This research highlights that the performance of global hydrology models can be significantly improved by running a parameter uncertainty assessment, and that in catchment scale studies, catchment specific calibration should be carefully considered. Furthermore, the uncertainty within individual global hydrology models is an important consideration that should not be overlooked as these models are increasingly included in ensembles and interdisciplinary studies.