The development of climate prediction systems for years to decades is an area of current research, as these time scales are important e.g. for the planning horizon of decision-makers. Those prediction systems can be improved by including the knowledge of past climate states. To get a better understanding of climate variations, paleoclimate models simulate climate for certain periods in the past, often key periods such as the Last Glacial Maximum (21,000 years before present), the Mid-Holocene (6,000 years before present), etc. For both decadal and paleoclimate applications, ensembles of climate predictions (a set of predictions instead of the most likely one) are evaluated to quantify the uncertainty of the predictions. The verification of such ensemble climate predictions is an ongoing field in climate research. In this thesis, the quality of decadal and paleoclimate ensemble predictions is assessed by a probabilistic evaluation that comprises different attributes such as reliability/calibration and skill.
Creating decadal climate predictions is challenging and still in an experimental stage due to little experiences (e.g. with the initialization of the model components) compared to weather forecasting. We consider three experiments (b1-LR, pr-GECCO, pr-ORA) of the MiKlip (Mittelfristige Klimaprognosen) decadal prediction system. These experiments differ in the way the atmospheric and oceanic model components are initialized, the number of ensemble members, etc. Each ensemble experiment is validated using one observational dataset, i.e. we assume no observational uncertainties. The threedimensional evaluation in the atmosphere and in the ocean shows skillful and reliable areas, especially in the subtropics and mid-latitudes. However, in all experiments, we detect deficiencies in the tropical Pacific region at higher altitudes, which may result from falsely generated dynamics in the model physics. For the ocean, we see clear differences between the experiments mostly caused by differences in the initialization data. In the Pacific and the subtropical belt around the equator, pr-GECCO outperforms b1-LR and pr-ORA also in deeper layers of the ocean whereas in the North Atlantic, b1-LR and pr-ORA are more reliable compared to pr-GECCO.
Pollen and macrofossils in sediment cores provide the basis for the local reconstruction of vegetation and, thus, climate for a state in the past. We determine probabilistic information of the observed pollen by estimating botanical climate transfer functions using the generalized linear model. This probabilistic information is used to optimize a multi-model ensemble created from members of PMIP3 (Paleoclimate Modelling Intercomparison Project Phase 3). For the Mid-Holocene, summer temperatures change clearly (up to 0.4 K over land) when assimilating the PMIP3 multi-model ensemble to the observed pollen data. The added value is evidenced by the predominantly positive Brier skill scores (improvement of ca. 20% on average).
Another approach to estimate climate transfer functions is the quadratic discriminant analysis as used in the Bayesian biome model. To apply the Bayesian biome model, the environmental vegetation needs to fulfill similar conditions as in the Dead Sea basin, where three vegetation zones (Mediterranean, Irano-Turanian, and Saharo-Arabian territory) are considered at the transition from arid to sub-humid climate. We apply the Bayesian biome model to a sediment core drilled at the Dead Sea, which encompasses the last ca. 220,000 years. For the Eemian warming phase (approx. 130,000 to 115,000 years before present), we find similar winter temperatures and annual precipitation as for today. For the Last Glacial, the reconstructed values show generally higher precipitation rates and lower winter temperatures compared to today's climate.
Universitäts- und Landesbibliothek Bonn Accessed 107 times | Last updated 28.09.2018
Stolzenberger, S. (2017): On the Probabilistic Evaluation of Decadal and Paleoclimate Model Predictions. University of Bonn
|Title||On the Probabilistic Evaluation of Decadal and Paleoclimate Model Predictions|
|School||University of Bonn|