Establishing the skill of climate field reconstruction techniques for precipitation with pseudoproxy experiments in Europe

Abstract:

In recent years important efforts were focused in the development of Climate Field Reconstructions (CFR). These techniques allow merging a limited number of local reconstructions to produce a regular grid which accommodates climatic information from different proxy sources. Pseudo Proxy Experiments (PPE) are idealized experiments performed with synthetic data, typically generated with comprehensive climate models. The PPEs aim at assessing the accuracy that can be expected by (statistical) state-of-the-art reconstruction techniques, but also at identifying potential shortcomings and improvements.

In this study a number of PPEs are investigated in order to assess the ability of three different CFR techniques to reconstruct precipitation over Europe. The methods comprise of a linear fit (Canonical Correlation Analysis, CCA), a simple non-linear approach (the Analog Method, AM) and a Bayesian model (Bayesian Hierarchical Method, BHM). Given the inherent complexity of this variable, hardly reproduced by state-of-the-art global circulation models, some downscaling technique is necessary to design meaningful PPEs. In this study the synthetic data consist of a high-resolution climate simulation performed with a Regional Climate Model over Europe for the last two Millennia.

Results indicate that unlike BHM, CCA systematically underestimates the variance. TheAM can be adjusted to overcome this shortcoming, presenting an intermediate behavior between the two aforementioned techniques. However, a trade-off between reconstruction target correlations and reconstructed variance is the drawback common to all CFR techniques. CCA (BHM) represent the largest (lowest) skill in preserving the temporal evolution, whereas the AM can be tuned to reproduce better correlation at the expense of a loss in variance. While BHM, in the form employed here, has been shown to perform well for temperatures, it does heavily rely on prescribed spatial correlation lengths. For temperature this assumption might be valid, it is hardly warranted for precipitation data. In general, none of the methods outperforms all others. All experiments suggest that a dense and regularly distributed proxy network is required to reconstruct precipitation accurately, reflecting its high spatial and temporal co-variance structure.