Present and LGM permafrost from climate simulations: contribution of statistical downscaling

We quantify the agreement between permafrost distributions from PMIP2 (Paleoclimate Modeling Intercomparison Project) climate models and permafrost data. We evaluate the ability of several climate models to represent permafrost and assess the inter-variability between them. Studying an heterogeneous variable such as permafrost implies to conduct analysis at a smaller spatial scale compared with climate models resolution. Our approach consists in applying statistical downscaling methods (SDMs) on large- or regional-scale atmospheric variables provided by climate models, leading to local permafrost modelling. Among the SDMs, we first choose a transfer function approach based on Generalized Additive Models (GAMs) to produce high-resolution climatology of surface air temperature (SAT). Then, we define permafrost distribution over Eurasia by SAT conditions. In a first validation step on present climate (CTRL period), GAM shows some limitations with non-systemic improvements in comparison with the large-scale fields. So, we develop an alternative method of statistical downscaling based on a stochastic generator approach through a Multinomial Logistic Regression (MLR), which directly models the probabilities of local permafrost indices. The obtained permafrost distributions appear in a better agreement with data. In both cases, the provided local information reduces the inter-variability between climate models. Nevertheless, this also proves that a simple relationship between permafrost and the SAT only is not always sufficient to represent local permafrost. Finally, we apply each method on a very different climate, the Last Glacial Maximum (LGM) time period, in order to quantify the ability of climate models to represent LGM permafrost. Our SDMs do not significantly improve permafrost distribution and do not reduce the inter-variability between climate models, at this period. We show that LGM permafrost distribution from climate models strongly depends on large-scale SAT. The differences with LGM data, larger than in the CTRL period, reduce the contribution of downscaling and depend on several factors deserving further studies