Robustness of Regional Patterns of Change in Multi-model Studies; Beyond Model Spread

Abstract:

Grid point percentiles have commonly been used to represent uncertainty in temperature and precipitation projections, while pattern scaling of temperature and precipitation has been used to demonstrate robustness in climate change projections.

Using CMIP5 we demonstrate a substantial disproportionality between global climate sensitivity and the regional distribution of temperature change. We argue that estimating a given probability for a change in any particular region prevents a probabilistic statement about the change in a different region, as this would potentially be in conflict with the global mean being the overall constraint in any given model realization: if the largest climate change signal is chosen in each point, the global mean temperature change exceeds the change projected by any individual model. This is not physically justified. For precipitation, spread is large in all grid points; consistent with no change and a large uncertainty everywhere. Each individual model on the other hand shows a clear and statistically significant pattern of change. This implies that the physical consistency between climate variables such as temperature and precipitation is basically lost in a statistical analysis based on grid-point statistics.

Here we propose an EOF approach identifying dominating patterns of regional climate change. These are used to construct globally consistent maps of the uncertainty in climate change scenarios. By going beyond the grid point level statistics, our method is designed to capture the spatial patterns in the uncertainty and maintain the physical correlations between variables. In particular, we identify that based on the EOF analysis of temperature changes; a bit more than 50% of the variance in the temperature change pattern is explained by the first EOF. To keep physical consistency, using the same PC loadings deduced from temperature on precipitation changes, only explains around 10% globally and between 15 - 30% of the variance in precipitation changes over land, depending on various means to separate out the most dominant individual model. This implies, however, that within the inner 50% of the explained temperature variance, the more physically consistent pattern of change in precipitation is much better constrained than it appears using grid point statistics.