Spatial and temporal dependence in distribution-based evaluation of CMIP6 daily maximum temperatures

Abstract

Climate models are increasingly used to derive localised, specific information to guide adaptation to climate change. Model projections of future scenarios are conferred credibility by evaluating model skill in reproducing large-scale properties of the observed climate system. Model evaluation at fine spatial and temporal scales and for rare extreme events is critical for provision of reliable adaptation-relevant information, but may be challenging given significant internal variability and limited observed data in this setting. Comparing distributions of physical variables from historical simulations of Coupled Model Intercomparison Project models against observed distributions provides a comprehensive, concise and physically-justified skill measure. Calculating divergence between distributions requires aggregation of data spatially or temporally. The spatial and temporal scales at which a divergence measure converges to a consistent value can indicate the scales at which a well-defined climate signal emerges from internal variability. Below this threshold, there may be insufficient data for robust evaluation, particularly for rare extremes. Here, the behaviour of several divergence measures in response to spatial and temporal aggregation is analysed empirically to give a novel evaluation of CMIP6 daily maximum temperature simulations against reanalysis. Some key insights presented here can inform methodological choices made when deriving adaptation-relevant information. Convergence varies according to model, geographic region and divergence measure; selection of the level of precision at which models can provide reliable information therefore requires a context-specific understanding. For this purpose, an interactive tool provided alongside this study demonstrates scale-dependent evaluation across several geographic regions. Commonly applied measures are found to be only weakly sensitive to discrepancies in the tails of distributions.