Large-scale emulation of spatio-temporal variation in temperature under climate change

Abstract

Future temperature variations under greenhouse gas (GHG) emission scenarios are critical to assess possible impacts on human society and make reasonable mitigation policies. Due to the huge running cost, Earth system models (ESMs) may be difficult to flexibly provide the temperature projections following some specific emission pathways for empirical analysis. This study develops the mean and variability filed emulators in the high-resolution land grids to approximate the temperature behavior conditioned on GHG emissions in ESM. The emulator of mean temperature response is modeled as a function of GHG emissions to represent the expected values for ESM output, and the associated high-dimensional spatial dependence across grid points is estimated by the nearest-neighbor Gaussian process. The variability emulator is constructed with the residuals between the mean temperature response and the ESM output, and the associated space-time correlation structure is decomposed by principal component analysis and discrete Fourier transform. The analysis shows that the emulators trained with the runs of ESM only from part of representative concentration pathways can efficiently reproduce the temperature variations under different emission scenarios. The emulated gridded temperatures would be easily taken for climate impact and risk assessment, and be incorporated in the integrated assessment model for climate policy analysis.

Keywords

• temperature emulation
• hierarchical Bayesian
• nearest neighbor
• Gaussian process