Revisiting model complexity: Space-time correction of high dimensional variable sets in climate model simulations

Abstract

Multivariate bias correction (BC) models are well-known to correct more statistical attributes in climate model simulations. However, their inherent complexity and excessive parameters can introduce higher uncertainty into future climate simulations. In contrast, univariate BC models, with fewer parameters, are limited to correcting certain attributes. An issue that has not been investigated in-depth is the impact of an increased number of variables in the multivariate BC has on the bias-corrected climate models’ stability. This study compares the performance of a multivariate BC approach, Multivariate Recursive Nested Bias Correction (MRNBC), and a univariate BC approach, Continuous Wavelet-based Bias Correction (CWBC), as the number of variables to be corrected increases, known as the “curse of dimensionality” (CoD). The analysis uses high-resolution climate model outputs for both current and future simulations of sea surface temperature and precipitation in the Niño 3.4 region. Results show both BC models effectively correct current climate biases. As the number of variables increases, CWBC remains robust and produces sensible future simulations, while MRNBC’s complexity leads to deterioration in standard deviations and spatial cross-correlation. CWBC, based on univariate correction, is relatively unaffected by the CoD.