On deep learning-based bias correction and downscaling of multiple climate models simulations

Bias correcting and downscaling climate model simulations requires reconstructing spatial and intervariable dependences of the observations. However, the existing univariate bias correction methods often fail to account for such dependences. While the multivariate bias correction methods have been developed to address this issue, they do not consistently outperform the univariate methods due to various assumptions. In this study, using 20 state-of-the-art coupled general circulation models (GCMs) daily mean, maximum and minimum temperature (T-mean, T-max and T-min) from the Coupled Model Intercomparison Project phase 6 (CMIP6), we comprehensively evaluated the Super Resolution Deep Residual Network (SRDRN) deep learning model for climate downscaling and bias correction. The SRDRN model sequentially stacked 20 GCMs with single or multiple input-output channels, so that the biases can be efficiently removed based on the relative relations among different GCMs against observations, and the intervariable dependences can be retained for multivariate bias correction. It corrected biases in spatial dependences by deeply extracting spatial features and making adjustments for daily simulations according to observations. For univariate SRDRN, it considerably reduced larger biases of T-mean in space, time, as well as extremes compared to the quantile delta mapping (QDM) approach. For multivariate SRDRN, it performed better than the dynamic Optimal Transport Correction (dOTC) method and reduced greater biases of T-max and T-min but also reproduced intervariable dependences of the observations, where QDM and dOTC showed unrealistic artifacts (T-max < T-min). Additional studies on the deep learning-based approach may bring climate model bias correction and downscaling to the next level.

Automated summary

This study comprehensively evaluates the Super Resolution Deep Residual Network (SRDRN) deep learning model for climate downscaling and bias correction. The SRDRN model effectively removes biases based on the relative relations among different GCMs and retains the intervariable dependences for multivariate bias correction. The results show that the SRDRN model outperforms other methods in reducing biases and reproducing the intervariable dependences of the observations.