Bias Correction of Ocean Bottom Temperature and Salinity Simulations From a Regional Circulation Model Using Regression Kriging

It is well known that climate and circulation model simulation output are often systematically biased. However, existing bias correction methods typically ignore spatial autocorrelation of the biases and correct only the overall mean and variance, resulting in mismatched spatial variability between bias‐corrected simulations and observations. In this study, we propose using regression kriging (RK) to correct for biased spatial patterns and apply this method to Regional Ocean Modeling System (ROMS) simulated ocean bottom temperature and salinity for the Mid‐Atlantic Bight, USA. RK combines modeling non‐stationary trends using (generalized) regression with ordinary kriging (OK) of the regression residuals. We compared the performance of RK to a simpler OK method and a quantile mapping (QM) method often used for bias correction of such model output. These methods were evaluated using the Structural Similarity (SSIM) index that can simultaneously evaluate model accuracy, precision, and spatial similarities. Our results show that while both RK and QM can correct for overall biases of the mean and variation, only RK can effectively reduce the spatial‐temporal autocorrelation of the biases. The RK method was able to bias correct while preserving the spatial‐temporal trends of the ROMS simulated bottom temperature and salinity surfaces. The RK approach can easily be applied to any similar climate and circulation model simulation output. This study has profound implications for studies that use the output from such a model for fine‐scale mapping, habitat suitability modeling, species distribution modeling, or predicting the effects of climate change.