Multivariate Bias Correction of reanalysis data for North-western Himalayan region

The hydro-climatic studies in the North-western Himalayan region primarily depend on the reanalysis datasets due to the inaccessible terrain. The reanalysis datasets appear biased compared to observation data, as physical processes in Numerical Weather Prediction (NWP) models are insufficiently represented. Although scientific progress in climate modelling is considerable, statistical attributes such as mean, variance, dependence structure between several variables and locations vary according to historical data. Therefore, raw simulations used in climate models can seriously affect the results of weather and climate models. Hence, Multivariate Bias Correction (MBC) approaches are used to overcome these inherent biases considering the covariance of multiple climatic variables and sites. This study applies MBC approaches to the reanalysis data of 12 stations of Indian states of Uttarakhand and Jammu Kashmir in the north-western Himalayas. The reanalysis data taken for this study is from the High Asia Reanalysis dataset and is compared with the observation data provided from the Princeton University global meteorological forcings at 0.25° horizontal resolution. The variables taken for the study are mean temperature, precipitation, specific humidity, wind speed, downward longwave radiation, downward shortwave radiation. The MBC approaches used are developed by (Cannon, 2017), in which MBCp method corrects Pearson correlation, MBCr method corrects Spearman rank correlation, MBCn method is based on N-dimensional PDF transform. The results show that all three MBC methods performed on the monsoon season data show a considerable shift towards observation for all the variables. The results analyzed with different error metrics such as RMSE, MAE, bias percentage show a significant reduction in error. The bias percentage in all variables is reduced to 0.01% in most of the variables for specific stations. The MBCp method seems to perform better than the other two methods in most of the sites and variables.