Bias-corrected Down-scaled Multi-Model Ensemble Prediction of the Indian Summer Monsoon Rainfall using Self-Organizing Map

English Abstract:

The probabilistic large scale forecasts though useful for Extended Range Prediction (ERP) of the Indian Summer Monsoon Rainfall (ISMR) over larger regions of Indian subcontinent, it has been found that the regional forecast outlook misrepresents the amplitude of rainfall variability. This is a common problem in a large scale Global Circulation Model. E.g. the rainfall prediction over north-east India, south-peninsular India and the regions with elevated orography show large underestimation of rainfall amplitude and limited forecast skill at fourth pentad lead time. Hence, bias correction of the large scale forecast and its downscaling to smallest possible regions need to be explored rigorously. In this study, Self-Organizing Map (SOM) has been used to bias-correct and downscale the CFS v2 generated ERP forecast of ISMR. SOM is used to identify the patterns from the low-resolution (1x1, IMD) observed data. Then dynamical model (CFSv2) generated data is projected onto these patterns to generate the bias-corrected forecast and reconstructed using a high resolution observed data (0.25x0.25, TRMM) so that we finally get the bias-corrected downscaled prediction of the ISMR.

Three versions of CFSv2 [T126, T382 and the bias corrected GFS (GFSbc) at T126] is used in this study to generate total 129 [(11 members of CFST126+11 members of CFST382+21 members of GFSbc) x3 different lattices] members. Mean of these are given as the deterministic forecast. The brier skill score and the pattern correlation shows significant improvement in the forecast skill.