Sparse Canonical Correlation Analysis Postprocessing Algorithms for GCM Daily Rainfall Forecasts

Abstract

Abstract We introduce two variants of canonical correlation analysis (CCA) for model output statistics of GCM forecasts of daily rainfall. These approaches link the coarse-gridded GCM forecasts with the reference field through a projection onto highly correlated basis vectors to address the recurrent errors in daily rainfall forecasts due to spatial bias and subgrid variability. The first model, namely, sparse CCA (SCCA), includes the sparsity feature into the ordinary CCA to provide a reduced number of canonical coefficients. The second model (B-SCCA) employs the bagging approach to reduce the variance in the predictions due to the sample variability in the derived canonical series. The models are tested using simulated data imposed with a strong spatial bias, and then using subseasonal rainfall forecasts provided by the NASA GMAO GEOS model under the SubX project, as well as gridded rainfall data (MSWEP product) for the region of South Korea. A linear regression model is chosen as the baseline postprocessing algorithm and ordinary CCA is also evaluated against the proposed models. As for the simulated data, the SCCA model confirms its ability to address spatial bias in forecast fields compared with the baseline model. For the actual forecasts, the leading improvements of SCCA and B-SCCA over the baseline model are for the S1 skill score, suggesting that these models offer a relative gain in reproducing the spatial gradient of the reference rainfall field, which is relevant in hydrological applications that require a sound representation of spatial variability. Our results also highlight the importance of prefiltering the input data before applying CCA in such settings.