Prediction of Indian Ocean Dipole with Canonical Correlation Analysis.

The statistical prediction skill of Indian Ocean Dipole/Indian Ocean Sea Surface Temperature (SST) is assessed using Canonical Correlation Analysis (CCA) applied to individual and combined predictors set. NOAA Extended Reconstructed (ER) Sea Surface Temperatures (SST) version 5 (ERSSTv5) data were used as predictand over the Indian Ocean and lagged MSLP and 10-m zonal and meridional winds over the Indian Ocean and SSTs over the Pacific and Indian Oceans were used as predictors. CCA Skills from retroactive predictions, with a model developed for 1982-2010 and verified for 2011-2022, and Leave-5-Year-Out cross validation at 2-month lag were compared with skills obtained from 7 NMME models at 2-month lead time. Results show that SSTs are highly predictable over the western Indian Ocean but correlate poorly with observations over the eastern Indian Ocean. Both CCA and NMME perform poorly in January-April and best during September-December. Correlations between monthly observed and CCA and NMME predicted IODs range between 0.76 to 0.94 during September to December. NMME ensemble correlates the most in September with more than 88% explained variance. While the NMME IOD skills are comparable with that obtained from the CCA, all NMME members show large biases with RMSE as high as 4σ for CFSv2 compared to RMSE of 0.34σ for CCA in December. The biases in NMME IOD prediction can be improved by calibrating the raw NMME forecasts with observations.