Forecasting all-India summer monsoon rainfall using regional circulation principal components: A comparison between neural network and multiple regression models

Pre-monsoon principal components (PCs) of circulation fields covering the S outh Asian subcontinent were used as predictors for all-India summer monsoo n rainfall (AISMR) over the period 1958-1993. Predictive skill of non-linea r neural network models and linear multiple regression models was compared using a bootstrap-based resampling procedure. Monsoon precursor signals rep resented by PCs were investigated and comparisons made with a recent observ ational and general circulation modelling study. Pre-monsoon PCs of the 200 hPa geopotential height field in May formed a co mpact, interpretable, and significant set of predictors for AISMR. Predicti ve skill was comparable to or better than that reported in prior modelling studies, each of which used optimized sets of regional and global predictor s. No improvement was noted when using data from multiple atmospheric level s, and skill at lead times more than 1 month prior to monsoon onset in June was poor. For May predictors there were only small differences in skill be tween the neural network and multiple regression models, although the neura l network results at longer lead times tended to be better than those shown by multiple regression. Interestingly, the 850 hPa PCs in January showed a maximum in predictive skill that was only evident in the neural network mo del results. The strength of this relationship suggests that further invest igation into the use of January 850 hPa predictors for the long-range forec asting of AISMR is warranted. Copyright (C) 1999 Royal Meteorological Socie ty.