Toward regional climate-change downscaling of weather statistics using a hidden Markov model

Climate change projections of daily rainfall statistics are investigated for the summer monsoon over India using a hidden Markov model (HMM) together with GCM projections under the A2 emissions scenario from the CMIP3 model database. The GCM projections exhibit a greater likelihood of more extreme rainfall on rainy days along with an increase in the probability of dry days. However, as is well known, the GCM daily rainfall marginal distributions are highly truncated, and exhibit too few wet days. An HMM trained on the MPI-ECHAM5 GCM's gridded daily rainfall for the end-of-20th vs end-of-21st century periods, indicates that the change in rainfall PDF is associated with (a) an increase in the prevalence of the dry state—thus leading to more dry days, together with (b) a general increase in wet-day mean amounts in all states. The latter changes indicate important non-stationarity in the NHMM's rainfall-state parameters, that need to be incorporated before the model could be used in climate-change downscaling as a function of GCM-projected dynamical and thermodynamic changes in circulation and specific humidity. To address this non-stationarity, we explore a modification to the NHMM in which an additional input variable is included that influences the exponential parameters of the state rainfall distributions, through a generalized linear model.