A model output statistic-based probabilistic approach for statistical downscaling of temperature

Large-scale temperature projections need to be downscaled to river basin scale to facilitate a regional scale climate change impact assessment. A multi-stage statistical downscaling procedure is proposed in the current study, the first stage captures the climate change signals from the simulations of general circulation models (GCMs) by spatially downscaling the monthly GCM simulations. The second stage disaggregates the spatially downscaled monthly series to a daily scale by a weather generator which adds the regional climatic information into the spatially downscaled time series. A distribution-free post-processing shuffling is finally performed to rebuild the intervariable correlation of downscaled temperatures with regional rainfall which is important in reliable projection of streamflow. The procedure is validated by downscaling the maximum and minimum temperatures over the Bharathapuzha catchment in India for the period 1951–2005. The downscaled series of temperature shows Normalised Root Mean Square Error (NRMSE) less than 0.09 and correlation coefficients greater than 0.4. The ability of the procedure in capturing non-stationarity in the climate is also analysed by its performance in different phases of ENSO.