Near Surface Air Temperature is an important climatic variable that affects the hydrological response of a river basin, and forms an input to most of the hydrological models. General Circulation Models (GCMs) simulate the response of temperature and other climate variables to the variations in emission concentrations, but their outputs are too coarse to be used in most hydrological models. A multi stage statistical downscaling approach is proposed for downscaling GCM predicted temperatures. In the first stage the Relevance Vector Machine (RVM) is used to develop a statistical model between the GCM simulated historical climate variables and the observed historical temperature for spatially downscaling monthly GCM simulations. A weather generator is then used to generate daily data from the spatially downscaled temperature data. On fine scales, lack of correlation between precipitation and temperature data used for hydrological modelling can lead to large uncertainties in the generated hydrological series. Thus, a distribution free post processing is performed for reproducing the observed regional correlation between temperature and precipitation, in the generated temperature data. The methodology is then applied to the Bharathapuzha catchment in Kerala, India, to downscale temperature from the climate models BNU-ESM, CESM1-BGC, CMCC-ESM2, FGOALS-G2, FIO-ESM-2.0 and MIROC4h. The statistical models set up using RVM show consistent performance during the calibration (1969-1980) and validation (1981-2005) phases, with Nash-Sutcliffe efficiency (NSE) between 0.64 to 0.83. The weather generator is then run to generate daily temperature data from the monthly downscaled series. Across the different climate models, daily maximum temperature is generated with RMSE between  2.5°C to 3.3°C, while the minimum temperature has RMSE ranging from 1.7°C to 2.0°C. The probabilistic nature of the procedure enables the generation of multiple series from the same set of predictors. The simulation band from the multiple GCMs is studied for the period 2016 to 2021 to understand the deviation in predicted temperature for the future scenario. The prediction band for maximum temperature has an average band width of 6.7°C and for minimum temperature, the average band width is 4.9°C.

Statistical downscaling plays an important role in reducing the uncertainty in regional-scale climate change studies. A major assumption of all statistical downscaling studies is the stationarity in the relationship identified from the historical observations. This assumption is difficult to validate since the future is truly unknown. The proposed methodology tries to overcome the limitations of this assumption by considering global physics that drive the regional climate to select the predictors to develop the statistical relationship. Since the statistical model is developed with climatic processes as a background, it can be said with higher confidence that the modelled relation would remain sound for the future as well. The proposed methodology is divided into two stages. In the first stage, a Relevance Vector Machine model is developed using historical observed global predictors that are identified to have teleconnections with the regional climate predictand. The spatial downscaling of this regional predictand has been done on a monthly scale. The bias associated with the downscaled predictand is removed by separating the anomalies from the prediction and adding that to the historical mean of the observed data. This monthly series is further disaggregated into daily series using a weather generator. The non-stationarity in the climate projections is accommodated in the weather generator and more regional features of the climate is integrated into the predictand in this stage. The proposed methodology is validated by downscaling rainfall over a river basin in India and its performance is analysed. The downscaled data is seen to reproduce characteristics of daily rainfall like consecutive wet days, the number of rainy days, wet spell duration, and dry spell duration with a bias of less than 10%.

A large number of General Circulation Models (GCMs) are currently available for modelling the atmospheric conditions over the Earth. However, there is a large variability in the future climate predicted by the available set of climate models. Hence, the climate data introduces the most amount of uncertainty in the climate change impact assessment. Regional-scale climate change impact studies based on these models may produce a wide range of possible impacts that becomes unusable for policymakers. A robust GCM selection procedure is introduced in the current study to bring the uncertainty to a realistic range. The proposed approach takes into account the process representation in the climate models by checking teleconnections in data along with their ability to predict the regional climate in spatial and temporal scale. The interdependence between the climate models are also accounted for in the proposed approach to avoid underestimation of uncertainty. The procedure is validated in the Bharathapuzha River Basin, Kerala, India. The study considers 22 GCMs that participated in the Coupled Model Inter-comparison Project-5 and 6 Regional Climate Models (RCMs) that are recommended for the Indian subcontinent. The climate models BNU-ESM, CMCC-CM, GFDL-ESM2G, GFDL-ESM2M and MPI-ESM-MR are found to be performing well for the prediction of both precipitation and temperature. The proposed climate model selection procedure can bring down the band width of uncertainty from 376 mm to 162 mm in monthly rainfall prediction with a containing ratio of 44%. The downscaling of the climate predictions can further increase the containing ratio by removing the systematic error. The bandwidth of uncertainty has reduced from 10.82 K to 3.83 K in the prediction of minimum temperature and from 8.35 K to 4.52 K for maximum temperature. The proposed GCM selection procedure provides more confidence in the predicted future climate since regionally significant correlations between climate variables are preserved in the selected models. The model selection procedure is validated for the period 2006-2018 with the observed climatic variables, and the selected models are found to be performing well.