A Multi-stage Stochastic Approach for Statistical Downscaling of Rainfall

The study proposes a statistical downscaling procedure for regional rainfall. In the multistage procedure, the spatial downscaling stage using Relevance Vector Machine (RVM) model captures the climate change signals in the monthly scale simulations from the climate models by considering a predictor set that is based on the regional climatic phenomena. The spatially downscaled monthly rainfall is then disaggregated into daily scale using a modified weather generator. In the temporal downscaling stage using a weather generator, the transition matrix is modified to account for the non-stationarities in the rainfall. The proposed methodology is validated in the Bharathapuzha River Basin, India by downscaling rainfall data from climate models, BNU-ESM, CESM1-BGC, CMCC-ESM2, FGOALS-G2, FIO-ESM-2.0, and MIROC4H. The process-based indices specified in the VALUE framework developed from the downscaled rainfall data closely matches the indices developed from the observed rainfall data for the historical period, with absolute PBIAS less than 14% across all the indices considered. The reduction in uncertainty achieved is studied using average band width of the simulations, which is seen to reduce from 186 mm to 52 mm after downscaling. The capability of the downscaling procedure for capturing the non-stationarity in the climate is tested by comparing the performance of the procedure over warm and cold phases of ENSO, and it is found to be satisfactory. The study proposes a multi stage, stochastic approach for statistical downscaling of rainfall.It combines the strength of RVM models and weather generators.The statistical downscaling relationship is based on the regional climatic phenomena.Downscaled rainfall series captures the occurrence and distribution characteristics of regional observed rainfall and its non-stationarity.