Statistical downscaling of precipitation using machine learning techniques

Statistical models were developed for downscaling reanalysis data to monthly precipitation at 48 observation stations scattered across the Australian State of Victoria belonging to wet, intermediate and dry climate regimes. Downscaling models were calibrated over the period 1950–1991 and validated over the period 1992–2014 for each calendar month, for each station, using 4 machine learning techniques, (1) Genetic Programming (GP), (2) Artificial Neural Networks (ANNs), (3) Support Vector Machine (SVM), and (4) Relevance Vector Machine (RVM). It was found that, irrespective of the climate regime and the machine learning technique, downscaling models tend to better simulate the average (compared to other statistics) and under-estimate the standard deviation and the maximum of the observed precipitation. Also, irrespective of the climate regime and the machine learning technique, at the majority of stations downscaling models showed an over-estimating trend of low to mid percentiles (i.e. below the 50th percentile) of precipitation and under-estimating trend of high percentiles of precipitation (i.e. above the 90th percentile). The over-estimating trend of low to mid percentiles of precipitation was more pronounced at stations located in dryer climate, irrespective of the machine learning technique. Based on the results of this investigation the use of RVM or ANN over SVM or GP for developing downscaling models can be recommended for a study such as flood prediction which involves the consideration of high extremes of precipitation. Also, RVM can be recommended over GP, ANN or SVM in developing downscaling models for a study such as drought analysis which involves the consideration of low extremes of precipitation. Furthermore, it was found that irrespective of the climate regime, the SVM and RVM-based precipitation downscaling models showed the best performance with the Polynomial kernel.