Hourly Precipitation Downscaling With Empirical Statistical Methods: Case Study near the US Eastern Coast

Abstract:

Statistical downscaling methods are widely used to reconcile the spatial and temporal difference between climate model outputs and input requirements of climate change impact assessments. Quantile-mapping, due to its simplicity and general good performance, is one of the most popular. However, its simplicity also leads to variations in the procedure among studies, and it is not always well cross-validated. There is also less attention paid to sub-daily precipitation downscaling which is important for urban hydrologic purposes. The study explores the potential of quantile-mapping by comparing four of its variations, and extends its application by linking it to an hourly k nearest-neighbor resampling disaggregation. The methods are trained with NCEP/NCAR 40-year reanalysis. Seven-fold cross-validation on various mean, variability, and extreme precipitation statistics is performed with hourly observations at rain gauges along the eastern coast of the United States. The best performance on mean monthly precipitation and annual maximum 5-day precipitation is found when quantile-mapping is trained without regard to seasonality. However, the best performance on all other statistics is contingent upon the method being trained separately for each month and combined with a monthly bias-correction factor. Deficiencies in quantile-mapping carry over to the hourly disaggregation step. When the downscaled daily precipitation is satisfactory, the hourly disaggregation performs very well on all statistics after careful selection of its two parameters. These results demonstrate the importance to consider seasonality in the implementation of quantile-mapping and highlight some clear, though limited, deficiencies. That it is feasible to generate realistic hourly precipitation series with easy-to-implement empirical methods is also of interest to water managers.