Introducing Quantile Mapping to a Regression Model Using a Multi-Model Ensemble to Improve Probabilistic Projections of Monthly Precipitation

Abstract

A new method was proposed for the probabilistic projection of future climate that introduced quantile mapping to a regression method using a multi-model ensemble (QM_RMME). Results of this method were then compared with those of the traditional regression method (RMME). Six stations in Japan where 100 year observation records were available were used to evaluate the performance of the methods. An initial 50-year period (1901–1950) was used to develop the regression models and the final period (1951–2000) was used for evaluation. Results showed that the estimation errors at the 50th and 90th percentile were smaller for QM_RMME as compared to RMME at most sites. Conversely, when the model development and evaluation periods were limited to 20 years (1901–1920 and 1951–1970, respectively), the 90th percentile error was larger for QM_RMME. This was attributed to quantile mapping resulting in over-fitting of the data during the model development period. Furthermore, the QM_RMME error increased when the difference of observations between the model development and verification periods was large. Therefore, results indicated that the RMME method was more stable for relatively short data verification periods.