Seasonal prediction of summertime rainfall in south China using multi-model ensemble products

Abstract

General circulation model (GCM)-based dynamical forecast systems are commonly used for seasonal predictions. Some large-scale dynamical vairables such as the mean sea level pressure, upper air temperature, zonal wind, etc., are well predicted. However, GCMs show very low skills in predicting some variables such as the local-scale precipitation. In order to interpolate the large-scale climate information from GCMs to the regional scale, downscaling techniques based on either dynamical or statistical models need to be used. This research addresses this issue by using statistical schemes to downscale GCM-based seasonal forecast outputs for local-scale rainfall prediction over South China (SC) in boreal summer. Hindcast experiments from 11 global models and their multi-model ensemble (MME) average were considered. Singular value decomposition analysis based on observed data showed that the precipitation variation over SC is strongly related to that in the mean sea level pressure (SLP) over Southeast Asia, the western north Pacific and Indian Oceans. Hence, SLP was chosen as the predictor for predicting the SC station-scale rainfall variations. A statistical downscaling scheme based on the "perfect prognosis" approach (PP), was first developed and evaluated. Based on the relationship between the observed SLP and observed station-scale precipitation using SVD analysis, a statistical scheme was constructed. Station-scale precipitation is then predicted by replacing observations with model outputs for SLP, in the precipitation SVD reconstruction. Downscaling based on hindcast data of SLP from 11 GCMs and their MME average indicated that PP-based downscaling has difficulties in improving the rainfall prediction in central SC. In general, rainfall prediction based on PP downscaling outperforms the direct model output (DMO) at western SC for most model except BCC and NCEP. The exceptional models show improvement in eastern SC. However, the improvement of the PP-based downscaling is limited since the scheme is unable to correct systematic biases of GCMs. A similar SVD analysis was then repeated to obtain a statistical relationship between model SLP and the observed station precipitation. Results showed that there is also strong covariability between model hindcast SLP and observed station precipitation. Hence another downscaling approach, namely model output statistics (MOS), was adopted for predicting the SC station-interpolated precipitation with model SLP as predictors. Compared with DMO, the improvement of rainfall prediction based on the MOS-type downscaling is limited in two distinct geographical locations. For first group of models (labelled as Type 1), improvement is mainly found in western SC near Guangxi. For second group of models (labelled as Type 2), improvement is mainly seen in the eastern coastal area. Further analysis revealed that dynamical models have difficulties in capturing the regional circulation details in SLP over SC, leading to erroneous prediction in some locations. The statistical method is able to map the large-scale circulation patterns to station-scale rainfall variations, thereby correcting some of the biases over land. To conclude, statistical downscaling helps to increase the prediction skill where DMO performs poorly.

Date of Award	2 Oct 2013
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chi Yung Francis TAM (Supervisor)