资源环境科技发展态势分析平台

Physical scaling (SP) method is a statistical downscaling approach where model-based climate data are downscaled taking into consideration large-scale climate, elevation and land-cover at the location of interest. In this study, the downscaling skills of an ensemble of SP method and its variants and Statistical DownScaling Model (SDSM) towards downscaling North American Regional Reanalysis (NARR) temperature data are compared. Two downscaling approaches: direct and indirect, two versions: SP and surrounding pixel information and three functional forms: linear regression, quantile regression and generalized additive models are considered to prepare the method ensemble. To evaluate method performance, a leave-one-out cross-validation approach is adopted. Results indicate that SP method and its variants have comparable skill to SDSM. Further method skill is found to be only marginally influenced by the choice of method version and functional form, and considerably influenced by the choice of approach. The ensemble of models is thereafter used to downscale future air temperature projections made by a global climate model: FGOALS-s2. It is found that downscaled future projections are most significantly influenced by the choice of version, followed by the choice of approach and the choice of functional form in the decreasing order of importance.