The objective of PCIC’s Downscaling Intercomparison Project was to provide a rigorous evaluation of the comparative strengths and weaknesses of several popular statistical methods of downscaling climate extremes.
Meeting the need for practical local-scale climate information is a key challenge for climate scientists working from large-scale global climate models (GCMs). GCMs typically simulate values for temperature and precipitation as averages over fairly large spatial areas, leaving analysts guessing about potentially important local variations within those areas. Such local-scale information is required to understand the possible impacts of climate variability and change on climate extremes.
To bridge the gap between global scale GCM output and local scale impacts, climate scientists have developed a number of 'downscaling' techniques, using either regional climate models or various statistical methods. A large number of statistical downscaling approaches are available, and thus some guidance on which method is most appropriate for a particular application and locale is desirable.
Methods
Initially, five methods were compared: Automated Statistical Downscaling (ASD), Bias-Corrected Spatial Disaggregation (BCSD), Expanded Downscaling (EDS), Quantile Regression Neural Networks (QRNN) and TreeGen (TG). These techniques were tested for their ability to represent local climate extremes in British Columbia, a region that presents a significant challenge for downscaling due to its widely varying topography. The performance of the various techniques was assessed through the use of the so-called Climdex indices, a set of 27 climate indicators used by the World Meteorological Organization for measuring climate extremes.
Temperature and precipitation data from seven meteorological stations in four different climate zones of British Columbia were used to determine values and distribution for each index. Three statistical tests were applied for each combination of downscaling method and index in order to ascertain the method's ability to reproduce the distribution of each index.
Results
Initial results appear to favour EDS overall as the most reliable method among the five downscaling techniques tested for representing climate extreme indices in the study areas. However, individual test results vary widely by region and by index. Figure 1 and Figure 2 show summarized results of the model intercomparison by study region and index, respectively. These results are being submitted for publication in a peer reviewed scientific journal. Additional downscaling methods and study areas are planned for inclusion in the project's final report, expected in December 2011.
Figure 1: Graph showing the percentage of tests passed for each of the five statistical downscaling methods, across all 27 Climdex indices and grouped by major climate region. EDS appears to have performed best overall, and overwhelmingly so on the BC coast, though QRNN performed better in the northern taiga region of the province.
Figure 2: Graph showing the percentage of tests passed across all five statistical downscaling methods and study areas for each of the 27 Climdex indices for representing climate extremes. For detailed definitions of each index see http://cccma.seos.uvic.ca/ETCCDI/list_27_indices.shtml.
Acknowledgements
- BC Ministry of Forests and Range, Future Forests and Ecosystems Scientific Council (FFESC)
- BC Ministry of Environment