Global downscaled projections for climate impacts research (GDPCIR): preserving extremes for modeling future climate impacts
Abstract. Global climate models are important tools for understanding the climate system and how it is projected to evolve under scenario-driven emissions pathways. Their output is widely used in climate impacts research for modeling the current and future effects of climate change. However, climate model output remains coarse in relation to the high-resolution climate data needed for climate impacts studies, and it also exhibits biases relative to observational data. Treatment of the distribution tails is a key challenge in existing downscaled climate datasets available at a global scale; many of these datasets used quantile mapping techniques that were known to dampen or amplify trends in the tails. In this study, we apply the trend-preserving Quantile Delta Mapping (QDM) bias-adjustment method (Cannon et al., 2015) and develop a new downscaling method called the Quantile-Preserving Localized-Analog Downscaling (QPLAD) method that also preserves trends in the distribution tails. Both methods are integrated into a transparent and reproducible software pipeline, which we apply to global, daily model output for surface variables (maximum and minimum temperature and total precipitation) from the Coupled Model Intercomparison Project Phase 6 (CMIP6) experiments (O’Neill et al., 2016) for the historical experiment and four future emissions scenarios ranging from aggressive mitigation to no mitigation: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5 (Riahi et al., 2017). We use European Centre for Medium-RangeWeather Forecasts (ECMWF) ERA5 (Hersbach et al., 2018) temperature and precipitation reanalysis data as the reference dataset over the Sixth Intergovernmental Panel on Climate Change (IPCC) Assessment Report (AR6) reference period, 1995–2014. We produce bias-adjusted and downscaled data over the historical period (1950–2014) and for four emissions pathways (2015–2100) for 25 models in total. The output dataset of this study is the Global Downscaled Projections for Climate Impacts Research (GDPCIR), a global, daily, 0.25° horizontal-resolution product which is publicly hosted on Microsoft AI for Earth’s Planetary Computer (https://planetarycomputer.microsoft.com/dataset/group/cil-gdpcir/).
Diana R. Gergel et al.
Status: final response (author comments only)
- CC1: 'Comment on egusphere-2022-1513', Damien Irving, 22 Jan 2023
- RC1: 'Comment on egusphere-2022-1513', Anonymous Referee #1, 27 Feb 2023
- CC2: 'Comment on egusphere-2022-1513', Naomi Goldenson, 02 Mar 2023
- RC2: 'Comment on egusphere-2022-1513', Anonymous Referee #2, 22 Mar 2023
Diana R. Gergel et al.
CIL Global Downscaled Projections for Climate Impacts Research https://planetarycomputer.microsoft.com/dataset/group/cil-gdpcir
Model code and software
R/CIL GDPCIR dataset codebase https://zenodo.org/record/6403794#.Y6t4sezMJAc
Dodola codebase https://zenodo.org/record/6383442#.Y6t5Y-zMJAc
Diana R. Gergel et al.
Viewed (geographical distribution)
One of the most simple methods for creating "application ready" climate projection data is to calculate an array of adjustment factors representing the change in each quantile between an historical (e.g. 1995-2014) and future (e.g. 2045-2065) GCM (i.e. relatively coarse spatial resolution) simulation. Using daily GCM data you might calculate quantile changes for each month, so you end up with a 100 by 12 array of adjustment factors. You can then directly apply those adjustments to observational data (i.e. over 1995-2014) that is at a much higher spatial resolution, in order to produce a new higher resolution climate projection dataset (i.e. for 2045-2065). For instance, if the first day (e.g. 1 January 1995) in your observational dataset at a particular grid point is 20 degrees Celsius and that corresponds to the 30th percentile of January temperatures in the observations, you simply apply the 30th percentile GCM adjustment factor (after regridding the adjustment factors to the observational spatial grid) for January to that 20 degree day (to get the temperature for 1 January 2045). This is basically the method used in the latest climate projections for Australia: https://www.climatechangeinaustralia.gov.au/en/obtain-data/application-ready-data/scaling-methods/
I'd be interested to know if the authors think the approach I've decribed above would produce similar results to QDM-QPLAD, as it's not an approach canvassed in their introduction? QDM-QPLAD is significantly more complicated, so it would be interesting to hear the benefits of that added complexity.