Downscaling — from global to local
What this page covers
Why raw climate-model output isn’t directly useful at country / admin-1 scale; the two methods that bridge the gap; how they differ in approach, cost, and the kind of answers they can support.
The resolution gap
To be written. CMIP6 GCMs are ~100-250 km native resolution. African admin-1 polygons (~50-200 km across) sit at the edge of one or two grid cells. Coastal effects, orographic precipitation, river basin scales all sub-grid. Hence the need to downscale.
Statistical downscaling
To be written. Big idea: use observed-climate patterns to map coarse model output to finer resolution. Methods: BCSD (Wood 2004), ISIMIP3BASD (Lange 2019), quantile mapping (Cannon 2015). Pros: cheap, fast, applicable to many GCMs at once. Cons: assumes stationarity of bias; can’t generate new physical detail (just spatial disaggregation).
Dynamical downscaling
To be written. Big idea: nest a regional climate model (RCM) inside a global GCM. Run the regional model at fine resolution; let physics generate the local detail. Pros: physically self-consistent; can resolve orography, sea breezes, monsoon variability. Cons: expensive (compute + storage); smaller ensembles; introduces another layer of model uncertainty (the RCM choice).
CORDEX — the dynamical downscaling community
To be written. Coordinated Regional Climate Downscaling Experiment. CORDEX-Africa (AFR-22 / AFR-44); CORDEX-CORE. CMIP5-era runs are mature; CMIP6-era runs are in progress. Reference: Giorgi & Gutowski 2015.
Which to use, when
To be written. Decision matrix. Statistical (NEX-GDDP-CMIP6, ISIMIP3b, CHELSA) for: most national / admin-1 adaptation work, large-ensemble framing, comparability across countries. Dynamical (CORDEX) for: extreme events / convection-permitting questions, physical-consistency-critical impacts, regional research with compute budget.
What the Atlas uses, and why
To be written. NEX-GDDP-CMIP6: 0.25° statistical downscaling. Cross-link to Dataset landscape for the full comparison.