Bias adjustment and downscaling
Global climate models cannot be used as they are for local impact studies, for two reasons. Their statistics differ from the observed climate: a model can be too warm, too dry or rain too often in a given place. And their grid cells are large, about 150 km across on average, while impact studies need a field, a city or a catchment. Bias adjustment deals with the first problem, downscaling with the second (Vrac et al., 2025).
Bias adjustment
Section titled “Bias adjustment”Bias adjustment, also called bias correction, compares the statistics of a climate model with those of a reference dataset over a calibration period, then corrects the model so that its statistics match: not only the mean but the variability and the whole distribution of values. The same correction is applied to the future simulation. The most widely used method is quantile mapping (Vrac et al., 2025); Maraun and Widmann (2018) review the methods.
Downscaling
Section titled “Downscaling”Downscaling increases the spatial resolution of climate model data. There are two families of methods.
- Dynamical downscaling. A regional climate model, driven by a global one, simulates the climate of a region at a finer resolution, for example about 12 km in the EURO-CORDEX simulations over Europe. It is physically based but needs a lot of computing, so it exists for a few global models and regions only, and its results often still need bias adjustment (Vrac et al., 2025).
- Statistical downscaling. Statistical relations between the large-scale climate and the local climate, learned from observations or a reanalysis, are applied to the climate model data.
Both in one step
Section titled “Both in one step”When the reference is finer than the model, bias adjustment also downscales: the model is corrected towards the local distribution of the reference, cell by cell, so the corrected data take on the reference’s resolution (Vrac et al., 2025). This is how our climate projections are made: the CDF-t method adjusts each climate model to the ERA5-Land reanalysis at 0.1°, about 10 km, from model grids of 0.9° to 2.5° (Noël et al., 2022). Unlike simple quantile mapping, CDF-t keeps the climate change signal of the model. See Data processing for the full chain and Limitations for what these methods cannot do.
References
Section titled “References”- Vrac, M., Loukos, H., Noël, T., & Defrance, D. (2025). Should we use quantile-mapping-based methods in a climate change context? A “perfect model” experiment. Climate, 13(7), 137. https://doi.org/10.3390/cli13070137
- Noël, T., Loukos, H., Defrance, D., Vrac, M., & Levavasseur, G. (2022). Extending the global high-resolution downscaled projections dataset to include CMIP6 projections at increased resolution coherent with the ERA5-Land reanalysis. Data in Brief, 45, 108669. https://doi.org/10.1016/j.dib.2022.108669
- Maraun, D., & Widmann, M. (2018). Statistical Downscaling and Bias Correction for Climate Research. Cambridge University Press. https://doi.org/10.1017/9781107588783