Downscaling climate data
Earth system models (ESMs) simulate the evolution of the climate system over the entire globe up to thousands of years in the future, at a horizontal resolution of around a hundred kilometres.
The computation time required to simulate the global climate at a finer spatial resolution, of the order of tens of kilometres or less, remains high, even with current computing resources. As a result, climate phenomena or processes operating on scales finer than that of the computational grid are parameterized at best, or they’re simply ignored. Despite their slightly coarser resolution, ESMs produce results that are adequate for representing the large-scale climate. However, the crying need for finer spatial scale climate projections for specific regions has encouraged researchers to turn to methods that refine the simulations produced by ESMs. This is known as the statistical downscaling of climate data.
To date, downscaling techniques stem from two schools of thought, both of which rely on the existence of a quality simulation produced by an ESM. These are:
Dynamical downscaling: This approach relies on the use of regional climate models (RCMs) which, like ESMs, are physical climate models based on fluid mechanics. They are used to refine the horizontal resolution of the climate of a selected region of the globe. Their finer resolution (for example, 50 to 25 km with the CORDEX-CMIP5 protocol and 25 to 12 km with CORDEX-CMIP6) enables them to develop more detailed climate characteristics, due in part to a much more accurate representation of land surface features (such as mountains, coastal contours and the presence of lakes and rivers). To stay connected with the global climate, an RCM must obtain large-scale values of certain climate variables from an ESM at its boundaries. In this case, the RCM is said to be driven by an ESM. Although costly in terms of computing time, downscaling ensures that the hundred or so climate variables produced are consistent in time and space. What’s more, physical climate models such as RCMs and ESMs have the ability to have greenhouse gases and aerosols (whose quantities are governed by scenarios of future emissions or concentrations) interact with the other components of the climate system. This makes them suitable for simulating the future climate in the aim of studying climate change.
Statistical downscaling: The premise is that the characteristics of the observed climate at a fine scale can be derived from a series of large-scale climate variables (predictors). Various techniques can be used, like quantile mapping, multiple regression, stochastic generators and neural networks. These enable statistical relationships to be established between the conditions observed on a fine scale and predictors derived from climate data from the recent past (e.g. 1971-2020). These methods not only allow us to refine the spatial scale of climate simulations, but also to correct certain biases in climate models. When it comes to statistically downscaling an ESM simulation of the future climate, however, we have to assume that the statistical relationships established in the recent past will remain the same in a future climate. Various techniques exist to ensure that the climate change signal simulated by physical models is preserved in the downscaled data. Statistical downscaling is an inexpensive and rapid approach compared to dynamical downscaling, and is more easily applicable to large ensembles of climate simulations incorporating a variety of greenhouse gas and aerosol models and scenarios. Univariate statistical downscaling techniques, which treat each variable separately, are the oldest and most widespread. They are suitable for many applications that are based on a single variable, or for those where consistency between variables is not an issue. But in many cases, preserving spatial and temporal coherence between several variables is essential. The calculation of the Fire Weather Index is a good example. In cases like that, multivariate techniques designed for the joint downscaling of several variables are preferred.
