linked in logo | IS-ENES | Contact

Exploring climate model data


Downscaling’ is a general concept that embraces various methods for increasing the spatial resolution and reduce some of the biases in order to improve the usability of climate scenarios. Often, scenarios from comparatively coarse-resolution global climate models (GCMs) are implied, even though this is not a prerequisite. Basically, there are two fundamentally different approaches to this, dynamical downscaling and statistical downscaling.

Dynamical downscaling

Dynamical downscaling makes use of a regional climate model (RCM) having higher spatial resolution (typically 10–50 km) over a limited area and ‘fed with large-scale weather’ from the GCM at the boundaries of the domain. A regional climate model is conceptually similar to a global climate model; both focus on the dynamical and physical processes that governs the weather and thus the climate, the basic physical processes are the same although the parameterisations differ because of the different resolution, and the numerical approaches for solving the equations are similar.

Statistical downscaling

Statistical downscaling, or empricial statistical downscaling (ESD), takes a statistical approach to the same problem. A reference dataset (often one or several time-series of meteorological station data) is used to calibrate the climate scenario data. The main principle is to find a statistical relationship between the observational dataset and a corresponding reference period of the climate model data. Originally, a major focus was to adapt GCM scenario data for use as input to impact models, typically a hydrological model.

A major source of difference – or bias – between GCM data and the observational reference dataset is the huge difference in spatial resolution. Because of this, and the history of development of these methods, “downscaling” is sometimes equated with “downscaling of GCM data”. Even though RCMs provide data at a much higher resolution than the GCMs, even higher higher precision is sometimes warranted. If this is the case statistical downscaling methods can equally well be applied to RCM data. In this case the term ‘further downscaling’ is sometimes used to distinguish this processing step from the dynamical downscaling achieved by the RCM.

Bias correction

Statistical downscaling is sometimes equated with bias correction. The reason is that in statistical downscaling the bias due to the different scales of the GCM output data and the reference dataset is lumped together with the true GCM biases (i.e. the model imperfections not related to scale issues). However, bias correction can equally well be applied to RCM data, in which case the focus is more on combined model biases of the GCM and RCM. As much of the scale transition is handled at the RCM modelling stage bias correction in this context is sometimes referred to as RCM calibration.

Pro's and con's of the two approaches

The two approaches, dynamical and statistical downscaling are not quite interchangeable alternatives, and there are many technical aspects to consider if one need to choose one of them. Here are some main characteristiscs listed.

Main advantages of dynamical downscaling (RCMs):

  • Individual variables are physically consistent in time and space, and the different variables are internally consistent.

  • The same fundamental physical principles are used in both an RCM and a GCM.

  • An RCM provides for large region a wealth of output data at high resolution compared to what can be obtained from a GCM.

  • No specific calibration data is required.

Main disadvantages of dynamical downscaling (RCMs):

  • RCMs are very complex and requires substantial computational resources, often at same level as required for GCM simulations (however, large datasets covering many regions are made freely available by modelling institutes).

  • Near the boundary of the RCM domain artefacts and spurious effects occur.

  • While removing much of the GCM bias that is related to the coarse resolution, an RCM also adds its own biases to the output data.

Main advantages of empirical statistical downscaling(ESD):

  • The methods are computationally inexpensive.

  • Many different statistical methods are available, allowing for substantial flexibility.

  • ESD typically includes bias correction as an integral part of the process.

Main disadvantages of empirical statistical downscaling(ESD):

  • A calibration dataset, typically a long meteorological record of high qulaity is required. Any quality problems in the calibration data will be tranferred to the downscaled GCM/RCM data.

  • The higher the requirements regarding spatial and temporal consistency, or inter-variable consistency are, the more complex and computationally demanding the statistical procedures become.

  • The ESD approach requires/assumes stationary statistical relationship, the relationship must remain constant under climate change.

Related links



The ENES3 project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 824084.