Frequently asked questions on downscaling

Downscaling refers to the process of transforming coarse-scale climate information, typically global climate model output, into fine-scale climate information. There are two approaches to downscaling: dynamical downscaling and statistical downscaling.

Typical spatial resolutions of current global climate models range from approximately 100 to 300 km. Such coarse resolutions cannot represent all of the regional characteristics that could affect local climates. Statistical and dynamical downscaling provide climate information that represents local details while remaining consistent with the evolution of climate as represented by global climate models.

Statistical downscaling refers to the use of empirical statistical models to represent relationships between large-scale climate data (provided by global models) and small-scale local effects, such as the temperature or rainfall variability at a given location.

Advantages of statistical downscaling:

• The statistical models are calibrated with observations. This removes biases when applied to present-day climate simulations, though not necessarily when applied to future projections. (Note that model error could potentially influence the results if an observational dataset used for calibration of the statistical scheme was produced with the aid of a climate or weather model. For example, atmospheric 'reanalyses' could be used as surrogates for large-scale atmospheric observations. A reanalysis is made by using observations to constrain a weather forecast model. The resulting dataset tends to agree much better with observations than the model would on its own, but model errors might still influence the results.)

• Spatial coverage and resolution is determined by the extent and density of the observing network (rather than by computer power as in the case of global and regional climate models).

Disadvantages of statistical downscaling:

• Statistical relationships between local variables and large-scale averages that are empirically derived from present-day data may not necessarily be valid in a changed future climate. The assumption that such relationships remain valid in a changed climate is often referred to as the stationarity assumption.

• Statistical relationships can only be obtained at locations where adequate observations are available (such as weather stations with long records spanning many decades).

• Downscaled variables are limited to those for which adequate observations are available to calibrate a statistical model.

Dynamical downscaling refers to the use of regional climate model to downscale the climate simulations produced by global climate models. A regional climate model represents the evolution of meteorological variables such as winds, temperature, and humidity at higher spatial resolution than a global climate model. Regional climate models are driven on the boundaries by global climate model data. A regional climate model is a self-consistent physical model of the climate system. It numerically solves the equations of physics that are appropriate for its spatial resolution. Global climate models and global numerical weather prediction models also do this, but there are differences in how the different types of models are implemented.

Advantages of dynamical downscaling:

• Dynamically downscaled results are not restricted to regions with adequate observational coverage, as opposed to statistically downscaled results (information on statistical downscaling). In Canada, for example, the south of the country is well covered by an extensive network of weather stations, but coverage in the north is much sparser.

• Dynamical downscaling does not rely on the assumption that the relationship between local variables and large scale averages stays constant over time (the stationarity assumption). Instead a physically-based model is used to predict how this relationship may change over time.

Disadvantages of dynamical downscaling:

• No climate model is perfect. The equations of physics as solved by global climate models involve approximations, mainly due to their limited spatial resolution. The same is true of regional climate models: although they have finer resolution than global climate models, approximations are still involved. This means that there is some amount of model-related uncertainty in all regional climate model results.

• Regional models also inherit uncertainty from the global climate models used to drive them. This is because there is model-related uncertainty in global model results, and regional models are strongly constrained to follow the large-scale atmospheric evolution that is prescribed by their parent global models. For example, if a global climate model predicts that the Pacific jet stream will shift northward under climate change, a regional model driven by this global model would predict a similar jet stream shift, but this fact does not provide any additional confidence that the predicted shift is correct.

A global climate model numerically solves the equations of physics that determine the atmospheric circulation over the whole planet. The output from such a model includes physical variables affecting climate such as winds, temperatures, and precipitation. A regional climate model does the same thing, but for a limited spatial domain (North America, for example, is one possible domain).

Reducing the domain from global to regional allows regional climate models to run at higher spatial resolution than global models using similar computational cost. Higher resolution is useful to represent fine-scale regional characteristics that may affect local climate but cannot be resolved by coarser global models. These could include local topographic variations found in mountainous regions such as the Rockies, convoluted shapes of coastlines like the coast of British Columbia, or the many small lakes found throughout much of Northern Canada.

Regional climate models are used to downscale the climate simulations made by global climate models. This process is referred to as dynamical downscaling.

It would be ideal to conduct global climate simulations at a higher spatial resolution so that regional topographic characteristics and fine-scale atmospheric features (individual clouds, for example) could explicitly be resolved by global climate models. Unfortunately this is not computationally feasible at the present time. The computational cost of a climate model increases by roughly a factor of ten each time its spatial resolution is doubled. For example, refining the spatial resolution of a model from 100 km to 10 km would require about a thousand-fold increase in computing power.

The domain is the region represented by a regional climate model. Examples of domains include North America, Africa, or Europe. An international project called the Coordinated Regional Climate Downscaling Experiment (CORDEX) exists to produce coordinate sets of regional climate projections and help scientists evaluate results from regional climate models. The standard domains that it uses are listed online, and include Arctic and North American domains.

Regional climate model domains are typically chosen to be large enough to represent coherent atmospheric circulation features across the region of interest, but small enough that the computational cost of the model is much less than that of a high-resolution global climate model. For example, a regional model covering North America only represents about 10% of the global surface area, allowing about twice the spatial resolution of a global model for a similar computational cost. Regional climate models are thus an efficient way to achieve higher spatial resolution in a chosen region.

An artificial constraint is introduced into a regional climate model because it is bounded at the edges of its domain. Since the real atmosphere circulates globally, climate in any region is affected not only by local characteristics of the region such as type and shape of its land surface but also by events occurring outside the region. For example, an air mass with distinct properties could enter a region, such as when an Arctic air mass moves into southern Canada. Another example is the remote influence of El Niño on North American temperature and rainfall during winter. Due to its limited spatial domain, a regional model cannot explicitly represent these kinds of influences from afar. Hence these influences must be communicated to the model by prescribing information on its boundaries. This is accomplished by driving the regional model on its lateral boundaries with climate data obtained from a global model.

Regional climate models should therefore be considered as extensions of global climate models: they extend the capabilities of coarse global models by adding fine regional-scale detail, and are constrained by the driving on their boundaries to closely follow the evolution of global models. The desired result is that a regional climate model provides a fine-resolution representation of a region's climate that is nevertheless consistent with the coarse-resolution representation that is provided by a global climate model. In this sense regional climate models add spatial detail to the predictions made by global climate models.