Climate Change Canada

 

3.5.1: Downscaling strategies

Climate projections must be made using global models because many of the processes and feedbacks that shape the response of the climate system to external forcing operate at the global scale. For many applications, where only the change in some climate quantity is needed, global model projections can be used directly. This is because climate change normally applies over a much larger area than the climate itself, which can vary markedly over short distances in some regions. This is particularly the case for projected change in temperature, which has a very broad spatial structure, although local temperature may differ between, for example, the bottom of a valley and the surrounding hillsides.

However, for other applications, global climate model projections are not adequate, as they typically have horizontal spatial resolution, or grid spacing, of 100 km or coarser (see Figure 3.2 for explanation of model characteristics) (e.g., Charon, 2014). As an example, when using climate projections to drive a detailed hydrological model at the scale of a drainage basin, one needs values of future climate variables at a scale that respects local topographic, coastal, and other features, and represents high-frequency variability and extremes. Users of climate projections must therefore first evaluate whether they really need high-resolution climate scenarios or whether they could make effective use of lower-resolution climate change scenarios. Higher resolution, in and of itself, does not necessarily indicate higher-quality or more valuable climate information. But, for many applications, higher resolution may be necessary, and can also facilitate better understanding among, and communication to, users. A caution, however, is that internal climate variability is reduced by averaging results over large areas, and so as one goes from global to regional to local scale, internal variability becomes larger, leading to larger uncertainty in projections at the local scale, relative to that at regional or global scale (Hawkins and Sutton, 2009).

When climate information at a higher spatial- or temporal-resolution is needed, there are several approaches available to take global climate model projections and “downscale” them to higher resolution for a region of interest (or even a single location). These generally fall into two categories: statistical and dynamical downscaling.

Statistical downscaling is a form of climate model “post-processing” that combines climate model projections with local or regional observations to provide climate information with more spatial detail (Maraun et al., 2010; Hewitson et al., 2014). Statistical post-processing methods typically downscale to higher resolution and correct systematic model biases. A simple example is the so-called “delta method,” in which the change in some climate quantity, obtained from a climate model projection, is added to the observed historical value of that quantity. This allows projected changes from different climate models to be used in a consistent manner, since each model’s climatological bias is eliminated. Bias correction is particularly important when using downscaled climate information to drive impact models that depend on crossing absolute thresholds. For example, snow accumulation is sensitive to whether temperature is above or below freezing.

Simple techniques like the delta method may be suitable for some quantities, such as mean temperature, but not for others, such as daily precipitation, for which biases may be manifested differently, in variability, extremes, or dry/wet spells (Maraun et al., 2010). More complex statistical downscaling approaches are required in such cases, making use of detailed high-resolution observational datasets that reflect local topographic influences. These high-resolution data are used to interpolate low-resolution climate change projections to much higher resolution. In some cases, bias correction and other refinements are applied to correct statistical properties such as variances (Werner and Cannon, 2016). Yet other statistical downscaling methods take advantage of observed relationships between large-scale atmospheric circulation patterns, which are often well simulated by climate models, and local variables. By assuming that these statistical relationships remain fixed under a changing climate, climate model projections of circulation patterns can be used to make projections of future climate at a particular location. The statistical relationships introduce some aspects of local climate that may not be well represented in the driving global model (e.g., local topography and proximity to a lake). Fundamentally, all statistical downscaling methods assume that relationships between a model’s historical simulation and observations do not change over time, and that the information provided by the climate model and historical observations at their respective spatial scales is credible. Thus, the quality of statistical downscaling is directly related to the quality and availability of observational data. Recent critical reviews provide more information on the strengths and weaknesses of statistical downscaling and bias correction methods (Hewitson et al., 2014; Maraun, 2016).

Dynamical downscaling involves the use of a regional climate model, which is a physically based climate model (of the same level of complexity as a global model) that operates at high resolution over a limited area. Regional climate models incorporate much the same physical processes and scientific understanding as global climate models and indeed often share much of the same computer code. The important distinction is that regional climate models are driven at their lateral boundaries by output from a global climate model, as shown in Figure 3.10. The regional model also inherits errors and biases that may be present in the global model whose results are provided at the boundaries. The main advantage of dynamical downscaling is that, due to its limited area, a regional model can simulate climate on a much higher resolution than is possible with a global model, using a similar amount of computing effort. This additional detail is often desirable, particularly where the regional model output is used to drive another model (e.g., a hydrological model in which detailed basin geometry, high-frequency precipitation and extremes, and other small-scale features are essential). However, it remains an ongoing research topic to determine whether, and under what conditions, regional models add value relative to the original global model results that are being downscaled. There are no agreed-upon measures for added value (Di Luca et al., 2015; 2016; Scinocca et al., 2016), although the evidence available at the time of the IPCC Fifth Assessment indicated that there is added value in some locations owing to the better representation of topography, land/water boundaries, and certain physical processes, and that extremes are better simulated in high-resolution regional models (e.g., Flato et al., 2013).

An additional advantage of dynamical over statistical downscaling is that physical relationships between different climate variables (such as temperature and precipitation) are maintained. For very high-resolution dynamical downscaling (at model resolution of a few kilometres), physical processes such as convection can be resolved explicitly and can lead to improved simulation of climate variables such as precipitation extremes. Several recent studies indicate added value, including a dynamical downscaling system that includes a detailed representation of the Great Lakes (Gula and Peltier, 2012), the potential for added value near well-resolved coastlines (Di Luca et al., 2013), and evidence for improved simulation of temperature and precipitation extremes (Curry et al., 2016a,b; Erler and Peltier, 2016).

3.5.2: Downscaling results for North America and Canada

Both statistical and dynamical downscaling approaches have been applied and evaluated in many areas of the world. For North America, coordinated dynamical downscaling comparisons have been undertaken as part of the North American Regional Climate Change Assessment Program (NARCCAP: http://www.narccap. ucar.edu/) and the Coordinated Regional Downscaling Experiment (CORDEX: https://na-cordex.org/). For CORDEX, simulations were run at resolutions of approximately 25 km and 50 km. In both NARCCAP and CORDEX, Canadian models are represented. Coordinated experiments like these provide results from different regional climate models, driven at their boundaries by output from different global climate models. They also allow scientists to determine whether regional differences in projected climate change are related to the differences in the global driving models or to differences in the regional downscaling models. However, the CORDEX ensemble is considerably smaller than the CMIP global model ensemble, and studies using the CORDEX ensemble tends to focus on sub-regions rather than on Canada as a whole.

For Canada, regional climate models with smaller domains and higher resolution are being used, particularly by the Ouranos consortium and the Centre pour l’étude et la simulation du climat à l’échelle régionale (ESCER) at the Université du Québec à Montréal. Some of these simulations provide results at 15 km resolution (e.g., https://www.ouranos.ca/en/program/climate-simulation-and-analysis/). Statistical downscaling results are also readily available for Canada (https://www.pacificclimate.org/data/statistically-downscaled-climate-scenarios), with daily temperature and precipitation data at approximately 10 km resolution. These state-of-the art downscaling approaches (Werner and Cannon, 2016) are driven by multiple global climate model projections. In addition to the more detailed spatial structure, sophisticated statistical downscaling approaches can also provide estimates of future changes in climate extremes and other indices (such as frequency of hot days, growing season length, and drought indices) that are particularly important for certain impact studies (see Chapter 4). Downscaled results can also be used as inputs to impacts models — such as hydrological, crop, and ecosystem models — that are sensitive to variability on small spatial scales and to biases in climate models (Wood et al., 2004).