Evaluation of the gridded data and simulated products
As an illustration, the left panel of Fig. 1 shows quantile-quantile (Q-Q) plots between grid and station observations at the 10 observing stations for three indices, the number of days with daily maximum temperature greater than 30 °C (TX30), the length of growing season (GSL) for warm season crops, and the number of days with precipitation amount greater than 1 mm (Prep1). Indices computed from the gridded observational data compare well with those computed from station data. There are no apparent systematic differences for most indices (not shown). The difference in the elevations of the grid box and the station at Agassiz (black dots) explains the differences in temperature indices, while the effect of smoothing by the gridding procedure can explain the slightly larger number of precipitation days from the gridded data.
To compare multiple model simulations with observations at a particular location for 1950–2010, we first estimate quantiles in the simulated data for individual models separately and then compute the ensemble average of the quantiles from the 24 GCMs. The central panels of Fig. 1 displays Q-Q plots that compare model ensemble average quantiles with quantiles from gridded observations for the three indices. There are systematic biases in indices directly computed from the GCM simulations. The temperature indices have different biases in both magnitude and direction at different locations. Compared with gridded observations, the models simulate too many precipitation days at all locations. Various biases also exist for other indices (not shown). This indicates the need to bias-correct the model data in order to compute indices properly.
The right panels of Fig. 1 shows Q-Q plots that compare ensemble-averaged quantiles from the downscaled datasets with those from the gridded observations. As expected, biases seen in the raw GCM simulations essentially disappear in the downscaled data. Biases are also minimal for the other indices, indicating that the downscaling procedure has corrected model biases as intended. As the downscaling procedure is not trained against the indices from the gridded data, correction of bias for these indices does provide confidence in the usefulness of the procedure. This is consistent with the findings of Cannon et al. (2015) and Ficklin et al. (2016) and satisfies a key requirement for the data to be suitable in projecting changes in climate indices at regional and local scales. Comparison with homogenized data at 80 observing stations (Vincent et al. 2012; Mekis and Vincent 2011) provides essentially the same results (not shown).
Projected changes in indices
Seneviratne et al. (2016) showed that changes in extreme temperature and precipitation over large regions scale with global temperature increase across emission scenarios. Linking regional or national impacts to specified global warming levels provide a convenient way to communicate climate change information to adaptation community. Figure 2 shows the projected changes in Canadian mean temperature plotted against global warming levels. The global warming level is defined in terms of the 20-year running average of multi-model ensemble mean of global temperature relative to its pre-industrial (1861–1900) level. It is clear that long-term changes in Canada’s mean temperature scale with corresponding changes in global mean temperature across the three emission scenarios and at the rate about twice of global mean temperature. Changes in regional averages of indices in turn scale well with changes in Canada’s mean temperature across emission scenarios. As an illustration, we plotted in the right panel of Fig. 2 changes in the length of the growing season in the Canadian Prairies (including Alberta, Saskatchewan, and Manitoba) as a function of Canadian mean temperature change. This is also the case for other indices as well, as can be seen by comparing projected changes at a global warming level of 2 °C under RCP4.5 and RCP8.5 (see Supplementary Figs. S2, S3, S4). We therefore present projected changes for the near term (2031–2050) and for the end of the twenty-first century (2081–2100) under the RCP8.5 scenario. In all cases, project changes in indices are taken with respect to a 1986–2005 baseline period. The corresponding global mean temperatures projected by the 24 GCMs are 2.1 and 4.5 °C above the pre-industrial level, respectively; which will be referred to as the 2.1 and 4.5 °C warming levels in the remainder of the paper.
To facilitate the discussion, we show projected ensemble median changes due to anthropogenic influence for selected indices as maps. In addition, we provide the median, the 5th, and the 95th percentiles from the ensemble for large regions in Tables S4 and S5. The spatial regions include British Columbia (BC), the Canadian Prairies (Prairie), Ontario (ON), Quebec (QC), the Atlantic Provinces (ATL; including New Brunswick, Nova Scotia, Newfoundland and Labrador, and Prince Edward Island), and the northern territories (NN; including Yukon, Northwest Territories, and Nunavut).
Figure 3 displays projected changes in the number of days with daily maximum temperature above 30 °C (TX30; “hot days”) and the number of days with daily minimum temperature below − 15 °C (TNm15; “cold nights”). Consistent with projected warming, models project an increase in the number of hot days and a decrease in the number of cold nights. Projected increases in hot days frequency can exceed 10 and 40 days at the 2.1 and 4.5 °C warming levels for southern Canada, with the largest increase in regions where hot days occur frequently in the current climate, such as in southern parts of Ontario, Quebec, and the Prairies. Even with large warming, however, temperatures in northern Canada will still be too low to have hot days. Models project a larger decrease in the number of cold nights for the north than in southern Canada, reflecting amplified high-latitude winter warming. Many regions in Canada will progressively experience unprecedented warmth. For example, hot nights (daily minimum temperature above 20 °C, TN20) occur historically only in the southern part of southeastern Canada. The location at which at least half of the downscaled model simulations have TN20 nights progressively expands northward with time, reaching almost 70° N by the end of this century under the RCP8.5 scenario (Fig. S5). Models project the disappearance of extreme cold nights (daily minimum temperature below − 25 °C, TNm25) by the end of this century in southern Ontario, New Brunswick, Nova Scotia, and southern Newfound and Labrador (not shown).
Models project the lengthening of the frost-free (FF) and killing-frost-free (KFF) periods, with the largest increase in British Columbia (BC) and eastern Canada and a smaller increase in the Prairies and northern territories (Fig. S6). The smaller changes in the Prairies and northern territories reflect the more pronounced annual cycle in regions with strongly continental climates, where the climatological crossing-time of the threshold occurs on the steepest part of the seasonal march of temperatures thus the change in the period above or below, a given temperature threshold is smaller. Models project an increase in the frost-free period for BC of up to 77 days at the 4.5 °C global warming level; even the interior of the province, which shows the smallest increase, is projected to have an increase in frost-free period of more than 55 days.
Models also project large changes for cooling and heating degree-days, two important indices for utility planning (Fig. S7). Simulations project a substantial increase in cooling degree-days for southern Canada, especially in the most populated area of the country. A large decrease in heating degree-days is projected, especially in the north. This means that there will be stronger energy demand to meet summer cooling needs but weaker demand to meet winter warming needs. This in turn has implications for the utility sector. In fact, such a shift in the timing of energy demand has already shown its impact in Ontario’s electrical energy market, there is now a summer peak, due to the recent growth in air conditioning (OPSE 2012).
Tables S4 and S5 summarize regional averages of projected changes in climate indices at 2.1 and 4.5 °C global warming levels. To aid interpretation, we provide a brief description of projected changes for some indices for Ontario at the 4.5 °C global warming level (Table S5). The median projected increase in the number of hot days (TX30) is 38 annually, with the variation in projected increases between different models ranging from 15 to 56 days. Correspondingly, the number of hot spell days (TXc30) and the length of the longest hot spells (TXL30) are projected to increase by 33 and 11 days, respectively. The median projected increase in the number of hot nights (TN20), number of hot spell nights (TNc20), and length of the longest hot spell nights (TNL20) are 17, 12, and 5 nights, respectively, a pattern similar to but less pronounced than that for hot days. The projected increase in hot day and hot night frequency implies more intense, prolonged future heat waves in Ontario. Accordingly, the number of heat wave days with a daytime high temperature above 30 °C and a nighttime low temperature above 20 °C is projected to increase by 15 days.
Precipitation related indices
Figure 4 shows multi-model median changes in the number of days with precipitation, number of days with more than 10 mm of precipitation (Prep10), and precipitation intensity. Overall, the models project an increase in the number of days with precipitation in the north, especially north of 60°N, and a small decrease in parts of southern Canada. The region with an increase in the number of days with precipitation expands southward towards the end of the twenty-first century. Averaged over regions, it projected about 5 more precipitation days annually. The northern region and Quebec are exceptions, where increases of more than 15 days are projected at the 4.5 °C global warming level (Table S5). Projected changes in the annual maximum dry spell length (defined as periods of consecutive days with precipitation less than 1 mm) have the opposite spatial pattern (not shown), with a projected decrease in the north and small or no change in the south. Precipitation intensity is projected to increase, with a larger percentage increase of over 25% in the north at the 4.5 °C global warming level. The smallest projected percentage increases are seen in the Prairies and Nunavut, perhaps indicating differences in the ability of the atmospheric circulation to deliver moisture to the different regions of Canada. Projected changes in the number of days with precipitation over 10 mm have a spatial pattern similar to that of projected changes in precipitation intensity.
Figure 5 displays projected changes in growing season length for three main crop types at two global warming levels. As expected, the models project a lengthening in the growing season for warm season and overwintering crops as the climate warms. Climate models project an increase in the length of growing season for warm season crops in southern Canada by more than 20 days and 50 days at the 2.1 and 4.5 °C global warming levels, respectively. In contrast, the ensemble median for the length of growing season for cool season crops decreases by up to 10 days in many regions, although not all models agree with the sign of changes. The decrease is due to earlier termination of the growing season in the summer. The accelerated physiological maturity of these crops with earlier occurrence of high temperatures (the growing season terminates on the first occurrence of daily maximum temperature greater than 30 °C for five consecutive days) is more than offsets increases from an earlier start to the growing season. Nevertheless, the length of this growing season, when averaged across regions, is projected to increase slightly (Tables S4 and S5). Models also project an increase in other heat-related indices (not shown), including crop heat units (CHU) and effective growing degree-days (EGDD).
Spring-seeded small grain crops, such as spring wheat, require at least 1200 EGDD (Centre for Land and Biological Resources Research Branch, Agriculture and Agri-Food Canada 1995). The models project significant northward expansion of area with at least 1200 EGDD, which may reach as far north as Yukon at the 4.5 °C warming level (see Fig. 5, left). The projected warming may therefore open up new opportunities for the northward expansion of agriculture if other elements such as soil conditions and water availability are suitable and crop physiology permits. Yields for crops that can take advantage of additional heat may also improve. Qian et al. (2017) showed that global mean warming by up to 2.0 °C may be beneficial for Canadian crop production, though yield starts to decrease beyond the 2.0 °C warming level because of added water stress associated with increased evaporative demand.
Uncertainty in the projections
For temperature-related indices, projections by different models at the 2.1 °C global warming level agree on the sign of changes for almost all regions, with only a few exceptions, consistent with warming. At the 4.5 °C global warming level, the models agree with the sign of changes for every temperature index and for every region, with only two exceptions (i.e., growing season length for cold season crops in the Prairies and Ontario). This indicates the robustness of projected changes in temperature indices for the future. In contrast, uncertainty in projected precipitation changes is large; the models do not fully agree on the sign of changes in most regions at the 2.1 °C global warming level, but the models do agree on the sign of changes in most regions at the 4.5 °C warming level. As summarized in Tables S4 and S5, the spread in the projected changes increases with greater global warming. To a first approximation, regional changes in the indices scale with global mean temperature change. Stronger warming in the projected global mean temperature is associated with larger uncertainty, which is largely due to model uncertainty since there is a lack of evidence for an increase in the natural internal variability with warming. It follows that the increase in the uncertainty of projected changes in the indices is also largely due to model uncertainty. Note that the perception of uncertainty may be affected by the choice of reference period (e.g., whether earlier or later in the observational record). This is because the mean anomaly is constrained to be zero during the reference period while the anomaly is not constrained in this way beyond the reference period. As a result, the spread amongst models widens as the distance in time between the base period and the projection horizon increases (see also Hawkings and Sutton 2016).
The downscaling procedure preserves observed spatial and temporal variability of precipitation and temperature at high-resolution. Preserving temperature trend is important as the models are able to simulate observed changes in Canadian temperature (Wan et al. 2018). It also removes model biases. Thus, the downscaled products are better suited for impact assessment. It should be noted however that the projected future changes by the downscaled products are not necessarily more creditable than those by the underlying climate model outputs. In many cases, especially for absolute threshold-based indices, projections based on downscaled data have smaller spread because of the removal of model biases (e.g., Fig. 1). As an example, in Ontario, the spread in projected changes in the number of days with daily maximum temperature greater than 30 °C computed from the statistically downscaled data is smaller than that computed directly from the GCMs (Fig. S8). However, this is not the case for all indices. Downscaling from GCM resolution to the fine resolution needed for impacts assessment increases the level of spatial detail and temporal variability to better match observations. Since these adjustments are GCM dependent, the resulting indices could have wider spread when computed from downscaled data as compared to those directly computed from GCM output. In the latter case, it is not the downscaling procedure that makes future projection more uncertain; rather, it is indicative of higher variability associated with finer spatial scale.