1. Introduction
Heavy precipitation events are likely to increase almost everywhere over the globe owing to a changing climate (e.g., Cubasch et al. 1995; Zwiers and Kharin 1998; Kharin and Zwiers 2005; Tebaldi et al. 2006; Meehl et al. 2007). The Fourth Assessment Report (AR4) of the Intergovernmental Panel on Climate Change (IPCC) has indicated that precipitation intensity is projected to very likely increase over the world late this century under global warming (Alley et al. 2007). Many studies have specifically focused on projecting changes in future annual mean precipitation and frequency of heavy rainfall events, using general circulation model (GCM) projections. For example, using a six-GCM-model ensemble, Emori and Brown (2005) showed that annual mean precipitation is projected to increase by 20%–50% at high latitudes (i.e., poleward of 50°N/S) by the end of this century. Tebaldi et al. (2006) employed simulations from nine IPCC AR4 GCMs to assess changes in future precipitation extremes and found that in the high latitudes of the Northern Hemisphere there are the most coherent regional patterns of significant increases in the intensity of precipitation extremes. Other studies (e.g., Zwiers and Kharin 1998; Kharin and Zwiers 2005) have determined that the increase in magnitude of future projected extreme precipitation is greater than that of annual mean precipitation. As pointed out by Kharin and Zwiers (2005), while the globally averaged annual mean precipitation rate is projected by the Canadian Centre for Climate Modelling and Analysis (CCCma) Coupled General Circulation Model (CGCM1 A2) to increase by less than 3% by the end of this century, the corresponding increase in the 20-yr return values of annual extremes of 24-h precipitation rates is projected to be more than 12%.
However, large-scale GCM simulations are not suitable for local-scale climate change impact analysis. To overcome the discrepancy between two scales, downscaling methodologies are widely used to derive local- or station-scale future rainfall information. One of the leading techniques for doing this is statistical (empirical) downscaling (Wilby and Wigley 1997). Many statistical techniques have been used to develop downscaling transfer functions, as reviewed by Fowler et al. (2007). The statistical downscaling methods are generally divided into three major groups: regression-based models, weather-typing schemes, and weather generators.
Previous multiple regression approaches to develop precipitation downscaling transfer functions had difficulty with simulating statistical properties of daily precipitation processes. For example, Nguyen et al. (2006) applied the statistical downscaling model (SDSM) built by the multiple regression method to downscale daily precipitation in Montreal region (Quebec, Canada); the coefficients of determination R2 after calibration of the downscaling transfer functions are very low, ranging from 0.062 to 0.098. One of the major reasons for this weak regression result is that multiple regression approaches are not suitable for some weather variables, like precipitation, with a nonnormal distribution. To overcome this problem, alternative techniques that can cater for nonnormal distributed data should be employed to downscale daily precipitation. To achieve this, a number of studies (e.g., Buishand et al. 2004; Abaurrea and Asín 2005; Fealy and Sweeney 2007) have constructed the downscaling transfer functions of daily precipitation by employing logistic regression as an occurrence model and generalized linear modeling (GLM) approach as a quantity model. The results obtained from these studies have shown that use of both logistic regression and GLM offers a significant improvement over multiple linear regression. However, some of the studies indicated that these daily rainfall downscaling transfer functions still possess low model R2s (Buishand et al. 2004) and have difficulties with predicting extreme precipitation events (Fealy and Sweeney 2007). Haylock et al. (2006) have employed six statistical downscaling methods to downscale daily precipitation over the United Kingdom. Of the six methods examined, four employed artificial neural networks, one used canonical correlation analysis, and another one is the regression-based statistical downscaling method—SDSM developed by Wilby et al. (2002). The results revealed that “the inter-model differences between the future changes in the downscaled precipitation indices were at least as large as the differences between the emission scenarios for a single model” (Haylock et al. 2006). This implies that when projecting future local-scale precipitation information, different types of downscaling methods should be considered.
Atmospheric circulation-type classifications, such as sea level air pressure patterns, have also been employed to construct their relationships with daily rainfall, which can be applied to derive local-scale/station-scale future daily rainfall information from large-scale GCM simulations (e.g., Goodess and Palutikof 1998; Kostopoulou and Jones 2007a,b). An alternative approach—automated synoptic weather typing (or air mass typing)—has been employed in evaluation of climate impacts on a number of environmental issues (Cheng et al. 2010). The synoptic weather typing is able to characterize a complex set of meteorological variables as a coherent index (Kalkstein 1979; Perry 1983), using not only hourly sea level air pressure but also hourly surface and upper-air observations of temperature, dewpoint, u wind, υ wind, and total cloud cover.
However, to date, it appears that the automatic synoptic weather typing has not been employed to downscale future daily rainfall climate information from large-scale GCM simulations in Canada. The current study employs the synoptic weather typing and a number of linear and nonlinear regression techniques to downscale future daily rainfall from large-scale GCM simulations to the selected river basins in Ontario, Canada. This method is dependent on the stationarity of the past 50-yr predictor–predictand relationships under future climate conditions. The downscaling scheme is built upon the previous studies (i.e., Cheng et al. 2008, 2010), which is made up of a three-step process. First, within-weather-type daily rainfall simulation models were developed and validated using synoptic weather typing with cumulative logit regression and nonlinear regression procedures (Cheng et al. 2010). As demonstrated in the study by Cheng et al. (2010), the results obtained imply that it is necessary to perform synoptic weather typing prior to the development of daily rainfall simulation models. Furthermore, for development of daily rainfall simulation models, the study has used a number of the atmospheric stability indices in addition to the standard meteorological variables (commonly used in most of the previous downscaling papers). Second, regression-based downscaling transfer functions developed by Cheng et al. (2008) are adapted to project station-scale future hourly meteorological variables (except rainfall) that were used in development of daily rainfall simulation models. Hourly meteorological variables include air temperature, dewpoint temperature, sea level air pressure, total cloud cover, and south–north and west–east wind speed. Future hourly climate projections were derived from temporal downscaling transfer functions (i.e., relationships between hourly and daily weather variables derived from observations). Third, using downscaled future hourly climate projections, future daily synoptic weather types and then future daily rainfall quantities can be projected by applying synoptic weather typing and within-weather-type rainfall simulation models.
This paper is organized as follows. In section 2, data sources and their treatments are described. Section 3 summarizes within-weather-type daily rainfall simulation modeling and statistical downscaling developed from recent studies (Cheng et al. 2008, 2010) since methods and results from both studies are used in the current analysis. Section 4 presents the analysis techniques as applied to projection of future weather types and assessment of climate change impacts on daily rainfall: method I and method II. Section 5 includes the results and discussion on future rainfall-related weather types, robustness of rainfall simulation models, changes in frequency of future daily rainfall events and seasonal rainfall totals, changes in future return-period values of extreme rainfall events, and uncertainty of the study. The conclusions and recommendations from the study are summarized in section 6.
2. Data sources
Historical observations for the period April–November 1958–2002 used in the study by Cheng et al. (2010) to develop daily rainfall simulation models were also used in this current study. The warm season (April–November) was selected since most of the extreme rainfall events in the study area occur during this period. These observations include hourly surface meteorological data and U.S. National Centers for Environmental Prediction (NCEP) six-hourly upper-air reanalysis weather data. These data consist of air temperature, dewpoint temperature, sea level air pressure, total cloud cover, south–north and west–east wind speed. In addition to hourly meteorological data, daily rainfall data observed at the climate stations within each of four selected river basins (Fig. 1) were used to calculate daily river basin–average rainfall quantities, representing average rainfall conditions for the catchment. A number of climate stations (13, 12, 13, and 9 for Grand, Humber, Rideau, and Upper Thames River basins, respectively) were selected for the analysis. As described in the study by Cheng et al. (2010), one of the major reasons for using river basin–average daily rainfall is that it is suitable for simulation of rainfall-related streamflow volumes, as part of the project.
Study area and location of four selected river basins in Ontario, Canada. Dots are climate stations having daily observations and stars are location of the cities with meteorological stations having hourly observations.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
In addition to historical observations, daily climate change simulations from three GCMs and two emission scenarios were used in the study, summarized in Table 1. These models and scenarios include the following: one Canadian GCM—CGCM2 [IPCC Special Report on Emissions Scenarios (SRES) A2 and B2]; one U.S. GCM—Geophysical Fluid Dynamics Laboratory Climate Model version 2.0 (GFDL CM2.0; IPCC SRES A2); and one German GCM—ECHAM5–Max Planck Institute Ocean Model (MPI-OM; IPCC SRES A2). A2 and B2 scenarios considered different assumptions of future greenhouse gas (GHG) emissions derived from a distinctly different direction for future population growth, economic development, and technological change. From Environment Canada’s Web site (Environment Canada 2006), it is seen that the scenario A2 is similar to the IPCC “business-as-usual” scenario. Compared to scenario A2, the scenario B2 produces much lower GHG emissions and aerosol loadings in the future and projects less future warming, especially in the second half of this century. Both scenarios were used in the study to generate a range of projections of possible climate change impacts on future daily rainfall events. From the GCM experiments of stationary simulations for the specified time slices, daily data for all used surface weather variables (i.e., maximum and minimum temperatures, sea level air pressure, and west–east and south–north wind speed) were included in the study. In addition to surface climate information, daily upper-air GCM simulations of temperature and west–east and south–north winds on standard atmospheric levels (i.e., 925, 850 700, 600, and 500 hPa) were used in the analysis. For these GCM simulations, the three time windows (1961–2000, 2046–65, 2081–2100) were used in the analysis because these data are only available from the Web site of the Program for Climate Model Diagnosis and Intercomparison (Program for Climate Model Diagnosis and Intercomparison 2006). Furthermore, for projections of future return-period values of extreme rainfall events, two Canadian GCM transient model simulations for a 100-yr period (2001–2100) were included (Table 1).
GCM simulations and scenarios used in the study.
3. Summary of the previous studies
As part of this research, Cheng et al. (2010) have developed simulation models of daily rainfall quantities and statistical downscaling transfer functions for standard meteorological variables (Cheng et al. 2008). Since both studies were used in this current paper to project changes in frequency of future daily rainfall events, it is necessary to outline major methods used in and results derived from both studies.
a. Summary of daily rainfall modeling
A recent study by Cheng et al. (2010) employed an automated synoptic weather typing as well as stepwise cumulative logit and nonlinear regression analyses to simulate the occurrence and quantity of daily rainfall events. The synoptic weather typing was developed using principal components analysis (PCA), an average linkage clustering procedure, and discriminant function analysis to identify the weather types most likely to be associated with daily rainfall events for the four selected river basins in Ontario. The entire suite of 144 weather variables during the period April–November 1958–2002 were used in synoptic weather typing, which are hourly surface weather observations of six elements: air temperature, dewpoint temperature, sea level air pressure, total cloud cover, and south–north and west–east wind speed. Using daily 13-component scores produced by the PCA that explained 91% of the total variance, the average linkage clustering procedure and discriminant function analysis resulted in 24 major synoptic weather types for the study area. Of these weather types, 10 synoptic weather types were identified over the 45-yr period as primary rainfall-related weather types. These 10 weather types can capture 73%–77%, 92%–93%, and 95%–98% of the rainfall events with daily rainfall greater than or equal to 0.2, 10, and 25 mm, respectively, across the selected river basins.
As described in the study by Cheng et al. (2010), 10 rainfall-related weather types are associated with synoptic weather patterns: cold front I, cold front II, cold front III, cold low, cyclone I, cyclone II, mesohigh, quasi-stationary front, warm front I, and warm front II. The weather pattern was identified for each of the weather types, based on a subjective examination of a number of surface weather maps associated with heavy rainfall events. As shown in Fig. 2, weather types or weather patterns occur in different seasons. The weather patterns labeled with “I” occur most frequently in the summer (June–August) and usually possess much warmer characteristics than do the same weather patterns labeled with “II,” which occur most often in the spring (April–May) and fall (September–November). Cold front III was identified for the weather type that can occur throughout all seasons and that typically has thermal conditions between cold front I and cold front II. In addition, within-weather-type frequency of daily rainfall events and mean daily rainfall quantities in the Thames River basin for the period April–November 1958–2002 are shown in Table 2. Similar results were also found for other selected river basins but not shown owing to the limitations of space.
Monthly mean number of days occurring with each of 10 identified rainfall-related weather types or weather patterns in Thames River basin for the period 1958–2002.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
Within-weather-type daily mean rainfall amount and number of daily rainfall events in Thames River basin for the period April–November 1958–2002.
Within-weather-type daily rainfall simulation modeling comprises a two-step process: 1) cumulative logit regression to predict the occurrence of daily rainfall events, and 2) using probability of the logit regression, a nonlinear regression procedure to simulate daily rainfall quantities. The 228 predictors used in development of daily rainfall event occurrence simulation models include not only the standard meteorological variables but also a number of the atmospheric stability indices (e.g., lifted index, Galway 1956; K-index, George 1960; total totals index, Miller 1972). To avoid effects of multicollinearity among explanatory variables, the PCA was used once again to convert intercorrelated meteorological variables into uncorrelated principal component scores, which were then used as predictors for the cumulative logit regression. In addition, to effectively distinguish heavy rainfall events, the atmospheric stability indices were rearranged as dummy variables based on their relationships with daily rainfall quantities. Across the four selected river basins, the daily-rainfall-event occurrence simulation models revealed that there are significant correlations between the occurrence of daily rainfall events and model simulations. As described in the study by Cheng et al. (2010), the models’ concordances, derived from cumulative logit regression, range from 0.82 to 0.96 (a perfect model would have a concordance value of 1.0). Of the total 40 simulation models (10 models for each of the 4 selected river basins), 11, 19, and 33 models possess concordances greater than 0.92, 0.90, and 0.87, respectively. To evaluate performance of daily rainfall quantity simulation models, the four correctness levels of “excellent,” “good,” “fair,” and “poor,” as shown in Table 3, were defined based on the absolute difference between observed and simulated daily rainfall amounts. As described in the study by Cheng et al. (2010), the proportion of simulations on daily rainfall quantities that fell into excellent and good categories was much higher than the proportion that fell into fair and poor categories. Cheng et al. (2010) have found that, across the four selected river basins, the percentage of excellent and good daily rainfall simulations ranged from 62%–84%. In addition, it is noteworthy that the rainfall simulation models are able to capture most of daily heavy rainfall events (i.e., ≥32.5 mm) with the percentage of excellent and good simulations: 62%, 68%, 70%, and 81% for Grand, Thames, Humber, and Rideau River basins, respectively.
Criteria used to evaluate daily rainfall quantity simulation models. Diff is the absolute difference between observed and simulated rainfall and obs is the river basin–average daily observed rainfall.
The entire methodology used in the study by Cheng et al. (2010), which is comprised of synoptic weather typing and rainfall simulation modeling, was validated by randomly selecting one-fourth of the total years for the weather data. The remaining three-fourths of the total years were used for model development. The validation dataset is therefore independent from the data sample used in the development of the models. The validation results showed that the models were successful at replicating occurrence and quantity of daily rainfall events with similar results to the model simulations outlined above (refer to Cheng et al. 2010 for details).
b. Statistical downscaling of meteorological variables
To project future daily rainfall, downscaled future hourly climate information for the standard meteorological variables (excluding rainfall) that were used in development of daily rainfall simulation models is needed. These meteorological variables include surface and upper-air temperature, dewpoint, west–east and south–north winds, sea level air pressure, and total cloud cover. To derive future hourly station-scale climate information from GCM-scale simulations, a regression-based downscaling method developed by Cheng et al. (2008) was adapted for this current study. This downscaling method comprises a two-step process: to spatially downscale daily GCM simulations to the selected weather stations in south-central Canada and then to temporally downscale daily scenarios to hourly ones.
The downscaling transfer functions were constructed using different regression methods for different meteorological variables since a regression method is suitable only for a certain kind of data with a specific distribution. A number of regression methods, such as multiple stepwise regression, cumulative logit regression, orthogonal regression, and autocorrelation correction regression, were used to develop downscaling transfer functions. Performance of the downscaling methods was evaluated by 1) analyzing model R2s of downscaling transfer functions, 2) validating downscaling transfer functions using a leave-one-year-out cross-validation scheme, and 3) comparing data distributions, diurnal/seasonal variations, and extreme characteristics of the weather variables derived from downscaled GCM historical runs with observations over a comparative time period of 1961–2000. The results showed that regression-based downscaling methods performed very well in deriving daily and hourly station-scale climate information for all weather variables. The strong correlations between station-scale predictands and GCM-scale predictors are similar between model calibrations and validations. Most of the daily downscaling transfer functions possess model R2s greater than 0.9 for surface temperature, sea level air pressure, upper-air temperature, and winds; the corresponding model R2s for daily surface winds are generally greater than 0.8. The hourly downscaling transfer functions for surface air temperatures, dewpoint, and sea level air pressure possess the highest model R2 (>0.95) of the weather elements. The functions for south–north wind speed (υ wind) are the weakest model (model R2s ranging from 0.69 to 0.92 with half of them greater than 0.89). For total cloud cover, hourly downscaling transfer functions developed using the cumulative logit regression have concordances that ranged from 0.78 to 0.87 with over 75% greater than 0.8. Details of the daily and hourly downscaling methodologies as well as the results of downscaling transfer functions’ calibration and validation are not presented in this current paper owing to the limitations of space (refer to Cheng et al. 2008 for details).
4. Analysis techniques
To project climate change impacts on frequency of future daily rainfall events, in addition to daily rainfall simulation models and future station-scale climate simulations, future synoptic weather types are required for this study. The principal methods and steps used in this study are summarized in Fig. 3. This section focuses on methodological description for projection of future weather types and assessment of climate change impacts on rainfall.
Flowchart of methodologies and steps used in the study.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
a. Projection of future weather types
Using downscaled future hourly climate simulations, discriminant function analysis is able to project future daily weather types. To determine future weather types, principal component scores for each of the future days were calculated by multiplying the posteigenvector matrix by the standardized future climate data matrix. The posteigenvector, derived from the developmental dataset by PCA for synoptic weather typing (constructed in the study, Cheng et al. 2010), was used, so that future component scores are comparable with the postscores since both used the same set of eigenvectors. In addition, to more effectively compare component scores from both historical and future datasets as well as to remove the GCM bias, future downscaled climate data were standardized using the mean and standard deviation of downscaled GCM historical runs (1961–2000). Using the centroids of the predetermined weather types derived from the observed data, discriminant function analysis can assign each of the future days into one of the predetermined weather types based on proximate component scores. Since the synoptic weather types and their respective characteristics have already been predetermined, discriminant function analysis is an appropriate tool to assign each day of the future two time periods (2046–65 and 2081–2100) into one of the predetermined weather types (Klecka 1980; Lam and Cheng 1998).
b. Assessment of climate change impacts on rainfall: Method I
To support this assumption, the relationships between observed seasonal rainfall totals (or the number of seasonal total rain days) and seasonal frequency of 10 rainfall-related weather types were evaluated for the period April–November 1958–2002, across the selected river basins. For the 45 individual seasons, seasonal rainfall totals (or the number of seasonal total rain days) versus seasonal occurrence frequency of 10 rainfall-related weather types were plotted to demonstrate the relationships. As shown in Fig. 4, in the Thames River basin, seasonal rainfall totals and frequency of daily rainfall events greater than or equal to 0.2 mm significantly increase with an increase in the seasonal frequency of the 10 rainfall-related weather types. Similar results were found for other selected river basins but not shown owing to limitations of space. Across the four selected river basins, seasonal rainfall totals increase by a range of 1.8 mm (Rideau) to 4.3 mm (Humber) (four-river-basin nonweighted average: 3.4 mm) per one day increase in frequency of 10 rainfall-related weather types. The corresponding increase quantity for the number of seasonal rain days ranges from 0.6 (Rideau) to 1.0 (Grand) (four-river-basin average = 0.8).
Upward trends of the (a) number of seasonal rain days and (b) seasonal rainfall totals against seasonal occurrence frequency of 10 rainfall-related weather types for the period April–November 1958–2002 in Thames River basin. The solid line is a regression trend with slope b and p value.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
c. Assessment of climate change impacts on rainfall: Method II
In addition to changes in the frequency of future rainfall-related weather types considered in method I for projecting changes in future daily rainfall events, method II also considered changes in future climate characteristics. Method II applies within-weather-type daily rainfall simulation models (developed in the study, Cheng et al. 2010) with downscaled future climate data to project future daily rainfall quantities. To project future daily rainfall quantities, a two-step process of daily rainfall simulation modeling was employed. First, daily rainfall occurrence simulation models were employed to project probability of future daily rainfall occurrence. To more effectively apply daily rainfall occurrence simulation models, the 228 predictors used in development of the models, as described above, are needed for the future dataset. The future atmospheric stability indices were set up as the dummy variables according to the same criteria used for the historical observations (refer to Cheng et al. 2010 for details). In addition, the future standard meteorological variables were converted into the principal component variables for each day of the future dataset by multiplying the posteigenvector matrix (using the historical model developmental dataset) by the future downscaled climate data matrix. Similar to projection of future synoptic weather types, the posteigenvector was used, so that future component scores are comparable with the postscores since both used the same set of eigenvectors. Moreover, the future downscaled climate data were standardized using the mean and standard deviation of downscaled GCM historical runs (1961–2000) to more effectively compare component scores from both historical and future datasets as well as to remove the GCM bias. Second, using the probability of future daily rainfall occurrence, within-weather-type nonlinear daily rainfall quantity simulation models were used to project future daily rainfall amounts.
In each of the selected river basins, there are 11 daily rainfall quantity simulation models involved. In addition to 10 simulation models for 10 rainfall-related weather types described above, an extra simulation model was developed for the weather grouping other I. Other I consists of the weather types associated with some rainfall events, which do not meet the selection criteria for a “pure” rainfall-related weather types, The remainder of the weather types are grouped into other II, which is related to days with no rainfall. Consequently, all future days falling into the weather grouping other II were considered to have no rainfall.
5. Results and discussions
a. Future rainfall-related weather types
Using downscaled hourly climate information, discriminant function analysis is able to project future daily weather types for two time periods: 2046–65 and 2081–2100 as well as a GCM historical run period: 1961–2000. As discussed above, 10 weather types were identified over the period April–November 1958–2002 as the primary rainfall-related types. Percentage occurrences of the 10 rainfall-related weather types are plotted in Fig. 5, as shown by observations and downscaled 3-GCM A2 ensemble for three time periods (1961–2000, 2046–65, 2081–2100). Percentage occurrences of the weather types have a good agreement between observations and downscaled GCM historical runs over a comparative time period 1961–2000 across the study area. This implies that the downscaling method is suitable to derive station-scaled future climate information, and discriminant function analysis performs well in projecting future rainfall-related weather types.
Percentage occurrences of the 10 rainfall-related weather types as shown by observations (obs, 1961–2000) and downscaled 3-GCM A2 ensemble for three time periods (HR—historical run, 1961–2000; 2046–65; 2081–2100).
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
Within-weather-type difference of percentage occurrences between downscaled GCM historical runs and observations over a comparative time period (April–November 1961–2000) on average of three cities (London, Ottawa, and Toronto). Negative percentage occurrence difference: to increase future rainfall projections; positive percentage occurrence difference: to decrease future rainfall projections.
After considering such a difference, the three-site-averaged percentage occurrences of the future 10 rainfall-related weather types are projected to be 49%–53% across two future time periods (the current average of 46%). Specifically for individual weather types and GCM scenarios, the changes might be different. Table 5 shows percentage changes in the frequency of future rainfall-related weather types derived from downscaled CGCM A2 and CGCM B2 on average of three cities (London, Ottawa, and Toronto). It is immediately apparent that the warmer weather types occurred usually in summertime, such as cyclone I, cold front I, quasi-stationary front, and warm front I, are projected to increase in the future. The colder weather types that usually occur in spring and autumn, such as cyclone II, cold low, cold front II, and cold front III, are projected to decrease in the future. In addition, the change magnitudes of projected future rainfall-related weather type occurrences also vary between the scenarios. The increase rates for the warmer weather types (especially for cyclone I and quasi-stationary front), derived from downscaled B2 scenario for future two periods 2046–65 and 2081–2100, are usually projected to be smaller than those from downscaled A2 scenario. The corresponding projected decrease rates by the period 2081–2100 for the colder weather types (especially for cyclone II and cold low), derived from downscaled B2 scenarios, could also be smaller than those from A2 scenario. The difference in future change rates between A2 and B2 scenarios are expected since B2 scenario projects less future warming than A2 scenario, especially in the second half of this century.
Percentage changes in the frequency of future rainfall-related weather types derived from downscaled CGCM2 A2 and CGCM2 B2 from the current conditions of April–November 1961–2000 on average of three cities (London, Ottawa, and Toronto).
In addition to frequency of the weather types, the within-weather-type meteorological characteristics derived from downscaled future climate information were evaluated to compare with historical observations. As an example, hourly meteorological variables (temperature, dewpoint, sea level air pressure, u wind, and υ wind) for observations and downscaled future CGCM2 A2 climate information are displayed in quartile box plots for the 10 rainfall-related weather types in London (Fig. 6). Similar results were also found for other locations and GCM simulations but not shown owing to the limitations of space. From Fig. 6, it is immediately apparent that within-weather-type variances of the meteorological characteristics derived from downscaled future GCM climate information are very similar to the observations. In addition, the characteristics of warm weather types (e.g., cyclone I, mesohigh, and cold Front I) are projected to be warmer and moister. These results further implied that the downscaling method is suitable to derive station-scaled future climate information and the discriminant function analysis performed well in projecting future weather types using downscaled hourly climate information.
Data distribution comparison between hourly observations (the first quartile box plot, 1958–2002) and downscaled future hourly CGCM2 A2 simulations (the next two box plots: 2046–65 and 2081–2100) for each of 10 rainfall-related weather types in London.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
It is noteworthy that the projected increase in frequency of the warmer rainfall-related weather types might be not only due to the higher future temperatures but also considering other weather elements. From Fig. 6, it can be seen that within-weather-type data distributions for sea level air pressure and winds derived from downscaled future GCM climate data are similar to the observations. Furthermore, future hottest days were assigned by the synoptic weather typing to two other hottest weather types that were merged into the weather grouping other II with no rainfall. On average of downscaled 3-GCM A2 simulations, the frequencies of these hottest and second hottest weather types for the period 2081–2100 in Toronto are projected to be over 9.0 and 3.6 times greater than the current conditions (1.7 and 6.5 days yr−1 observed during 1961–2000), respectively. In addition, the projected decrease in frequency of the colder weather types might be partly due to the selected time period (April–November). As future temperature increases, as projected by the GCMs, the colder weather types that currently occur in April or November could, in the future, occur earlier, such as in March or occur later in December, respectively. However, in this study, March and December are not included in the analysis.
b. Robustness of rainfall simulation models
(a)–(d) Q–Q plots of daily rainfall quantities derived by simulation models using independent GCM historical runs vs observations over a comparative time period (April–November 1961–2000) in the four selected river basins. A 45° reference line suggests that both datasets come from populations with the same distribution.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
As described by Cheng et al. (2010), the robustness of rainfall simulation models was examined using historical observations of all days without synoptic weather typing. It was concluded that it is better to perform synoptic weather typing prior to the application of cumulative logit and nonlinear regressions for development of daily rainfall simulation models. In this paper, the robustness of rainfall simulation models is examined once again to ascertain this conclusion is still valid for projection of future daily rainfall quantities. To achieve this, a rainfall simulation model was redeveloped using all days without synoptic weather typing. The same methods and potential predictors used in the earlier rainfall simulation modeling (Cheng et al. 2010) were employed to develop rainfall simulation test model for each of the river basins. These rainfall simulation models were tested with downscaled GCM historical runs to project daily rainfall quantities for the period April–November 1961–2000. The resulted daily rainfall distribution is compared with observations, as shown in Q–Q plots (Fig. 8). It is seen that, from Fig. 8, the rainfall simulation models without weather typing usually underestimate quantities of daily heavy rainfall events (e.g., daily rainfall greater than 35 or 40 mm). This implies that it is necessary to perform synoptic weather typing altogether with daily rainfall simulation modeling for development of daily rainfall downscaling transfer functions.
As in Fig. 7, but using the different rainfall simulation models developed using all days without synoptic weather typing.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
c. Changes in frequency of future daily rainfall events and seasonal rainfall totals
Following the determination of weather types for future days based on downscaled GCM climate data, changes in the frequency of future daily rainfall events and seasonal rainfall totals could be quantitatively projected. In this study, to consider using different downscaling methods for local climate change impact analysis (Haylock et al. 2006), two independent methods described above were employed: method I—comparing frequencies of historical and future rainfall-related weather types—and method II—applying daily rainfall simulation models. The number of future seasonal rain days and future seasonal rainfall totals projected by two methods versus historical observations are graphically illustrated in Figs. 9 and 10. Method II usually projects greater increases in the frequency of future daily rainfall events and seasonal rainfall totals (April–November) than method I. In addition, the rainfall projections were evaluated by comparing differences in the number of seasonal rain days and seasonal rainfall totals derived from downscaled historical runs and observations during a comparative time period 1961–2000. As shown in Figs. 9 and 10, the values from both datasets are very similar for both method I and method II. This implies that both daily rainfall downscaling methods are suitable to be used for projecting changes in the number of future seasonal rain days and future seasonal rainfall totals.
Method I (based on changes in frequency of future rainfall-related weather types): the number of future seasonal rain days [(a) ≥0.2, (b) ≥15, and (c) ≥25 mm] and (d) future seasonal rainfall totals vs the observed values during the period April–November 1961–2002. The shaded bar is for observations, and the following three bars represent 3-GCMA2–averaged values for future three time periods; in order from left to right: 1961–2000, 2046–65, and 2081–2100.
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
As in Fig. 9, but using method II (based on daily rainfall simulation models).
Citation: Journal of Climate 24, 14; 10.1175/2011JCLI3764.1
To more clearly present changes in the frequency of future daily rainfall events and seasonal rainfall totals, four river basin average relative increases from the current conditions of April–November 1961–2002, derived from both methods, are shown in Table 6. The relative increases are presented by CGCM2 B2 and 3-CGM A2 averages with the range across GCMs. The projections of relative increases in the frequency of future daily rainfall events and seasonal rainfall totals vary across GCMs and across scenarios. Across three GCMs, lower and upper boundaries of the range derived from method I for the period 2046–65 shown in Table 6 are projections derived from CGCM2 A2 and ECHAM5 A2, respectively, and vise versa for the period 2081–2100. For method II, lower and upper boundaries of the range generally represent projections derived from CGCM2 A2 and GFDL A2, respectively, for both future time periods. Between scenarios A2 and B2, the relative increase rates for the period 2046–65, projected from CGCM2 A2, are slightly greater than those from CGCM2 B2, and vise versa for 2081–2100, with less than 10% difference. For method II, the relative increase rates for both future periods projected from CGCM2 B2 are greater than those from CGCM2 A2. For more detailed differences, regarding intermodel and interscenario uncertainties, refer to section 5e.
Four river basin–average percentage increases in the frequency of future seasonal rain days (≥0.2, ≥15, and ≥25 mm) and future seasonal rainfall totals from the current conditions of April–November 1961–2002, presented by 3-GCM A2 ensemble and CGCM2 B2.
The results with respect to increases in frequencies of daily rainfall events and seasonal rainfall totals presented in this study are consistent with the findings of previous studies (e.g., Zwiers and Kharin, 1998; Kharin and Zwiers, 2005; Haylock et al. 2006; Tebaldi et al. 2006; Meehl et al. 2007). There are three possible reasons for increases in the frequency of future rainfall events. First, a warmer and moister future climate is projected by the GCM scenarios. Second, the projected warmer future climate could bring about more frequent and vigorous atmospheric convection in the future. Third, the hydrological cycle under the projected future climate could be modified from the current condition.
d. Changes in future return-period values of extreme rainfall events
A return period analysis was employed to project changes in future return values of the extreme rainfall events for a number of return periods. A return period known as a recurrence interval is an estimate of the likelihood of events such as heavy rainfall of a certain intensity. Return values are thresholds that will be exceeded on average once every return period. The design criteria of stormwater infrastructure are constrained by the largest precipitation event anticipated during a fixed design period (e.g., 20, 50, or 100 years). To take into account climate change impacts on infrastructure, scientific information on changes in future return values of extreme rainfall events is needed for developing adaptation strategies and policies.
Annual maxima of the river basin–average three-day accumulated rainfall totals for the period 1961–2002 were fitted to the Gumbel (extreme value type I) distribution for each of the selected river basins. The use of three-day accumulated rainfall totals is arbitrary. Future projections of the river basin–average daily rainfall data, using two downscaled CGCM transient model simulations (Table 1) for three 50-yr periods (2001–50, 2026–75, and 2051–2100), were used to project future return-period values for this century. As shown in Table 7, the return values of three-day accumulated rainfall extremes for all evaluated return periods are projected to increase by about 20%–70% over the present century. For example, in the Humber River basin, the 20-yr return period values of three-day accumulated rainfall extremes for future three 50-yr periods are projected to increase by 50%, 63%, and 69%, respectively, from the observed value of 80 mm for the past 40 years. The percentage increase rates for longer return periods are usually greater than shorter return periods (except Thames River basin). Between two downscaled CGCM2 simulations, the projected percentage increases in the return values are similar, usually with slightly greater values derived from downscaled CGCM2 A2 than those from downscaled CGCM2 B2. The difference between the percentage increases derived from downscaled CGCM2 A2 and CGCM2 B2 on average of four river basins, as shown in Table 8, is usually less than 10% for future three 50-yr periods, with a few exemptions. To more effectively account for differences, the “difference” here represents the absolute difference between CGCM2 A2 and CGCM2 B2 to avoid negative values canceling out positive values. In addition, from Table 7, it can be seen that the 95% confidence interval for future projected return-period values is similar to the observed ones, which implies that the future projected return-period values are plausibly reliable.
Percentage increases in future three-day accumulated rainfall extremes for nine return periods (ensemble of two Canadian GCMs) from current observed values (95% confidence interval in parentheses). To effectively compare the 95% confidence intervals between observed and future projected return values as the same as the future projections, the 95% confidence intervals derived from observations are presented as percentages below or above the return values.
Differences of percentage increases in return values of future three-day accumulated rainfall extremes between downscaled CGCM2 A2 and CGCM2 B2 on average of the four river basins. The percentage increases in the return values derived from downscaled CGCM2 A2 are usually slightly greater than those from downscaled CGCM2 B2 with a few exceptions. To more effectively account for differences, the “difference” here represents the absolute difference between CGCM2 A2 and CGCM2 B2 to avoid negative values canceling out positive values.
e. Uncertainty of the study
Considerable effort was made in this study to transfer GCM-scale model simulations to station-scale climate information using statistical downscaling transfer functions. Through the downscaling process, the GCM bias was removed using about 50-yr historical relationships between regional-scale predictors and station-scale weather elements (Katz 2002). As a result, the quality of future GCM climate projections, following downscaling, was much improved. For example, the data distribution (including extreme events) of the downscaled GCM historical runs was similar to that of observations over a comparative time period 1961–2000. In addition, any differences between rainfall simulations driven by downscaled historical runs and observations were considered for further correction of future rainfall projections.
However, conclusions made in this study about the impacts of climate change on future rainfall still rely on GCM scenarios/projections and, as a result, there is corresponding uncertainty about the study findings. One of the most important sources of uncertainty in climate change impact studies comes from GCMs (Katz 2002). Because of the model resolution and complexity, the GCMs must have inevitably omitted some factors that affect climate; in turn, the GCMs are unable to resolve subgrid-scale processes and generate uncertainties through the model parameterizations. To quantitatively assess inter-GCM and interscenario uncertainties of future rainfall projections, we have analyzed the four river basin–average absolute difference between three selected GCMs under the SRES A2 scenario as well as absolute difference between two selected scenarios (CGCM2 A2 versus CGCM2 B2). The absolute difference used in analysis is to avoid negative values canceling out positive values. As shown in Table 9, overall, the intermodel uncertainties of percentage increases in the frequency of future seasonal rain days and seasonal rainfall totals are similar to the interscenario uncertainties, for both projection methods. Both intermodel and interscenario uncertainties from method II, overall, are about as twice as the uncertainties from method I with 11% versus 5%–6%. However, the uncertainties are generally much smaller than the projection of percentage increases in the frequency of future seasonal rain days and future seasonal rainfall totals as shown in Table 6. The overall mean projected percentage increases are about 2.6 times greater than overall mean intermodel and interscenario uncertainties from method I; the corresponding projected increases from method II are about 2.2–3.7 times greater than overall mean uncertainties.
Four river basin–average intermodel and interscenario uncertainties of percentage increases in the frequency of future seasonal rain days (≥0.2, ≥15, and ≥25 mm) and future seasonal rainfall totals from the current conditions of April–November 1961–2002.
Although the models developed from this study can simulate most of daily heavy rainfall events, as described above, it was found that the models have difficulty in capturing some summer localized convective rainfall events (Cheng et al. 2010). This model limitation is also reflected by simulation difficulty of summer localized convective cloud cover. As pointed out by Cheng et al. (2008), total cloud cover downscaling transfer functions performed better in the winter season than in the summer season since localized convective activities commonly occur in summer. Furthermore, this study has attempted to downscale the river basin–average daily rainfall information, which might underestimate the future extreme rainfall events at the individual stations. It is likely that projections in frequency of future heavy rainfall events offered by this study will represent the lower bound values for the study area. Therefore, southern Ontario could in the future possibly receive more heavy rainfall events than is currently projected by the study.
In addition to uncertainty of GCM scenarios and limitation of rainfall simulation models, the observed data used in the study have their limitations. Hourly meteorological data are essential to develop synoptic weather typing and rainfall simulation models. However, in the Grand River basin hourly meteorological observations are not available. For the Grand River, weather types classified using meteorological data gathered at London International Airport (located in Thames River basin) were used to link with the river basin’s daily rainfall data. Furthermore, predictors (e.g., atmospheric stability indices) derived from weather data observed at the London International Airport were used to simulate daily rainfall quantities for the Grand River basin. Therefore, the rainfall simulation results derived for the Grand River basin were not as accurate as they might be were hourly meteorological data for the Grand River basin available.
6. Conclusions and future work
The overall purpose of this study is to project possible changes in the frequency of daily rainfall events late this century for four selected river basins (i.e., Grand, Humber, Rideau, and Upper Thames) in Ontario, Canada. To achieve this goal, automated synoptic weather typing as well as cumulative logit and nonlinear regression analyses was applied together with downscaled GCM simulations to project future daily rainfall quantities. A formal verification process of model results has been built into the whole exercise, comprising synoptic weather typing, rainfall simulation modeling, and statistical downscaling. The results of the verification, based on historical observations of the outcome variables simulated by the models, showed good agreement. As a result, a general conclusion from this study is that a combination of synoptic weather typing, cumulative logit and nonlinear regression analyses, and regression-based downscaling can be useful to project changes in frequency of future daily rainfall events. The authors believe that such daily rainfall downscaling methods are useful to derive river basin–scale daily rainfall quantities for applications of hydrological and meteorological modeling. In addition, the modeled results from this study found that the frequency of future daily rainfall events could increase late this century due to the changing climate projected by GCM scenarios. The implication of the increases should be taken into consideration when adjusting engineering infrastructure requirements and developing adaptation strategies and policies.
As discussed earlier, the rainfall simulation models have demonstrated significant skill in the discrimination and prediction of the occurrence and quantity of daily rainfall events with a few exceptions of localized convective storms. Some of these convective storms might be impacted by the lake breeze effect. To more effectively capture these cases, future studies need to consider more analyses, using hourly rainfall data, for some stations surrounding the Great Lakes area, and rainfall simulation models could be rebuilt by taking into account part of the lake breeze effect. To more effectively capture and simulate current and hence future localized rainfall events, a dense network of hourly meteorological recording stations is essential to properly define or initialize the rainfall simulation models. The current Environment Canada’s observation network might not be fine enough to achieve this goal. In addition, future work could consider applying the methods used in this study to regional climate model (RCM) simulations for projection of future daily rainfall, which might overcome some of the limitations using GCM simulations.
Rainfall intensity–duration–frequency (IDF) curves, which consider short-duration (usually less than one day) rainfall intensity, are commonly used for hydrological engineering design standards. Since future daily rainfall totals are projected in this study, it is difficult to incorporate the 24-h value-based results into the shorter than 24-h-duration IDF curves. To more effectively include projections of future rainfall in the design standard, future short-duration rainfall data are necessary. Since six-hourly NCEP reanalysis data are available for the past 50 years, the methods used in this study to project future daily rainfall have potential to be adapted to project six-hourly rainfall quantities as a step toward even finer-scale temporal projections. The results of short-duration rainfall intensity analyses could be potentially used to develop more functional tools, integrating climate change with the building code design standard using the IDF curves.
Acknowledgments
This study was funded through the Government of Canada’s Climate Change Impacts and Adaptation Program (CCIAP), which made this research project (A901) possible. The authors gratefully acknowledge the suggestions made by the Project Advisory Committee, which greatly improved the study. We also would like to thank two anonymous reviewers for providing detailed comments that significantly improved the original manuscript.
REFERENCES
Abaurrea, J., and J. Asín, 2005: Forecasting local daily precipitation patterns in a climate change scenario. Climate Res., 28, 183–197.
Alley, R. B., and Coauthors, 2007: Summary for policymakers. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 1–18.
Buishand, T. A., M. V. Shabalova, and T. Brandsma, 2004: On the choice of the temporal aggregation level for statistical downscaling of precipitation. J. Climate, 17, 1816–1827.
Cheng, S. C., G. Li, Q. Li, and H. Auld, 2008: Statistical downscaling of hourly and daily climate scenarios for various meteorological variables in south-central Canada. Theor. Appl. Climatol., 91, 129–147, doi:10.1007/s00704-007-0302-8.
Cheng, S. C., G. Li, Q. Li, and H. Auld, 2010: A synoptic weather typing approach to simulate daily rainfall and extremes in Ontario, Canada: Potential for climate change projections. J. Appl. Meteor. Climatol., 49, 845–866.
Cubasch, U., J. Waszkewitz, G. Hegerl, and J. Perlwitz, 1995: Regional climate changes as simulated time-slice experiments. Climatic Change, 31, 273–304.
Emori, S., and S. J. Brown, 2005: Dynamic and thermodynamic changes in mean and extreme precipitation under changed climate. Geophys. Res. Lett., 32, L17706, doi:10.1029/2005GL023272.
Environment Canada, cited 2006: Data. [Available online at http://www.cccma.bc.ec.gc.ca/data/data.shtml.]
Fealy, R., and J. Sweeney, 2007: Statistical downscaling of precipitation for a selection of sites in Ireland employing a generalised linear modelling approach. Int. J. Climatol., 27, 2083–2094.
Fowler, H. J., S. Blenkinsop, and C. Tebaldi, 2007: Linking climate change modeling to impacts studies: Recent advances in downscaling techniques for hydrological modeling. Int. J. Climatol., 27, 1547–1578.
Galway, J. G., 1956: The lifted index as a predictor of latent instability. Bull. Amer. Meteor. Soc., 37, 528–529.
George, J. J., 1960: Weather Forecasting for Aeronautics. Academic Press, 673 pp.
Goodess, C. M., and J. P. Palutikof, 1998: Development of daily rainfall scenarios for southeast Spain using a circulation-type approach to downscaling. Int. J. Climatol., 18, 1051–1083.
Haylock, M. R., G. C. Cawley, C. Harpham, R. L. Wilby, and C. M. Goodess, 2006: Downscaling heavy precipitation over the United Kingdom: A comparison of dynamical and statistical methods and their future scenarios. Int. J. Climatol., 26, 1397–1415.
Kalkstein, L. S., 1979: A synoptic climatological approach for environmental analysis. Proc. Middle States Div. Assoc. Amer. Geogr., 13, 68–75.
Katz, R. W., 2002: Techniques for estimating uncertainty in climate change scenarios and impact studies. Climate Res., 20, 167–185.
Kharin, V. V., and F. W. Zwiers, 2005: Estimating extremes in transient climate change simulations. J. Climate, 18, 1156–1173.
Klecka, W. R., 1980: Discriminant Analysis. Sage University Press, 71 pp.
Kostopoulou, E., and P. D. Jones, 2007a: Comprehensive analysis of the climate variability in the eastern Mediterranean. Part I: Map-pattern classification. Int. J. Climatol., 27, 1189–1214.
Kostopoulou, E., and P. D. Jones, 2007b: Comprehensive analysis of the climate variability in the eastern Mediterranean. Part II: Relationships between atmospheric circulation patterns and surface climatic elements. Int. J. Climatol., 27, 1351–1371.
Lam, K. C., and C. S. Cheng, 1998: A synoptic climatological approach to forecast concentrations of sulfur dioxide and nitrogen oxides in Hong Kong. Environ. Pollut., 101, 183–191.
Meehl, G. A., C. Covey, T. Delworth, M. Latif, B. McAvaney, J. F. B. Mitchell, R. J. Stouffer, and K. E. Taylor, 2007: The WCRP CMIP3 multimodel dataset: A new era in climate change research. Bull. Amer. Meteor. Soc., 88, 1383–1394.
Miller, R. C., 1972: Notes on analysis and severe storm forecasting procedures of the Air Force Global Weather Central. U.S. Air Force Air Weather Service Tech. Rep. 200(R), 190 pp.
Nguyen, V. T. V., T. D. Nguyen, and P. Gachon, 2006: On the linkage of large-scale climate variability with local characteristics of daily precipitation and temperature extremes: An evaluation of statistical downscaling methods. Advances in Geosciences, Vol. 4, Academic Press, 1–9.
Perry, A., 1983: Growth points in synoptic climatology. Prog. Phys. Geogr., 7, 90–96.
Program for Climate Model Diagnosis and Intercomparison, cited 2006: WCRP CMIP3 multi-model dataset. [Available online at http://www-pcmdi.llnl.gov/ipcc/about_ipcc.php.]
Tebaldi, C., K. Hayhoe, J. M. Arblaster, and G. A. Meehl, 2006: Going to the extremes: An intercomparison of model-simulated historical and future changes in extreme events. Climatic Change, 79, 185–211.
Wilby, R. L., and T. M. L. Wigley, 1997: Downscaling general circulation model output: A review of methods and limitations. Prog. Phys. Geogr., 21, 530–548.
Wilby, R. L., C. W. Dawson, and E. M. Barrow, 2002: SDSM — A decision support tool for the assessment of regional climate change impacts. Environ. Model. Software, 17, 147–159.
Zwiers, F. W., and V. V. Kharin, 1998: Changes in the extremes of the climate simulated by CCC GCM2 under CO2 doubling. J. Climate, 11, 2200–2222.
