The Effects of Climate Change on Extreme Precipitation Events in the Upper Thames River Basin: A Comparison of Downscaling Approaches

Abstract

Future changes in climatic conditions from increasing greenhouse gas concentrations will have a major impact on the hydrologic cycle. It is important to understand and predict future changes in temperature and precipitation in order to effectively manage water resources. Atmosphere-Ocean coupled Global Climate Models (AOGCMs) are used widely to predict the effects of greenhouse-gas forcing on global climate conditions. However, their spatial and temporal resolutions are quite large so their outputs must be modified to represent local climate conditions. This process is called downscaling, and there are a variety of tools available to achieve this goal. This study compares three downscaling approaches, namely the Statistical DownScaling Model (SDSM), Long Ashton Research Station Weather Generator (LARS-WG), and the K-NN Weather Generator with Principal Component Analysis (WG-PCA). Each weather generator is used to simulate the historical climate for the Upper Thames River Basin in Ontario, Canada for use in a comparison of downscaling tools. Future climate conditions are simulated by LARS-WG and WG-PCA from six different AOGCMs, each with two to three emissions scenarios, for a total of 15 different models. In simulation of historical climate variability, the models generally perform better in terms of mean daily precipitation and total monthly precipitation. LARS-WG simulates precipitation events well but cannot reproduce means and variances in the daily temperature series. SDSM adequately simulates both temperatures and precipitation events. WG-PCA reproduces daily temperatures very well but overestimates the occurrence of some extreme precipitation events. Results are variable for the downscaling of AOGCMs; however, the downscaling tools generally predict a rise in winter, spring and fall precipitation totals, as well as an overall increase in mean annual precipitation in future decades.

Les changements futurs lis aux conditions climatiques qui dcouleront des concentrations croissantes de gaz effet de serre auront une incidence majeure sur le cycle hydrologique. Il est important de comprendre et de prdire les changements futurs lis aux tempratures et aux prcipitations afin de pouvoir grer efficacement les ressources en eau. Les modles de circulation gnrale coupls atmosphre-ocan (AOGCM) sont largement utiliss pour prdire les effets des gaz effet de serre sur les conditions climatiques mondiales. Cependant, leurs rsolutions spatiales et temporelles sont trs grandes. Leurs rsultats doivent donc tre modifis afin de reprsenter les conditions climatiques locales. Ce processus a pour nom rduction dchelle et il existe divers outils permettant d'atteindre cet objectif. La prsente tude compare trois mthodes de rduction dchelle, c'est--dire le modle de rduction dchelle statistique (SDSM), le gnrateur stochastique de climat LARS-WG et le gnrateur de climat KNN avec analyse en composantes principales (GC-ACP). Chaque gnrateur de climat sert simuler le climat historique pour le bassin hydrographique du cours suprieur de la rivire Thames en Ontario, au Canada, des fins de comparaison des outils de rduction dchelle. Les conditions climatiques futures sont simules l'aide du gnrateur LARS-WG et du GC-ACP partir de six AOGCM diffrents, chacun s'accompagnant de deux trois scnarios dmissions, pour un total de 15 modles diffrents.Pour la simulation de la variabilit climatique historique, les modles offrent en gnral un meilleur rendement en ce qui concerne les prcipitations quotidiennes moyennes et les prcipitations mensuelles totales. Le LARS-WG simule bien les vnements de prcipitations mais ne peut reproduire les moyennes et les carts dans les sries de tempratures quotidiennes. Le modle SDSM simule convenablement la fois les tempratures et les prcipitations. Le GC-ACP reproduit trs bien les tempratures quotidiennes mais surestime la survenance de certains vnements de prcipitations extrmes. Les rsultats sont variables pour la rduction dchelle des modles AOGCM; cependant, les outils de rduction dchelle prdisent en gnral une hausse des prcipitations totales d'hiver, de printemps et d'automne, ainsi qu'une hausse globale des prcipitations annuelles moyennes dans les dcennies venir.

Introduction

Industrialization and a growing dependency on fossil fuels have caused an increase in atmospheric greenhouse gas concentrations over the past century. A predicted consequence of this increase is that the average global temperature will rise by 1.8 to 4.0C by the year 2100 (Intergovernmental Panel on Climate Change (IPCC), 2007). The most severe temperature increase is predicted in the mid to high latitude regions of the Northern Hemisphere (IPCC, 2007). Rising temperature has a major impact on the amount and frequency of precipitation an area receives. This is particularly important for watersheds where runoff from extreme precipitation events causes rising stream flows (Zhang et al., 2008; Kwon et al., 2011).

The IPCC has predicted that runoff in high latitudes and wet tropical regions could increase by as much as 1040% by the 2080s (IPCC, 2007). As such, the assumption that historical rainfall and stream flow records can predict future events is no longer valid. It is crucial to perform hydrologic impact assessments of climate change in an attempt to better understand the frequency and magnitude of future extreme events. Such assessments can be used by engineers and policy-makers to ensure effective water resource management and adequate flooding protection.

Atmosphere-Ocean coupled Global Climate Models (AOGCMs) are widely-used tools that predict climatic responses to greenhouse gas forcing at a global scale. Most models divide the atmosphere into a grid with resolutions typically greater than 2 latitude and longitude, with horizontal spacing on the order of 100 km and up to 20 vertical layers (Dibike and Coulibaly, 2005; Prodanovic and Simonovic, 2007; Liu et al., 2011; Jeong et al., 2012).

While daily outputs are available for some AOGCMs, these are often of inferior quality and are unavailable for many of the models, so monthly values are often used in climatic vulnerability assessments (Prodanovic and Simonovic, 2007; Kwon et al., 2011). Because of their relatively large spatial and temporal resolutions, the models are inappropriate for direct application in hydrologic impact assessments at a watershed scale (Liu et al., 2011; Ying et al., 2011; Schoof, 2012). Furthermore, the underlying assumptions and inputs vary from model to model, so inclusion of several AOGCMs is favourable for a comprehensive climate change impact assessment (Schoof, 2012).

One way of converting globally scaled AOGCM outputs to a watershed scale is through the use of downscaling techniques. A wide variety of methods have been employed to downscale AOGCM data. Two major groups of statistical downscaling tools are: 1) regression based (transfer function) methods and 2) stochastic weather generators (Dibike and Coulibaly, 2005; Ying et al., 2011).

Regression-based techniques attempt to quantify relationships between local predictands (temperature, precipitation) and larger-scale atmospheric variables like wind speed, humidity, and pressure (Wilby et al., 2004; Jeong et al., 2012). Stochastic weather generators can be classified broadly in three main categories: parametric, semi-parametric, and non-parametric. An advantage of the weather generators is that they can be used to generate synthetic time series of any length, and thus the frequency and probability of extreme events can be examined.

Parametric weather generators typically use a Markov chain to determine the probability of wet or dry days, and then use probability distributions to determine precipitation amounts, temperatures, and other secondary variables (Wilby et al., 2004). Many of the models are extensions of the original WGEN approach, proposed by Richardson (1981). Examples include CLIGEN, WXGEN, GEM, and SIMMETEO (Hanson and Johnson, 1998; Soltani and Hoogenboom, 2003; Kuchar, 2004; Schoof et al., 2005).

While these parametric approaches have the advantage of being computationally simple, there are several drawbacks. When used with a low-order Markov dependence, some models cannot adequately simulate wet and dry spell lengths, and often underestimate prolonged drought or rainfall events (Semenov and Barrow, 1997; Dibike and Coulibaly, 2005; Sharif and Burn, 2007; Mehrotra and Sharma, 2007). Moreover, an underlying probability distribution must be assumed for precipitation amounts, temperatures and other variables, and the models are sensitive to the choice of that distribution (Sharif and Burn, 2007).

In order to address the issues associated with parametric weather generator algorithms, semi-parametric models have been developed. Two common semi-parametric models are the Long Ashton Research Station Weather Generator (LARS-WG) and the Statistical DownScaling Model (SDSM), both of which are readily accessible for use by the broader climate-change science community (Canadian Climate Change Scenarios Network (CCCSN), 2011).

SDSM is a regression-based downscaling model combined with a stochastic weather generator (Wilby and Dawson, 2007). An advantage of the approach is that AOGCM outputs are not used in the simulation of precipitation and temperature, which have very fine spatial and temporal variability and thus are not accurately reproduced by the climate models. Instead, AOGCM outputs of synoptic-scale variables, such as upper-level wind speed and air pressure, are used to linearly condition local-scale variables.

SDSM has been applied to several basins in the UK and North America, and has been shown to adequately reproduce historical climate (Dibike and Coulibaly, 2005; Wetterhall et al., 2007; Wilby and Dawson, 2007; Koukidis and Berg, 2009; Chen et al., 2010). Most studies use large-scale predictor data from the National Centre for Environmental Prediction (NCEP), screening up to 26 variables as potential predictors of precipitation or temperature (Dibike and Coulibaly, 2005; Chen et al., 2010; Hashmi et al., 2011; Liu et al., 2011). Some applications found that the variability in wet and dry spell lengths is underestimated by the model (Dibike and Coulibaly, 2005; Wetterhall et al., 2007; Liu et al., 2012). Other studies found that SDSM underestimates the standard deviation in daily precipitation amounts for some months, as well as the occurrence of extreme precipitation events (Chen et al., 2010; Hashmi et al., 2011; Liu et al., 2012). A major limitation of the SDSM approach is that each variable is simulated independently. As such, the model cannot preserve inter-variable correlations, and spatial correlations must be assumed for multi-site application. Furthermore the choice of predictor combinations and domain spaces can be critical to the results of downscaling (Liu et al., 2011).

LARS-WG is another popular and user-friendly downscaling tool developed for agricultural impact assessments (Racsko et al., 1991; Semenov and Porter, 1995; Semenov and Barrow, 1997). Downscaling is achieved by applying AOGCM change-factors to the observed weather series to develop AOGCM-modified daily inputs to the weather generator. LARS-WG employs a series approach in which the distribution of wet and dry spell sequences in the observed data are used to derive a spell distribution for the simulated data. Precipitation amounts are then modelled based on a mixed-exponential distribution. Distributions for the other variables are conditional on whether the current day is wet or dry (Semenov and Barrow, 1997). In this way the observed climate characteristics are preserved by the simulated data (Dibike and Coulibaly, 2005; Khan et al., 2006).

LARS-WG has been shown to adequately reproduce daily precipitation and temperature variability in Canada (Khan et al., 2006), New Zealand (Hashmi et al., 2011) and Europe (Semenov and Barrow, 1997). Hashmi et al. (2011) found that LARS-WG tends to underestimate the standard deviation of daily precipitation amounts, as well as the return periods of extreme precipitation events. A drawback associated with this downscaling tool is that spatial correlations must be assumed for multi-site application.

Non-parametric weather generators are another popular method for downscaling AOGCM data. They differ from the parametric and semi-parametric models discussed above in that no statistical assumptions are made about the probability distributions of the climate variables. Non-parametric weather generators are computationally simple, using a K-Nearest Neighbour (K-NN) resampling procedure that selects the following day's weather from a subset of the historical record with similar characteristics to the current day (Beersma et al., 2001; Wojcik and Buishand, 2003; Sharif and Burn, 2006). Downscaling requires the use of AOGCM change factors to modify historical observed datasets. Multi-site application is simple as simulations are run concurrently and spatial correlations are preserved. This is particularly advantageous for hydrologic impact assessments.

Various adaptations of the K-NN downscaling model have been found to successfully simulate observed climate variability for sites in North America (Yates et al., 2003; Sharif and Burn, 2006; Eum and Simonovic, 2012), South America (Apipattanavis et al., 2007), Asia (Eum et al., 2010) and Europe (Brandsma and Buishand, 1998; Beersma et al., 2001; Wojcik and Buishand, 2003). One drawback of the traditional K-NN approach is that historical values are merely re-shuffled; therefore, no unprecedented daily precipitation amounts are produced in simulation (Yates et al., 2003; Wojcik and Buishand, 2003). To address this issue, Sharif and Burn (2007) modified the K-NN algorithm to include a perturbation component. It was also modified by Eum and Simonovic (2008) to include more variables in the selection of nearest neighbours through the use of principal component analysis (and renamed the Weather Generator with Principal Component Analysis, WG-PCA).

This study provides a comparison and evaluation of three widely-used downscaling methods: SDSM, LARS-WG and WG-PCA, in terms of their ability to simulate extreme precipitation events at the Upper Thames River Basin (UTRB) in Ontario, Canada. Downscaling of daily precipitation and temperature is performed for the 2050's climate (20412070) using monthly data from six different AOGCMs, each with two to three emissions scenarios. Due to the inherent limitations in each AOGCM and the various downscaling approaches, it is important to consider a wide variety of models and downscaling tools for a comprehensive climate change impact assessment. The rest of the paper is organized as follows: first, descriptions of the study area and datasets used are given. Following are details of each downscaling application as well as methods of analysis for their comparison. Next is a discussion of the results, followed by conclusions of the study.

Study Area and Data

The Upper Thames River Basin

The Upper Thames River Basin (UTRB) is located in southwestern Ontario, Canada, between the Great Lakes of Erie and Huron. The river drains an area of 3,482 km², running 273 km in two main branches that meet in London, the major urban centre, and continue as a single channel through the Lower Thames River into Lake St. Clair. Figure 1 shows a map of the basin.

Figure 1. The Upper Thames River basin, in southwestern Ontario, Canada.

The total population of the basin is 460,000 with 350,000 of that population residing in London (Simonovic, 2010). The river is located in a highly developed area of southwestern Ontario where land use is predominately agricultural (80%) and urban, with some remaining forested areas. Average yearly precipitation in the basin is about 1,000 mm/year, with roughly 60% lost through pond and wetland storage, groundwater recharge, evaporation, and evapotranspiration (Wilcox et al., 1998).

The UTRB has a long history of seasonal flooding events. These most frequently occur in March due to snowmelt, and in the summer months due to intensive flood-producing rainfall events. Low-flow periods typically occur between June and September with the highest drought risk in July and August (Simonovic, 2010).

Flooding infrastructure in the UTRB includes three major reservoirs: Wildwood and Fanshawe on the North Thames, and Pittock on the South Thames. The primary purposes of these are to control flooding in London during spring snowmelt and summer storms, as well as for low flow augmentation and recreational purposes. There are also extensive diking systems throughout the City of London as well as diversion channels in order to reduce flood damages.

Recent studies have suggested that climate change has the potential to increase flooding risk in the basin (Sharif and Burn, 2006; Prodanovic and Simonovic, 2007; Solaiman et al., 2010; Simonovic, 2010; Eum and Simonovic, 2012). Higher flood levels could cause major social and economic impacts in the basin, particularly for London, as existing infrastructure becomes damaged and families are forced to relocate from flooded areas. As such, it is important to have an accurate assessment of potential climate impacts on the basin in order to manage the risks associated with increasing flood levels.

Data

Input data used in this study were collected from several sources. Table 1 provides a description of the climate variables used in this study. Historical precipitation and temperature data for 19792005 were collected from Environment Canada's Canadian Daily Climate Data (CDCD) archive for the London Airport station (Environment Canada, 2011).

Table 1.  Environment Canada and NARR climate variables used in the study.

Daily Variable	Description
PPT^a	Precipitation (mm)
TMAX^a	Maximum daily temperature (C)
TMIN^a	Minimum daily temperature (C)
PRMSL^b	Pressure reduced to mean sea level (Pa)
SPFH^b	Specific humidity (kg/kg)
UGRD^b	Eastward wind speed (m/s)
VGRD^b	Northward wind speed (m/s)
^aData collected from Environment Canada.
^bData collected from NARR.

In order to provide predictor variables for SDSM, North American Regional Reanalysis (NARR) data were collected from the Data Access Integration website of the Canadian Climate Change Scenarios Network (CCCSN, 2012). The NARR variables employed in this study are specific humidity, mean sea level pressure, eastward wind component and northward wind component.

Table 2 provides details of the AOGCMs selected for this study, including the country of origin and the atmospheric resolution of each model. AOGCMs were chosen based on the availability of variables to provide consistency with the NARR and Environment Canada data. Each model includes two to three of the emissions scenarios from IPCC's Special Report on Emissions Scenarios (SRES) storylines A1B, A2 and B1 (IPCC, 2000). The variables include maximum and minimum temperature, precipitation, northward wind component, eastward wind component, mean sea level pressure, and specific humidity.

Table 2.  Description of AOGCM and emissions scenarios used in the study.

			Atmospheric Resolution
AOGCM	Sponsors, Country	SRES Emissions Scenarios	Lat.	Long.
CGCM3T47, 2005	Canadian Centre for Climate Modeling and Analysis, Canada	A1B, A2, B1	3.75	3.75
CGCM3T63, 2005		A1B, A2, B1	2.81	2.81
CSIROMK3.5, 2001	Commonwealth Scientific and Industrial Research Organization (CISRO) Atmospheric Research, Australia	A2, B1	1.875	1.875
GISSAOM, 2004	National Aeronautics and Space Administration (NASA)/ Goddard Institute for Space Studies (GISS), USA	A1B, B1	3	4
MIROC3.2HIRES, 2004	Centre for Climate System Research (University of Tokyo), National Institute for Environmental Studies, and Frontier Research Centre for Global Change (JAMSTEC), Japan	A1B, B1	1.125	1.125
MIROC3.2MEDRES, 2004		A1B, A2, B1	2.8	2.8

Monthly AOGCM data are available from the CCCSN GCM/RCM monthly data download interface (CCCSN, 2011). Data from each AOGCM for the time period 20412070, representative of a 2050's climate with forcing from each SRES emissions scenario, were collected. AOGCM data representative of the baseline climate (19712005) were also gathered in order to compute monthly change factors.

Methodology

The Canadian Climate Change Scenarios Network website (CCCSN, 2012) provides public access to AOGCM datasets as well as two popular downscaling tools, SDSM and LARS-WG. There are extensive resources on the website to aid users in the application of these downscaling approaches.

In order to use monthly AOGCM data for local-scale impact studies on daily weather, pre-processing of the AOGCM outputs is required. This is carried out in two steps. Firstly, the gridded AOGCM values are interpolated to the station of interest using the Inverse Distance Weighting Method. NARR gridded variables are also interpolated to the station of interest in order to provide single-station input data for LARS-WG and WG-PCA.

Next, change factors between the baseline (19712005) and future (20412070) climate from each AOGCM are calculated using the monthly averages. For temperatures, wind speeds and mean sea level pressure, the change factors are computed as the magnitude of difference between the baseline and future monthly averages. For precipitation and specific humidity, percent changes are used.

The monthly change factors from each AOGCM and emissions scenario are applied to the daily historical datasets from Environment Canada and NARR to produce AOGCM-modified datasets for input into LARS-WG and WG-PCA. For consistency between the two downscaling tools, change factors are applied in the same way to the daily data to generate identical inputs to both downscaling tools.

The variables simulated by the weather generators in this study are daily precipitation, maximum temperature, and minimum temperature. These are simulated for the London Airport weather station, with a historical record of 27 years. In order to investigate the statistical characteristics of the output, twelve ensembles are generated for a total of 324 years of synthetic climate data for each simulation.

Validation of the downscaling models is performed using the first half (14 years) of the historical record as an input, and comparing the simulated output to the second half (13 years) of the historical data. Monthly box plots of precipitation, wet days, and average temperatures are generated as well as wet-spell length frequency distributions to determine which calibration best matches the historical data.

Application of the Statistical Downscaling Model (SDSM)

SDSM is a widely-used downscaling tool which is a hybrid between a stochastic weather generator and a regression-based method, where atmospheric variables (predictors) such as humidity and wind speeds from reanalysis datasets are used to linearly condition local-scale variables (predictands) such as precipitation amounts (Wilby and Dawson, 2007; Koukidis and Berg, 2009). Precipitation is simulated using a conditional process where local amounts are correlated with the occurrence of wet days, which are correlated with the atmospheric predictor variables (Khan et al., 2006). Temperatures are modelled as unconditional processes, which assume a direct relationship between the predictand and the atmospheric predictors.

There are several tools within the program that can be used to select predictors. Predictor variable data should be standardized prior to screening the various combinations. Correlation and partial correlation analysis are used to determine potential combinations of predictor variables. Knowledge of the physical relationships between weather variables is important to ensure a reasonable set of predictors is used (Dibike and Coulibaly, 2005; Liu et al., 2011). Furthermore, the selected predictors should not be highly correlated with one another as this can result in over-fitting (Wilby and Dawson, 2007; Liu et al., 2011).

There are a variety of transfer functions available to improve correlations between the predictand and predictors. For daily precipitation, which follows a skewed distribution, the fourth root transformation is applied in order to normalize the distribution (Wilby and Dawson, 2007). Temperatures are normally distributed, so no transformations are required. The model can also be classified as monthly or annual depending on the variable being simulated. For temperatures, a monthly model is used, with one regression equation for each month. Precipitation is also simulated as a monthly model based on calibration results and the recommendations of similar studies (Khan et al., 2006).

Relevant combinations of predictors and transfer functions were tested for each predictand, using different values for bias correction and variance inflation to best replicate the observed data. Predictor combinations were chosen based on the ability of the model to adequately simulate climatic variability using an independent dataset for validation. The selected combinations of predictors for each predictand are shown in Table 3 .

Table 3.  Summary of selected large-scale predictor variables corresponding to selected predictands for SDSM. See Table 1 for definitions of predictor variables.

Predictand	Precipitation	Max Temperature	Min Temperature
Selected predictors	VGRD	VGRD	VGRD
	PRMSL	SPFH	SPFH

A major limitation of the SDSM simulation presented here is that interpolated single-point predictor data are employed to ensure consistency with the other downscaling tools. For best results, gridded NARR predictor data should be used and the optimal grid spacing determined, as the choice of domain size (grid box area) can significantly affect performance of SDSM (Tomozeiu et al., 2007; Liu et al., 2011). Furthermore screening of more predictor variables from the reanalysis data might improve results, as most researchers use up to 26 potential predictors (Khan et al., 2006; Chen et al., 2010; Liu et al., 2011; Hashmi et al., 2011).

The stochastic component of SDSM allows for the simulation of multiple ensembles of climate data, where each ensemble has the same length as the input dataset. For downscaling of climate scenarios with SDSM, daily AOGCM gridded data are used in place of the reanalysis predictors. Because only monthly AOGCM data are available for the models used in this study, synthetic historical data are generated for a performance validation.

Application of Long Ashton Research Station Weather Generator (LARS-WG)

LARS-WG was developed by Dr. Mikhail Semenov in the UK as a tool for agricultural impact assessments (Racsko et al., 1991; Semenov and Porter, 1994; Semenov and Barrow, 1997). The model uses time series of precipitation, maximum and minimum temperatures, and solar radiation as inputs. LARS-WG analyzes the observed precipitation series in order to determine the statistics of wet-day occurrence and mean daily precipitation. From this, semi-empirical distributions are developed to simulate wet and dry-spell lengths with daily precipitation amounts conditional on the spell length (Semenov and Barrow, 2002; Khan et al., 2006; Hashmi et al., 2011).

For temperature simulation, LARS-WG uses a conditional process, whereby temperatures are derived based on the day's wet or dry status. The historical temperature records are analyzed to determine the mean and standard deviation for monthly wet and dry days. Annual variation in temperature standard deviation is approximated using Fourier series, and the normal distribution is used for residuals (Semenov and Barrow, 2002; Khan et al., 2006). LARS-WG uses the statistical parameters derived from the historical input record to simulate ensembles of synthetic climate data.

Future scenarios are generated using modified temperature and precipitation input data, where change factors from the monthly AOGCM data are applied to the observed record. This differs from the SDSM approach in which large-scale gridded daily AOGCM data are used as predictor variables. For this study, LARS-WG is used to generate synthetic historical climate data as well as data for each AOGCM and emissions scenario to provide a comparison between the downscaling approaches.

Application of the K-NN Weather Generator with Principal Component Analysis (WG-PCA)

The WG-PCA downscaling model is an extension of the K-NN approach proposed by Yates et al. (2003) and Sharif and Burn (2006). The model can be used to simulate multiple variables for any number of stations in a watershed.

The approach works by creating a subset of data from the historical record that lie within a temporal window centered on the current day. The temporal window can be chosen by the user, and in regions with highly seasonal climates, such as southwestern Ontario, it is important to select a window that is not too large; otherwise, unreasonable values could be produced. A value of 14 days is selected for the temporal window length at the UTRB, based on calibration results.

Next, the average weather condition across stations is computed for the current day and all days in the temporal window. Principal components analysis is used to determine which days in the subset have the most similar characteristics to the current day (Eum et al., 2010). The closest K days are retained and a geometric probability distribution is used with a random variable to select the next day's weather from this subset.

WG-PCA can be used to generate ensembles of climate data, each with the same length as the historical input record. AOGCM scenarios are created by applying change factors to the historical record and using the modified dataset as an input. Because no underlying probability distributions are assumed by the model, the only calibration required is the selection of a temporal window for the site. All seven variables found in Table 1 are used as inputs to the model to provide more information in the selection of the nearest neighbour. The weather generator is used to generate synthetic historical climate data as well as several AOGCM outputs.

Evaluation of Downscaling Approaches

In order to evaluate the utility of the three downscaling approaches for application to the Upper Thames River Basin, the ability of each model to simulate historical precipitation and temperature variability is investigated. Twelve ensembles are simulated using the 27-year input data for a total of 324 years of synthetic climate for each downscaling tool. Line plots of monthly means and standard deviations of daily precipitation as well as boxplots of total monthly precipitation are compared with the historical values ( Figure 2 ). Boxplots contain information about the spread of the simulated data, as the interquartile range (IQR) is shown by the height of the box with the median shown as a thick black line within the box. The whiskers extend to 1.5 times the IQR and outliers are shown as black dots. The ability of each downscaling approach to simulate extreme precipitation and temperature events is also investigated ( Figures 3 and 4 ). Table 4 shows the temperature and precipitation indices selected to compare the downscaling tools. Precipitation and temperature thresholds are selected based on those commonly used in the literature (Brown et al., 2010; dos Santos et al., 2011; Vincent et al., 2011; Hu et al., 2012). Boxplots of the simulated indices are used to compare the outputs with the median of the observed data.

Figure 2. Comparison of generated and observed climate variability from three weather generators. The first column shows results from SDSM, the second column shows LARS-WG and the third column shows WG-PCA. The observed values are represented with a solid line and the simulated values are represented with a dashed line. The top row of graphs show the absolute maximum TMAX (top lines) and absolute minimum TMIN (bottom lines). The second row contains boxplots of simulated total monthly precipitation from the weather generators, with historical medians plotted in black. The third row contains simulated and observed standard deviations and means (heavy lines) of daily precipitation.

Figure 3. Boxplots of simulated temperature indices from SDSM, LARS-WG and WG-PCA with the observed historical median and quartiles shown.

Figure 4. Boxplots of simulated precipitation indices from SDSM, LARS-WG and WG-PCA with the observed historical median and quartiles shown.

Table 4.  Definitions of selected temperature and precipitation indices.

	Definition	Units
Precipitation
Heavy PPT days	Number of days with PPT greater than 10 mm	days/yr
Wet days	Number of days with PPT greater than 0.5 mm	days/yr
Very wet days	Number of days with PPT greater than the 95^th percentile of observed wet days	days/yr
Highest 5-day PPT	Maximum PPT sum for a 5-day interval	mm/5 days
Temperature
Warm days	Number of days with TMAX greater than the 90^th percentile of observed data	days/yr
Warm nights	Number of days with TMIN greater than the 90^th percentile of observed data	days/yr
Cold days	Number of days with TMAX less than the 10^th percentile of observed data	days/yr
Cold nights	Number of days with TMIN less than the 10^th percentile of observed data	days/yr

Results

Preservation of Historical Climate Variability

A comparison of simulated and historical climate variability is provided in Figure 2 . Table 5 presents coefficients of determination (R-squared values) for the results shown in Figure 2 . The first row shows the absolute maximum TMAX and minimum TMIN, with the generated values plotted as dashed lines. SDSM produces values of up to 8C above and below the historical range, but captures the seasonality of the historical record well. SDSM has the lowest R-squared value (0.904) for maximum temperatures in Table 5 , and ranks second of the models for simulation of absolute minimum temperature. Model performance of SDSM for temperature simulation could likely be improved if gridded NARR data were used and more potential predictors tested (Khan et al., 2006; Chen et al., 2010; Liu et al., 2011; Hashmi et al., 2011). LARS-WG can adequately simulate the range of values; however, slightly lower absolute maximum temperatures are predicted in February and May to August with higher absolute minimum temperatures in March, August and September. The LARS-WG coefficient of determination value is second of the downscaling tools based on the results from Table 5 for extreme maximum temperature, and third in terms of the minimum temperatures. For WG-PCA there is very little deviation from the historical values except for a lower absolute minimum TMIN in November. This is due to the reshuffling procedure which allows for a slight deviation from the observed range, as the temporal window can include data from a neighbouring month. The WG-PCA model produces the best results for extreme temperatures with the highest values of the coefficients of determination, 0.997 and 0.988 for maximum and minimum temperatures, respectively.

Table 5.  Coefficients of determination for the climate characteristics presented in Figure 2 .

Climate Characteristics, London Airport	SDSM	LARS-WG	WG-PCA
Absolute maximum TMAX	0.904	0.964	0.997
Absolute minimum TMIN	0.978	0.970	0.988
Total monthly precipitation medians	0.546	0.808	0.692
Daily precipitation mean	0.934	0.929	0.890
Daily precipitation standard deviation	0.929	0.987	0.921

The second row in Figure 2 contains boxplots of the simulated total monthly precipitation values, with line plots showing historical medians. SDSM overestimates total monthly precipitation for JanuaryMay, August and October, with underestimations in September, November and December. However, all historical median values lie within the interquartile range of the simulated data. SDSM results for median values of total monthly precipitation generate an R-squared value of 0.546, the lowest of the three downscaling tools. LARS-WG reproduces most historical medians very well and outperforms the other weather generators based on the R-squared values. There are very slight overestimations in February, March, and AugustOctober. WG-PCA's R-squared values rank second for total precipitation simulation based on the results in Table 5 , with slight overestimations in February-April, September and October and underestimations in June and November.

The third row of Figure 2 shows the standard deviation and mean (heavy lines) of daily precipitation values by month. Results differ slightly for each downscaling approach. SDSM-generated mean daily precipitation values are close to the observations, with very small overestimations for August to October. SDSM-simulated values of standard deviations are fairly close to the historic record with underestimations of up to 1 mm for January, February, March, June, July and November. SDSM simulates the mean daily precipitation values slightly better than LARS-WG and WG-PCA based on the R-squared values in Table 5 , and has the second highest coefficient of determination for simulation of standard deviations. LARS-WG performs very well in simulation of both the mean and standard deviation of daily precipitation ( Figure 2 ). The model ranks second for mean daily precipitation simulation, with an R-squared value of 0.929, and outperforms the other two models for standard deviations, with a value of 0.987. WG-PCA-simulated precipitation standard deviations show slight overestimations from the observations for April, July and October. For the mean values, there is a small overestimation in October values and an underestimation in November. WG-PCA is outperformed by SDSM and LARS-WG for simulation of precipitation means and standard deviations with values of 0.890 and 0.921, respectively. In general, the fit for all models is fairly good and the coefficients of determination are relatively high, indicating the ability of each of the weather generators to sufficiently reproduce the precipitation characteristics.

In order to statistically evaluate the ability of the three weather generators to effectively reproduce the mean values of precipitation, maximum temperature and minimum temperature, the Wilcoxon rank-sum test is used on their monthly outputs; the p-values are presented in Table 6 . In all months, both SDSM and WG-PCA errors for precipitation means are found to be insignificant at the 95% confidence level (all p-values greater than 0.05). For LARS-WG, errors are significant in the months of January, June and September. For maximum temperatures, results show that the errors from WG-PCA and SDSM are insignificant. LARS-WG errors are also insignificant in all months but August at the 95% confidence level. SDSM and WG-PCA errors for minimum temperatures in all months are also insignificant. Again, the errors for LARS-WG are insignificant for all months but August.

Table 6.  Test results (p-values) of the Wilcoxon rank-sum test for the difference of means in observed and simulated daily precipitation, maximum temperatures and minimum temperature. Insignificant results are indicated in bold.

	Precipitation			Maximum Temperature			Minimum Temperature
Month	SDSM	LARS-WG	WG-PCA	SDSM	LARS-WG	WG-PCA	SDSM	LARS-WG	WG-PCA
Jan	0.765	0.035	0.511	0.879	0.3	0.953	0.765	0.235	0.987
Feb	0.448	0.312	0.318	0.918	0.144	0.795	0.86	0.674	0.767
Mar	0.798	0.698	0.869	0.103	0.327	0.073	0.756	0.012	0.072
Apr	0.716	0.863	0.297	0.497	0.132	0.711	0.122	0.175	0.444
May	0.688	0.384	0.274	0.671	0.785	0.851	0.255	0.252	0.472
Jun	0.612	0.041	0.91	0.711	0.825	0.94	0.673	0.593	0.875
Jul	0.307	0.129	0.788	0.803	0.839	0.908	0.711	0.163	0.92
Aug	0.176	0.447	0.992	0.745	0.000	0.306	0.968	0.000	0.564
Sep	0.302	0.000	0.734	0.55	0.476	0.416	0.752	0.702	0.385
Oct	0.582	0.099	0.64	0.745	0.494	0.132	0.12	0.234	0.173
Nov	0.745	0.737	0.399	0.298	0.428	0.123	0.29	0.106	0.369
Dec	0.739	0.410	0.874	0.285	0.126	0.322	0.961	0.683	0.338

In addition to the equality of means, the Levene's test is used to calculate p-values for the equality of variances in each of the weather generators for daily precipitation, maximum temperature and minimum temperature; the results are shown in Table 7 . For all weather generators, the monthly errors in estimations of daily precipitation variance are found to be insignificant at the 95% confidence level. For both SDSM and WG-PCA the errors for maximum and minimum temperature are found to be insignificant at the 95% confidence level. For all months from LARS-WG, errors in the daily temperature values are found to be significant. Perhaps this is due to the internal assumptions made in the generation of semi-empirical distributions for temperature.

Table 7.  Test results (p-values) of the Levene's test for the equality of variances in observed and simulated daily precipitation, maximum temperatures and minimum temperatures. Insignificant results are indicated in bold.

	Precipitation			Maximum Temperature			Minimum Temperature
Month	SDSM	LARS-WG	WG-PCA	SDSM	LARS-WG	WG-PCA	SDSM	LARS-WG	WG-PCA
Jan	0.979	0.349	0.694	0.874	0.000	0.707	0.768	0.000	0.403
Feb	0.692	0.787	0.6	0.528	0.000	0.599	0.744	0.000	0.897
Mar	0.806	0.981	0.937	0.657	0.000	0.561	0.539	0.000	0.941
Apr	0.676	0.807	0.403	0.679	0.000	0.937	0.932	0.000	0.925
May	0.582	0.477	0.341	0.744	0.000	0.934	0.466	0.000	0.89
Jun	0.948	0.575	0.93	0.568	0.000	0.274	0.685	0.000	0.673
Jul	0.881	0.361	0.905	0.83	0.000	0.863	0.465	0.000	0.498
Aug	0.337	0.628	0.568	0.813	0.000	0.444	0.591	0.000	0.264
Sep	0.427	0.831	0.782	0.521	0.000	0.425	0.905	0.000	0.833
Oct	0.574	0.262	0.826	0.497	0.000	0.367	0.428	0.000	0.535
Nov	0.804	0.662	0.355	0.378	0.000	0.522	0.966	0.000	0.468
Dec	0.558	0.889	0.992	0.455	0.000	0.59	0.589	0.000	0.974

Figure 3 contains boxplots of simulated temperature indices from all three downscaling tools, with the historical median shown as a dashed line and the lower and upper quartiles (25^th and 75^th percentiles) of the historical data shown as dotted lines. Definitions of the indices can be found in Table 4 .

Performance in the simulation of warm days is poor from each of the weather generators as the median of the historical record falls outside of their simulated interquartile ranges. Moreover, the interquartile range of the observed data (the distance between the dotted lines) is not matched by any of the weather generators. All models significantly underestimate the 75^th percentile, with not even their whiskers extending to the observed upper quartile value. Both SDSM and WG-PCA simulations overestimate the median, while LARS-WG underestimates it.

Simulation of warm nights is more accurate for SDSM, with a very small underestimation in the median of the historical record; however, the variability in results is still less than that of the observed data, which has larger interquartile range (dotted lines). LARS-WG greatly underestimates the number of warm nights, as the median of the historical record is close to the highest whisker of the simulated data. The upper quartile of warm nights is close to the lower quartile of the observed data, indicating poor performance of LARS-WG for extreme temperature simulation. WG-PCA simulates the median of observed warm nights quite well; however, this model also underestimates the interquartile range observed in the historical record.

Similar results are found for WG-PCA generation of cold days, with a close median, but again, an underestimation in the interquartile range of the observed data. SDSM slightly overestimates the number of cold days per year and also underestimates the interquartile range of the observed data, shown by the dotted lines. LARS-WG simulation of cold days and cold nights is poor, with the median of the historical values about 7 days above the simulated 75^th percentile. SDSM underestimates the median of cold nights per year by about 4 days per year, WG-PCA by about 5 days per year, and LARS-WG by 8 days per year; all models underestimate the interquartile range from the observed data.

Simulation of extreme temperature indices by LARS-WG is poor, as each index is underestimated and, in three of the cases, the whiskers of simulated data do not exceed the 75^th percentile from the observed record. This is also reflected in the LARS-WG results from the Levene's test for equality of variances in daily maximum and minimum temperatures where the weather generator showed significant errors (Tables 10 and 11). Results vary for SDSM and WG-PCA, but overall warm night and cold day simulations are better than the generated warm day and cold night values. All models are found to underestimate the variability in the observations, simulating a much smaller interquartile range for all temperature indices.

Boxplots of simulated precipitation indices are shown in Figure 4 , with the historical median plotted as a line. Table 4 contains definitions of the indices used. Simulation of heavy precipitation days is satisfactory for all models. LARS-WG performs best for heavy precipitation days, accurately simulating the quartiles and median of the observed record. SDSM and WG-PCA overestimate the median by two to three days, respectively, with WG-PCA slightly overestimating the quartiles from the historical record.

The number of wet days per year is best simulated by LARS-WG, which slightly underestimates the median and quartiles by about three days/year. SDSM overestimates the observed median and quartiles of yearly wet days by almost five days per year. The median of WG-PCA simulations is equal to the 25^th percentile of the observed record, with the upper quartile only a few days above the observed median, indicating a major underestimation by WG-PCA in terms of the number of wet days per year.

LARS-WG and WG-PCA produce similar results for the number of very wet days, slightly overestimating the value by one day per year. Both models simulate an upper quartile of about two days higher than observed. SDSM slightly underestimates the observed median of very wet days by about one day per year, also underestimating the quartiles.

SDSM performs very well for simulation of the highest 5-day precipitation amount, accurately reproducing the observed median and lower quartile, but slightly underestimating the upper quartile. WG-PCA simulates slight overestimations of about 8 mm in the observed median and quartile values. LARS-WG overestimates the median by about 10 mm and the upper quartile by more than 15 mm. Overall, the model performances are acceptable for the simulation of extreme precipitation indices; this is also indicated by the Levene's test for equality of variance in Table 9 where the errors are shown to be insignificant from all three of the weather generators.

Overall, LARS-WG and SDSM simulate precipitation more accurately than temperature. LARS-WG in particular does not simulate extreme temperature indices or the variability of temperature values well. This is consistent with results from other comparative studies (Khan et al., 2006; Qian et al., 2008). WG-PCA is able to simulate both precipitation and temperature with relative accuracy. Both SDSM and WG-PCA passed all statistical tests at the 95% confidence level, while the LARS-WG results showed significant errors in some months for precipitation and temperature means, and in all months for temperature variability.

Downscaling of AOGCM Precipitation Data

In order to predict potential future extreme precipitation events, LARS-WG and WG-PCA are used to downscale six monthly AOGCM outputs, each with two to three emissions scenarios, for a total of fifteen scenarios. Because SDSM requires daily, gridded AOGCM predictor inputs, downscaling is not performed with this tool as consistent inputs to both LARS-WG and WG-PCA are used in this study. The AOGCM descriptions are in Table 2 . As the weather generators generally perform better in the simulation of total precipitation and mean daily precipitation amounts, a comparison of these is provided. For consistency, the same input datasets are used with both downscaling tools.

Table 8 shows AOGCM-predicted percent changes in total seasonal precipitation and mean precipitation amounts from LARS-WG and WG-PCA for the 2050's time period. It can be seen from the table that a range of values is predicted by the weather generators.

Table 8.  AOGCM predicted percent change in average total seasonal precipitation and mean daily precipitation values, as simulated by SDSM, LARS-WG, and WG-PCA for the 2050's time period.

		Percent Change in Total Seasonal Precipitation (%)								Percent Change in Mean Daily Precipitation (%)
		Winter		Spring		Summer		Fall
Emissions Scenario	AOGCM	LARS-WG	WG-PCA	LARS-WG	WG-PCA	LARS-WG	WG-PCA	LARS-WG	WG-PCA	LARS-WG	WG-PCA
A1B	CGCM3T47	16.5	26.9	20.5	25.1	0.6	1.7	6.3	9.7	10.7	15.4
	CGCM3T63	13.0	22.2	15.8	16.8	1.1	4.6	41.5	48.4	18.0	23.6
	GISSAOM	5.5	10.6	13.1	15.4	13.1	15.5	11.5	14.6	11.1	14.2
	MIROC3.2HIRES	18.4	23.5	6.0	7.4	3.0	9.9	9.7	16.0	4.6	9.0
	MIROC3.2MEDRES	0.4	9.1	12.9	20.7	14.5	10.6	2.7	0.6	0.4	4.7
Average A1B	10.8	18.5	11.3	17.1	1.0	0.3	14.3	17.9	9.0	13.4
A2	CGCM3T47	21.7	21.8	15.0	20.9	1.2	1.7	10.6	8.5	11.9	12.9
	CGCM3T63	4.9	26.7	11.0	3.5	1.2	16.1	46.2	49.4	16.3	24.6
	CSIROMK3.5	9.8	13.9	15.2	18.5	20.9	28.8	7.6	6.7	13.4	16.9
	MIROC3.2MEDRES	0.9	10.3	7.5	31.8	8.2	1.1	4.1	5.3	1.5	12.0
Average A2	9.2	18.3	12.0	18.4	2.3	9.6	14.9	17.6	9.8	16.0
B1	CGCM3T47	11.9	33.5	18.4	26.8	2.4	0.3	6.3	11.9	8.4	17.5
	CGCM3T63	12.9	10.9	3.4	15.9	11.9	4.0	54.6	44.5	20.2	19.7
	CSIROMK3.5	4.1	21.6	16.1	22.5	27.4	18.4	9.2	9.0	14.4	17.3
	GISSAOM	1.2	13.2	0.5	9.8	12.8	17.6	6.2	9.2	4.7	12.4
	MIROC3.2HIRES	5.2	13.5	0.6	8.9	5.9	7.5	2.3	1.9	1.0	3.8
	MIROC3.2MEDRES	2.2	7.5	23.0	18.4	1.2	5.8	1.1	3.1	5.8	3.8
Average B1	3.8	13.3	6.9	15.1	9.5	5.3	13.8	12.3	8.8	11.4

For the A1B emissions scenario in the winter, AOGCM simulations predict an increase in total seasonal precipitation of 0.4% to 26.9%. The percent change simulated by each weather generator varies. For example, in the CGCM3T47 simulations, LARS-WG predicts a 16.5% increase and WG-PCA predicts a much larger increase of 26.9%. The closest agreement between the downscaling tools is about 5.1% between GISSAOM and MIROC3.2HIRES predictions. The average increase predicted from the A1B winter simulations is 11% from LARS-WG and 19% from WG-PCA.

For A1B spring simulations from MIROC3.2HIRES, LARS-WG predicts a decrease in total spring precipitation of 6%, while WG-PCA generates a 7.4% increase. Predictions are fairly close between the two downscaling tools for CGCM3T63 and GISSAOM. The average A1B change in spring precipitation is 11% from LARS-WG and 17% from WG-PCA.

Percent changes in summer precipitation for A1B vary greatly between the AOGCMs. The simulated percent changes for LARS-WG and WG-PCA, respectively, range from decreases of 14.5% and 10.6% from MIROC3.2MEDRES to increases of 13.1% and 15.5% from GISSAOM. While the AOGCM predictions vary greatly, the downscaling tools generally predict similar results from each AOGCM, with the highest discrepancy in predictions a 6.9% difference for MIROC3.2HIRES.

Fall precipitation totals from the different AOGCMs show a wide range of results for the A1B scenario. For CGCM3T63, LARS-WG and WG-PCA predict increases of 41.5% to 48.4%, respectively, while the models predict increases of only 2.7% and 0.6% for MIROC3.2MEDRES. The highest discrepancy between downscaling tools for the same AOGCM is 6.9% for CGCM3T47.

For winter from the A2 scenario, there is slightly more disagreement between the downscaling tools. While predictions are close for CGCM3T47, with both models predicting increases of about 22%, LARS-WG predicts a 4.9% increase for CGCM3T47, in contrast to a 26.7% increase from WG-PCA.

All models and downscaling tools predict increases from 7.5% to 31.8% in spring precipitation from the A2 scenario. The largest disagreement between LARS-WG and WG-PCA is 24.3% for the MIROC3.2HIRES model.

For summer precipitation changes from the A2 scenario, each AOGCM produces very different results, with increases of 20% to 29% predicted for CSIROMK3.5 and a decrease of 8.2% predicted by LARS-WG for MIROC3.2MEDRES. There is a 17.1% discrepancy between the results (1.1% and 16.2%) predicted by LARS-WG and WG-PCA, respectively, for CGCM3T63.

For the fall simulations for A2, LARS-WG and WG-PCA predict increases in most simulations around 15% to 18% on average. The largest increase for both models is 46.2% to 49.4%, predicted by CGCM3T63. MIROC3.2MEDRES simulations show a decrease in total precipitation of 4.1% from LARS-WG with an increase of 5.3% predicted by WG-PCA.

Results are again variable for the B1 simulations in all seasons. For winter, the total precipitation changes vary between the AOGCMs and downscaling tools from 2.2% to 33.5%. There is a 21.6% discrepancy between the results for CGCM3T47.

For spring precipitation, LARS-WG predicts a slight decrease for some models (3.4% for CGCM3T63) and a significant increase for others (23% for MIROC3.2MEDRES), and WG-PCA predicts increases for all models ranging from 9% to 27%. Summer predictions show a 5% to 10% increase on average, with some AOGCMs predicting increases (22.5% to 27.4% for CSIROMK3.5) and others predicting decreases (5.9% to 7.5% for MIROC3.2HIRES).

LARS-WG and WG-PCA predict increases in fall precipitation for most AOGCM simulations (except for both MIROC3.2 models). There is generally a relatively close agreement between the downscaling tools. CGCM3T63, LARS-WG and WG-PCA predict drastic increases in precipitation of 54.6% and 44.5%, respectively.

The three right-hand columns of Table 5 show the percent change in the mean daily precipitation values from the historical average. LARS-WG generally predicts a slightly smaller increase in daily precipitation than WG-PCA. There is a fairly large disparity in some models, for example, with the A2 scenario, MIROC3.2MEDRES-predicted changes are 1.5% and 12% from LARS-WG and WG-PCA, respectively. The averages for each emissions scenario from LARS-WG show increases of about 9% to 10%. For WG-PCA, increases of 11% to 16% on average are predicted.

Overall, the results are highly variable depending on the AOGCM and downscaling tool. Each AOGCM has a set of internal assumptions that affect the outputs, despite similar emissions-scenario inputs. While the same AOGCM-modified input datasets are used for both LARS-WG and WG-PCA, the statistical process of each of these tools varies significantly so the simulations often produce slightly different outputs for the same AOGCM. In general, a rise in fall to spring precipitation is predicted by most AOGCMs, and an overall increase in mean daily precipitation. For winter, precipitation changes of 2.2% to 33.5% are predicted by the AOGCMs, with only three of the 30 total AOGCM simulations projecting decreased precipitation. In spring, 6% to 31.8% is projected, and 26 of the 30 simulations project increasing precipitation. For summer precipitation, there is less agreement on the sign and magnitude of change, with a range of 14.5% to 28.8%, and 11 of the models project decreasing amounts. In the fall, a range of precipitation changes from 4.1% to 49.4% is projected, and all but three of the simulations show increases. Daily mean precipitation is predicted to change from 1.5% to 24.6%, and only two of the models predict decreasing amounts.

For the Upper Thames River Basin, which is prone to major flooding events in winter and spring, the quantification of these expected changes is particularly important so that municipalities can employ updated Intensity-Duration-Frequency (IDF) curves and calculate expected changes in the return period flood levels. The basin has a significant amount of flood protection infrastructure, including an extensive diking system and several reservoirs so it is essential that climate change impacts on this infrastructure are taken into consideration by the Upper Thames River Conservation Authority and the municipalities of the basin in their water resources management strategies.

Conclusions

The main objective of this study was to investigate the potential of three downscaling tools, namely SDSM, LARS-WG and WG-PCA, to simulate historical and future climate conditions for the Upper Thames River Basin in Ontario, Canada. Downscaling is necessary for climate change impact assessments as AOGCMs are spatially too large for direct application for watershed analysis.

In order to investigate the effectiveness of each downscaling tool, it is important that each weather generator adequately reproduce the historical climate. Preservation of historical statistics provides a means for comparison between the three tools, as each method downscales future AOGCM data in a different way. Each of the three downscaling tools was used to simulate twelve ensembles (324 years) of synthetic historical climate data. LARS-WG and WG-PCA were then used to simulate synthetic climate data for London, Ontario from the AOGCM-modified datasets.

Each downscaling tool requires a very different calibration procedure. LARS-WG is computationally very simple, as the user simply supplies the historical data and varies the random seed to produce the best possible output.

SDSM calibration is a time-consuming process as the proper selection of predictors and domain size is critical in order to adequately reproduce historical climate statistics. Many reanalysis predictors should be screened and domain-field data used for the best result. The choice of domain size should be varied to ensure the best possible calibration. SDSM offers an extensive number of transfer functions, different model types (seasonal, monthly or annual), as well as the option for conditional or unconditional processes. Various combinations of these must be tested to determine which conditions and predictors are able to best simulate historical climate variability.

WG-PCA requires only the selection of a temporal window length, since there are no statistical assumptions made about the historical climate. The user simply varies the length of the temporal window and inputs variables from any number of stations. The selection of the temporal window is important, as too wide of a window can result in unrealistic values being produced due to the strong seasonality of the study area.

Overall, calibration of WG-PCA and LARS-WG can be performed quickly and easily, while SDSM is more time consuming, requiring some background knowledge and a good understanding of the program in order to apply it effectively.

In general, all three downscaling tools perform best in simulation of total monthly precipitation and mean daily precipitation amounts. Simulations of maximum and minimum temperatures are good for SDSM and WG-PCA, which are shown to simulate means and variances in the observed record within the 95% confidence interval. LARS-WG performance is less acceptable in this regard as the errors are found significant in all months for the variances of daily temperatures and, in some months, for the means. For precipitation simulation, LARS-WG performance is best with fairly high coefficients of determination between the observed and simulated means of the selected indices. WG-PCA and SDSM also performed fairly well in this aspect. Overall, WG-PCA and SDSM simulate most climate indices fairly well. LARS-WG simulation of precipitation indices is good; however, the weather generator does not perform as well for simulation of temperature. For SDSM, gridded reanalysis outputs should be used instead of the single-point interpolated station data, and more predictors should be tested; the authors are confident that SDSM results could be improved this way.

The ability of each downscaling tool to simulate historical precipitation and temperature indices was also investigated. All models generally underestimate the observed variability in temperature indices. LARS-WG performs poorly in terms of temperature simulation, underestimating each of the indices. WG-PCA performs better in simulation of warm nights and cold days per year; however, there is an overestimation in the number of warm days and an underestimation in the number of cold nights. SDSM also under-simulates the number of cold nights and overestimates the number of warm days each year.

Simulation of precipitation indices is adequate for all models, with slight over or underestimations in the medians. SDSM performs best in simulation of the highest 5-day precipitation amount, accurately predicting the median of the observed data. LARS-WG slightly outperforms SDSM and WG-PCA in simulation of heavy precipitation days and wet days. WG-PCA and LARS-WG simulations of very wet days and highest 5-day precipitation amounts are very similar, with slight overestimations in the observed median. The models generally simulate indices of precipitation better than temperature, and results are variable for each index.

Downscaling of 15 AOGCMs was performed with LARS-WG and WG-PCA, and the results were found to be highly variable. For the same AOGCM scenarios, results can vary significantly between the downscaling tools. Discrepancies between the different AOGCMs are high, with some models predicting major increases and others predicting decreases for the same season and the same IPCC emissions scenario. Each AOGCM has its own unique set of assumptions and although their inputs are the IPCC's SRES emissions scenarios, results can be very different. Furthermore, the downscaling tools include different statistical assumptions. For a single AOGCM with the same input files used by both tools, there can be significant discrepancies in predictions of total monthly and mean daily precipitation.

In general, London fall, winter and spring rainfall totals are predicted to increase by most AOGCMs and downscaling tools, with less conclusive results for the summer period. The magnitude of change predicted by the models for these periods is less certain due to the differences between the downscaling tools and the AOGCM. For winter, a range of precipitation changes from 2.2% to 33.5% are projected and most models show increases. In spring, changes range from 6% to 31.8%, and again the majority of models show increasing precipitation. Models are less conclusive for summer precipitation changes, which range from 14.5% to 28.8%. In fall, projected changes range from 4.1% to 49.4% with most models showing an increase. On average, most simulations show an increase in mean daily precipitation amounts from both LARS-WG and WG-PCA; however, the magnitude of change varies depending on the downscaling tool and ranges from 1.5% to 24.6%. For effective management of the Upper Thames River Basin, it is important that these expected changes be taken into consideration for flood protection strategies and for the design of storm water collection systems.

Because each weather generator and AOGCM produces such different results, it is important to consider multiple downscaling approaches and various AOGCMs for a comprehensive climate change impact assessment. Inclusion of only one downscaling method and AOGCM would result in a biased prediction of future climatic conditions since each tool has its own inherent strengths and limitations.

In general, WG-PCA performance is satisfactory for simulations of both temperature and precipitation variability and indices. LARS-WG performs very well for precipitation simulation, but could not adequately simulate the variance and indices of daily temperatures. The AOGCM outputs show considerable variability between models and between the downscaling tools. Future research should focus on improving the downscaling methods to provide more consistent precipitation results that can be interpreted with an increased level of confidence.

Acknowledgements

The authors are thankful to the North American Regional Reanalysis (NARR) and Environment Canada for providing the climate data used in this study, as well as the Canadian Foundation for Climate and Atmospheric Sciences (CFCAS) and Ontario Graduate Scholarship (OGS) for funding this research.