Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract: General Circulation Models (GCMs) are widely used tools to assess potential impacts of
global climate warming. However, their outputs are difficult to use in regional impact studies with regard
to water resources because of their coarse spatial resolution. Downscaling techniques have emerged as
useful tools to reduce the problem of discordant scales by deriving regional climate information from
global climate data. The objective of this study is to test the capability of one of these techniques, the
Statistical DownScaling Model (SDSM), to derive local scale temperature and precipitation data series
that can be used as inputs to a hydrologic model for streamflow modelling. Three river basins located
in the province of Québec are analyzed. Results show that the SDSM provides reasonable downscaling
data when using predictors representing the observed current climate. However, the performance is less
reliable when using GCM predictors.
Résumé : Les modèles de la circulation générale (MCG) sont des outils utilisés pour évaluer les impacts
potentiels du réchauffement climatique global. Cependant, il est difficile d’utiliser directement leurs
données dans le cadre d’études d’impacts régionales, tel que celles reliées aux ressources en eau, en
raison de leur résolution spatiale grossière. Le développement des techniques de réduction d’échelle
spatiale a permis de réduire le problème d’échelles discordantes en dérivant l’information du climat
régional à partir de données sur le climat global. L’objectif de cette étude est de tester la capacité d’une
de ces techniques, le Statistical DownScaling Model (SDSM), à fournir des données à échelle réduite
adéquates de température et de précipitation à un modèle hydrologique pour la modélisation des débits
en rivière. Trois bassins versants de la province de Québec sont à l’étude. Les résultats démontrent que
SDSM réduit raisonnablement l’échelle spatiale des données en utilisant les variables atmosphériques
à grande échelle représentant le climat actuel observé. Cependant, la performance diminue avec celles
simulées par un MCG.
1
Département de géographie, Université de Montréal, Montréal, QC H3C 3J7
2
Département des génies civil, géologique et des mines, École Polytechnique de Montréal, Montréal, QC
H3C 3A7
3
Hydro-Québec, Unité Conception, Hydraulique et Géotechnique, Montréal, QC H2L 4P5
Submitted March 2005; accepted September 2005. Written comments will be accepted until June 2006.
Canadian Water Resources Journal Vol. 30(4): 297–314 (2005) © 2005 Canadian Water Resources Association
Revue canadienne des ressources hydriques
298 Canadian Water Resources Journal/Revue canadienne des ressources hydriques
For many decades, scientists were concerned with the The study area consists of three river basins located in
possibility that the earth’s climate could be modified by the province of Québec: Vermillon, Sainte-Marguerite
human activities such as fossil fuel burning, agriculture and Grande-Baleine (Figure 1). The distances between
and deforestation. During the 20th century, the global these basins are large and they have different climatic,
surface temperature has increased by more than half hydrologic and geographic features. Grande-Baleine is
a degree Celsius and the 1990s was, on average, the one of the main basins in northern Québec, where the
hottest decade for the last 1,000 years (IPCC, 2001). effects of climate change could be more important than
Several studies show that the increase in temperature in southern areas. The Vermillon and Sainte-Marguerite
cannot be entirely explained by natural climate basins are used for hydroelectric power production. It
variations. For instance, Levitus et al. (2001) compared is likely that water resources and hydroelectric power
global temperature data simulated by a General production could be affected by climatic change
Circulation Model (GCM) with those observed for (Robinson, 1997).
the second half of the 20th century. They showed that
a simulation based on both natural and anthropogenic
radiative forcing is in closer agreement with observed Data Used
data than one that only takes into account natural
climate variability. Data from many sources are used throughout the
GCMs are also used in global warming downscaling procedure. They are described prior to
impact studies with regard to water resources. The the procedure itself.
combination of GCM outputs and hydrologic models
has been an approach commonly used during the last 1. Climate station data (predictands): For each basin,
decade to assess possible future hydrologic changes temperature and precipitation series were extracted
of drainage basins (e.g., Roy et al., 2001; Loukas et for three stations from the Environment Canada
al., 2002; Whitfield et al., 2002). GCMs proved to climatic database (Table 1). These stations have data
be efficient on the global scale. However, their use in series covering the 1961-1990 period (the maximum
regional applications, such as river flow simulations range permitted by SDSM) and have a minimum
when coupled with hydrologic models, is limited by of missing data. The choice was constrained by the
their coarse spatial resolution. Downscaling techniques sparse station network in Québec.
have emerged as useful tools to reduce the problem
of discordant scales between coupled models. These 2. Large-scale atmospheric variables (predictors):
techniques can be classified into four groups: Regional Observed and modelled predictors come, respectively,
Climate Models (RCM), regression-based techniques, from the National Centre for Environmental
weather patterns and stochastic weather generators. Prediction (NCEP) reanalysis (Kalnay et al., 1996)
The strengths and weaknesses of these techniques and the first-generation Canadian GCM (CGCM1).
are described in the literature (Giorgi et al., 2001). They were extracted for the grid point closest to
Downscaling methods need to be tested prior to their each climate station from the Canadian Institute for
use in climate change studies. Climate Studies (CICS) website (http://www.cics.
This paper presents an application of a downscaling uvic.ca/scenarios). Table A1 (see Appendix) lists
approach developed in recent years, the Statistical available predictors.
DownScaling Model (SDSM; Wilby et al., 2002).
The paper aims to test whether the SDSM is useful 3. Raw CGCM1 data: Unlike predictors, raw
in deriving adequate local climate data for streamflow surface temperature and precipitation data series
modelling from GCM information. were extracted directly from the CGCM1 model for
comparison needs (downscaled versus raw).
Sainte-Marguerite
Grande-Baleine
Users are allowed to simulate, through combinations of practices. Monthly percentages of explained variance
regressions and weather generators, sequences of daily show the capability of a given predictor to explain
climatic data for present and future periods by extracting local climate variability. Partial correlation coefficients
statistical parameters from observed data series. The applied to the most suitable predictors help eliminate
stochastic component of SDSM permits the generation those for which the weights are not important enough
of 100 simulations. Wilby et al. (1998; 1999) describe to influence the regression equations.
the mathematical formulations of SDSM. The ‘Calibration’ step involves the establishment
Table 3 summarizes the step-by-step procedure. The of statistical relationships between the selected
‘Screening’ process is crucial for the creation of credible predictors and the surface predictands. In this study,
downscaling scenarios. SDSM provides quantitative 12 regression models are created, that is, one for each
tools to assist in choosing a realistic set of predictors, month (seasonal or annual models are possible). The
even though local climate knowledge is part of best simulation process is conditional for precipitation, as
Step Description
1. Screening Selection of NCEP predictors related to the strength of each predictor – predictand relationship,
which is analyzed with monthly percentages of explained variance and partial correlation
coefficients.
2. Calibration Creation of monthly regression models using selected predictors. A simulation is performed using
the first half of each temperature/precipitation data series. A percentage of explained variance and
a value of standard error report the quality of the calibration, which is maximized with mean bias
correction and variance inflation parameters.
3. Validation Simulation of daily temperature/precipitation data series for the second half of each data series
with the calibrated regression models. t-tests and F-tests compare simulated and observed monthly
means and variances.
4. Climate scenario Simulation of daily temperature/precipitation data series for the whole 1961-1990 period with
both “observed” (NCEP) and modelled (CGCM1) predictors. Monthly means and variances are
compared using t-tests and F-tests. Raw CGCM1 temperature/precipitation data are used for
comparison of monthly residuals with downscaled data.
5. Hydrologic application Downscaled temperature and precipitation data series serve as inputs to the SSARR hydrologic
model for the simulation of streamflow hydrographs.
the amount for a given day is conditioned by the wet- (from ‘Climate scenario’) are used as inputs for the
day occurrence, which acts as an intermediate variable. Streamflow Synthesis And Reservoir Regulation
Precipitation data are transformed (fourth root or (SSARR) hydrologic model. SSARR has low data
natural log) prior to calibration. requirements and ease of use for the purpose of this
The stochastic component of SDSM allows for study. It is a conceptual model that requires only daily
performing 100 simulations in the ‘Validation’ process. temperature and precipitation data to perform long-
In this study, 20 simulations are performed and the term continuous hydrologic simulations. The main
one that most closely fits the observed data is chosen. focus of this study being the SDSM software, SSARR
The authors are aware that this approach is unusual. features are not described (see USACE, 1991).
Even though variations amongst simulations can be SDSM allows for the simulation of future climate
important, the closest agreement is privileged. Some scenarios, but this is not within the scope of this paper.
limitations may arise from this approach, notably for
future scenarios, which have no observed equivalent. In
such case, the simulation has to be clearly representative Results
of the GCM’s predictions. The ‘Climate scenario’
procedure is similar, except that the whole 1961-1990 Table A2 (see Appendix) shows screening results by
data series is used and that modelled predictors and providing, for each climate station, partial correlation
raw CGCM1 inputs are involved. For the sake of coefficients associated with a series of predictors.
conciseness, the study focuses on mean and variance These predictors were chosen following a preliminary
although other statistics (extreme values, percentage of analysis of monthly percentages of explained variance
wet-days, etc.) could also be evaluated (see Gagnon, for all available predictors. For surface temperature, the
2004). regionalized temperature (temp) predictor presents the
To help visualize the performance of SDSM, highest correlation coefficients (from 0.738 to 0.857).
downscaled temperature and precipitation data Thus, this predictor is the only variable selected for
the following step (calibration). For precipitation, also modified for most of the stations, showing that the
correlation coefficients are always low (below ± 0.2). variance of precipitation was generally underestimated
Predictors selected for model calibration are bold in by SDSM.
Table A2.
Table 4 presents the results of the calibration at Table 5. Mean bias and variance inflation parameters.
each meteorological station by giving the percentage
of explained variance and the standard error. For Temperature Precipitation
temperature, the best and the worst performances are Basin
observed at Nitchequon (80.5%) and Sept-Iles (67.4%), Variance Mean Variance
respectively. For precipitation, these are seen at Sept- Climate Station Inflation* Bias* inflation*
Iles (48%) and Nitchequon (28.8%). Except for Sept-
Iles, percentages for precipitation are all under 40%. Vermillon
Grande-Baleine
Kuujjuarapik 74.7 2.938 31.2 0.773 Table 6 compares the performance of SDSM for
Nitchequon 80.5 2.538 28.8 0.781 simulating temperature for both validation and
Schefferville 72.1 2.840 38.3 0.777 climate scenario steps with t-tests (means) and F-tests
(variance). The numbers of months with a statistically
significant difference are indicated. Concerning
Table 5 shows the best combinations of mean bias validation, which involves the use of NCEP predictors,
correction and variance inflation parameters for most of the stations have a maximum of one significant
each station. For temperature, many stations have a difference in means, with the exception of Saint-
value of variance inflation lower or higher than 12, Michel-des-Saints (two differences). Results are even
indicating that the variance was first overestimated better for variance, as only one significant difference is
or underestimated by SDSM. There is no mean bias exhibited for Gagnon, Kuujjuarapik and Schefferville.
correction as temperature only involves an unconditional With regard to the climate scenario, which involves
process. For precipitation, which involves a conditional the use of CGCM1 predictors, the results are different
process, many stations required a reduction of the bias from those obtained following validation. Most of
correction parameter (lower than one), indicating a the stations exhibit a minimum of seven statistically
previously overestimated mean. Variance inflation was significant differences in means. The performance is
especially significant at Gagnon and Sept-Iles, for Table 7. t-test and F-test results shown by statistically
which there is a difference for almost every month. significant differences for precipitation. Numbers
However, the variance simulation is similar in many indicate the number of cases where a significant
instances to the validation results. difference was found between monthly means (t-test)
and variances (F-test).
Table 6. t-test and F-test results shown by statistically
significant differences for temperature. Numbers Validation Climate Scenario
indicate the number of cases where a significant Basin
difference was found between monthly means (t-test) Mean Variance Mean Variance
and variances (F-test). Climate Station (n = 12) (n = 12) (n = 12) (n = 12)
Kuujjuarapik 0 1 7 1
Nitchequon 1 0 8 0 modelling in northern basins. It can be seen that the
Schefferville 1 1 8 1 30 year-averaged temperatures are lower (below 0°C),
for all stations, when SDSM uses CGCM1 instead
of NCEP predictors. Temperatures generated using
Concerning precipitation, variance is generally well NCEP predictors are closer to reality.
reproduced for most months for both the validation In order to test the ability of SDSM to
and climate scenario steps (Table 7). However, only realistically downscale GCM data, a comparison is
one station exhibits no significant differences for made between downscaled (with CGCM1 predictors)
validation (La Tuque). The t-tests show that SDSM and raw CGCM1 data and the new “observed” series
failed many times to reproduce means, as a minimum (downscaled with NCEP predictors) for the 1961-
of three significant differences are observed at many 1990 period. Figure 2 shows the mean monthly
stations for both steps. temperature residuals for three representative
As shown in Table 6, t-tests revealed many stations, one in each of the selected basins. The
significant differences in temperature means (‘Climate graphs do not show the sign (positive or negative)
scenario’), with some differences having important of the residuals, as they only aim at showing whether
implications. Table 8 presents a comparison of downscaled data are closer to reality than raw data.
temperatures at the start of the snowmelt season, According to these results, SDSM adequately
which is of paramount importance in streamflow performed the task of downscaling temperature, as
most of the downscaled data residuals are lower than This is due to the difference in temperature (Table 8).
those of raw CGCM1 data. This is also generally The overestimated magnitude of peak flows is caused
true for precipitation residuals for the same stations by the accumulation of snow that was supposed to fall
(Figure 3). Nonetheless, these data are mean monthly as rain. Nonetheless, the main limiting factor for these
values. Other results (Tables 6 and 7) showed that two basins appears to be the poor simulation of spring
the distribution of downscaled variables using SDSM temperature.
may be statistically different from observed data.
Figure 4. Mean monthly streamflow hydrographs (see Table 2 for the exact time
period): a) Vermillon, b) Sainte-Marguerite, and c) Grande-Baleine basins.
residuals between observed and NCEP predictors- Dibike, Y.B. and P. Coulibaly. 2005. “Hydrologic
based data (‘Validation’), but not between the latter and Impact of Climate Change in the Saguenay Watershed:
CGCM1 predictors-based data (‘Climate scenario’), as Comparison of Downscaling Methods and Hydrologic
performed in the present study. Models.” Journal of Hydrology, 307: 145-163.
Busuioc, A., D. Chen and C. Hellstrom. 2001. Levitus, S., J.I. Antonov, J. Wang, T.L. Delworth,
“Performance of Statistical Downscaling Models K.W. Dixon and A.J. Broccoli. 2001. “Anthropogenic
in GCM Validation and Regional Climate Change Warming of Earth’s Climate System.” Science, 292:
Estimates: Application for Swedish Precipitation.” 267-290.
International Journal of Climatology, 21: 557-578.
Loukas, A., L. Vasiliades and N.R. Dalezios. 2002. Wilby, R.L., L.E. Hay and G.H. Leavesley. 1999. “A
“Potential Climate Change Impacts on Flood Comparison of Downscaled and Raw GCM Output:
Producing Mechanisms in British Columbia, Canada, Implications for Climate Change Scenarios in the San
Using the CGCMA1 Simulation Results.” Journal of Juan River Basin, Colorado.” Journal of Hydrology, 225:
Hydrology, 259: 163-188. 67-91.
Nguyen, T.D., V.T.V. Nguyen, P. Gachon and Wilby, R.L., C.W. Dawson and E.M. Barrow. 2002.
A. Bourque. 2004. “An Assessment of Statistical “SDSM – A Decision Support Tool for the Assessment
Downscaling Methods for Generating Daily of Regional Climate Change Impacts.” Environmental
Precipitation and Temperature Extremes in the Greater Modelling & Software, 17: 147-159.
Montreal Region.” In Proceedings of the 57th Annual
Conference of the Canadian Water Resources Association,
Montréal, QC, June 16-18, 2004, 10 p.
Appendix
Predictor Symbol
Surface 500 hPa 850 hPa
Vermillon
La Tuque p500 0.133 0.0000 p_v 0.064 0.0000
s850 -0.145 0.0000 p_zh 0.004 0.5236
sphu 0.081 0.0000 p8_v 0.061 0.0000
temp 0.738 0.0000 p8zh 0.008 0.3901
s500 0.111 0.0000
Sainte-Marguerite
Gagnon p500 0.291 0.0000 p_v 0.072 0.0000
s850 -0.159 0.0000 p8_v 0.083 0.0000
sphu 0.025 0.0600 s500 0.161 0.0000
temp 0.795 0.0000
Grande-Baleine
Kuujjarapik p_zh -0.012 0.2416 mslp -0.041 0.0000
p500 0.137 0.0000 p_z 0.060 0.0000
p8zh 0.059 0.0000 p8_z 0.039 0.0002
s500 0.017 0.1073 p850 0.020 0.0567
s850 -0.074 0.0000 s500 0.136 0.0000
sphu -0.065 0.0000 s850 0.100 0.0000
temp 0.809 0.0000