Volume 42, Issue 2
Regular Article
Free Access

A framework for assessing uncertainties in climate change impacts: Low-flow scenarios for the River Thames, UK

R. L. Wilby

R. L. Wilby

Climate Change Unit, Environment Agency, West Bridgford, UK

Search for more papers by this author
I. Harris

I. Harris

Climatic Research Unit, University of East Anglia, Norwich, UK

Search for more papers by this author
First published: 28 February 2006
Citations: 13


[1] A probabilistic framework is presented for combining information from an ensemble of four general circulation models (GCMs), two greenhouse gas emission scenarios, two statistical downscaling techniques, two hydrological model structures, and two sets of hydrological model parameters. GCMs were weighted according to an index of reliability for downscaled effective rainfall, a key determinant of low flows in the River Thames. Hydrological model structures were weighted by performance at reproducing annual low-flow series. Weights were also assigned to sets of water resource model (CATCHMOD) parameters using the Nash-Sutcliffe efficiency criterion. Emission scenarios and downscaling methods were unweighted. A Monte Carlo approach was then used to explore components of uncertainty affecting projections for the River Thames by the 2080s. The resulting cumulative distribution functions (CDFs) of low flows were most sensitive to uncertainty in the climate change scenarios and downscaling of different GCMs. Uncertainties due to the hydrological model parameters and emission scenario increase with time but were less important. Abrupt changes in low-flow CDFs were attributed to uncertainties in statistically downscaled summer rainfall. This was linked to different behavior of atmospheric moisture among the chosen GCMs.

1. Introduction

[2] General circulation models (GCMs) are one of the primary instruments for obtaining projections of future global climate change. Whether directly or indirectly, outputs from GCMs also underpin most climate change impact assessments. However, it is widely acknowledged that disagreements between different GCMs over regional climate changes represents a significant source of uncertainty [Jenkins and Lowe, 2003]. Although common practice, overreliance on a single GCM could lead to inappropriate planning or adaptation responses. In this paper we present a probabilistic framework for blending climate projections from different GCMs, greenhouse gas emission scenarios and downscaling techniques with impact model uncertainty, in order to produce probabilistic information on future river flows for water resource planning.

[3] Research into probabilistic forecasts of climate change has been advancing rapidly on several fronts. For example, there have been systematic evaluations of uncertainties due to climate model projections using: multimodel ensembles [Räisänen and Palmer, 2001; Giorgi and Mearns, 2003]; multiensemble experiments with one GCM [Murphy et al., 2004]; megaensemble runs supported by the general public [Stainforth et al., 2002, 2005]; and sensitivity analyses of key processes in climate model emulators [Wigley and Raper, 2001]. Bayesian methods have been applied to multimodel ensembles to characterize uncertainty and probability distribution functions (PDFs) for future climate changes at regional scales [Allen et al., 2000; Benestad, 2004; Stone and Allen, 2005; Tebaldi et al., 2004, 2005]. However, probabilistic information presents new opportunities as well as challenges for policy makers more accustomed to working with fewer climate change realizations [Dessai and Hulme, 2003; Reilly et al., 2001; Webster, 2003; Webster et al., 2003]. To date, there have been relatively few impact assessments using probabilistic climate change information [e.g., Luo et al., 2005].

[4] Environmental agencies will therefore need to develop new risk-based frameworks for handling probabilistic data and for estimating uncertainties inherent to impact models used for air quality (including diffuse pollution loads), land quality (including soil status, erosion and pollution pathways), and water resources (including surface-groundwater coupling and concomitant impacts on water quality and ecosystems). Otherwise, the range of risks posed by climate change to major assets and infrastructure developments may be underestimated. This could have significant consequences for policy and planning, by affecting the timing and cost of adaptation responses [e.g., Hall et al., 2005]. Although the probabilistic framework set out below illustrates uncertainty in future low flows for the River Thames basin, the case study is offered as a possible “road map” for other environmental assessments.

2. Case Study River Basin and Data Sources

[5] The Thames basin is relatively sheltered from the influence of midlatitude depressions and, as a consequence, is one of the driest and sunniest parts of mainland Britain. Around the Thames estuary annual rainfall totals average just 500 mm, rising to 600–650 mm across the Thames Valley, and over 1000 mm on the South Downs [Mayes, 1997]. In the Thames region, presently 55% of the effective rainfall that falls annually is deployed, amounting to about 5000 ML/d, of which 86% is used for public water supply. Even without climate change, the present balance of supply and demand is in deficit by some 180 ML/d [Environment Agency, 2001] and drought is considered a regional hazard, particularly in summer when average potential evaporation exceeds total precipitation [Marsh, 2004].

[6] Daily precipitation, potential evaporation and naturalized river flow data were obtained for the River Thames upstream of a gauge at Kingston (51:25:02N, 00:20:51W) for the period 1961–1990. Daily area average precipitation (PPT) amounts were derived from a network of 12 gauges distributed throughout the basin [Davis, 2000]. A wet day threshold of 0.3 mm was applied in order to remove trace rainfall amounts. Monthly Penman potential evaporation (PET) amounts were obtained from the MORECS operational system for the Thames basin and converted to daily values using the sine curve method [Davis, 2001].

[7] All atmospheric predictor variables used for calibrating the regional climate scenario tool (see below) originate from the National Centers for Environmental Prediction (NCEP) reanalysis data set [Kalnay et al., 1996], but were processed to conform to the 2.5° latitude × 3.75° longitude grid of the Hadley Centre's coupled ocean/atmosphere climate model HadCM3 [Harris, 2004]. The archive contains daily predictors (describing atmospheric circulation, thickness, and moisture content at the surface, 850 and 500hPa), for nine regions covering the British Isles, for the period 1961–2000. Airflow predictors were derived from an array of grid boxes centered on the southeast England (SE) grid box (Figure 1). This locale maximizes the amount of climate model information drawn from the surrounding land mask, and best reflects the near-continental climate of the Thames Basin. Furthermore, previous studies consistently show maximum downscaling skill for this region using predictors drawn from the south and west of the target area [e.g., Harpham and Wilby, 2005].

Details are in the caption following the image
River Thames basin upstream of the Kingston gauge in relation to GCM grid boxes used for the statistical downscaling.

[8] Future climate scenarios were generated using predictor variables for the same grid boxes supplied by four GCMs (Table 1) and two emission scenarios (the medium-high (A2) and medium-low (B2) Emissions of the Intergovernmental Panel on Climate Change Special Report on Emission Scenarios (IPCC SRES)) for the period 1961–2100. All GCM predictors were normalized with respect to their respective 1961–1990 climatologies, and regridded to conform to the Hadley Centre model grid shown in Figure 1.

Table 1. Four GCMs Used in the Analysis
Model Source Climate Sensitivity References
CGCM2 Canadian Center for Climate Modelling and Analysis 3.59 Flato and Boer [2001]
CSIRO Mk2 Commonwealth Scientific and Industrial Research Organisation in Australia 3.50 Gordon and O'Farrell [1997]
ECHAM4 Max-Planck-Institut for Meteorology and Deutches Klimarechenzentrum in Hamburg 3.11 Stendel et al. [2000]
HadCM3 UK Meteorological Office's Hadley Centre 3.38 Gordon et al. [2000]

3. Methodology

[9] The trial framework for obtaining probabilistic information (on future river flows) involved three steps: (1) downscale GCM output to produce regional climate change scenarios, in this case for the Thames basin, (2) assign weights to the components of uncertainty (emission scenario, GCM, downscaling technique, hydrological model parameters and structure), and (3) perform Monte Carlo simulations of impacts using scenarios from task 1 and weights from task 2. Each element is described below.

3.1. Downscale GCM Output

[10] Climate change scenarios for the the Thames basin were downscaled using two techniques: (1) the change factor (CF) method and (2) the statistical downscaling model (SDSM) [Wilby et al., 2002]. The CF method is a relatively straightforward procedure for constructing regional climate change scenarios and has been widely used for rapid assessment of climate change impacts [e.g., Arnell, 2004; Diaz-Nieto and Wilby, 2005]. First, the baseline weather series (in this case daily PPT and PET) were obtained for the Thames Basin for the period 1961–1990. Second, monthly mean changes in equivalent variables from the two emission scenarios and four GCMs were calculated for the SE grid box. Finally, percentage changes in monthly mean PPT and PET (for the 30-year blocks centered on the 2020s, 2050s and 2080s) were used to scale the baseline weather series. Although CF scenarios combine detail from the observed series with the average climate changes projected by the GCMs, scaled and baseline scenarios differ only in terms of their respective means, maxima and minima; all other properties of the data, such as the range and variability remain unchanged.

[11] This limitation is overcome by the second downscaling technique. SDSM is best described as a hybrid of the stochastic weather generator and regression-based methods, because large-scale daily circulation patterns and atmospheric moisture variables are used to condition local-scale weather generator parameters at individual sites. The stochastic component of SDSM enables generation of multiple simulations with slightly different time series attributes, but the same overall statistical properties. (For full technical details and split sample tests of SDSM, see Diaz-Nieto and Wilby [2005] or Harpham and Wilby [2005].)

[12] Downscaling using SDSM involved two main steps. First, statistical relationships were established between the target variables of interest (i.e., daily PPT and PET) and large-scale indices of regional weather over SE obtained from the NCEP reanalysis for the current climate. Table 2 shows partial correlation coefficients for significant predictor-predictand relationships. These reflect the amount of explanatory power unique to individual predictors after considering cross correlations among variables. On average, local PPT is most strongly correlated with the strength of southerly airflows and humidity, whereas PET is most heavily weighted toward regional temperatures. Second, the empirical predictor-predictand relationships for the observed climate were used to downscale ensembles of the same local variables for the future climate, using data supplied by the four GCMs listed in Table 1 driven by the two emission scenarios (A2 and B2) for the full period 1961–2100.

Table 2. Large-Scale Atmospheric Predictor Variables Used to Downscale Daily Precipitation (PPT) and Potential Evapotranspiration (PET) for the River Thamesa
Predictand Predictors (NCEP Reanalysis) Notation Partial r
PPT mean sea level pressure MSLP +0.20
PPT near-surface specific humidity SHUM +0.26
PPT near-surface southerly wind component VSUR +0.29
PPT 850 hPa geopotential height H850 −0.24
PET near-surface westerly wind component USUR −0.28
PET near-surface specific humidity SHUM −0.50
PET mean regional temperature at 2m TEMP +0.72
PET 500 hPa geopotential height H500 −0.28
  • a The partial correlation coefficient (r) shows the explanatory power that is specific to each predictor. All are significant at p = 0.01.

3.2. Assigning Weights to Components of Uncertainty

[13] Assigning weights to different components of the Monte Carlo analysis is one of the most problematic steps. Hence Dessai and Hulme [2003] recommend that such value-laden assumptions should be transparent, coherent and consistent with scientific understanding. Accordingly, the weighting of each component is explained in turn.

3.2.1. Emissions Scenario Uncertainty

[14] At present, there are no universally accepted methods of assigning probabilities to different emission pathways because of uncertainty in key drivers of emissions, such as assumptions about the future world economy [e.g., Webster et al., 2002]. Following Giorgi and Mearns [2003], the A2 and B2 emission scenarios were given equal probability (i.e., p(A2) = 0.5, p(B2) = 0.5) in both conditional and unconditional experiments. Note that the choice of emission scenarios was constrained by the availability of archived daily data required for the SDSM downscaling.

3.2.2. Climate Model Uncertainty

[15] Murphy et al. [2004] introduced a climate prediction index (CPI) as an objective means of weighting different GCMs according to their relative ability to reproduce present-day climate variables. Their CPI was based on a broad suite of surface and atmospheric variables. We refine the concept by calculating an impact relevant climate prediction index (IRCPI) based on the skill of each GCM/SDSM pair at reproducing meteorological variables critical to low-flow estimation (in the River Thames Basin). This means that the IRCPI is weighted according to each GCMs' ability to reproduce climate information most pertinent to the impact assessment.

[16] Effective rainfall (PPT minus PET) is the single most important determinant of (naturalized) annual low flows in the River Thames. This is convenient because it provides a combined metric for weighting the performance of GCM/downscaling. Despite a significant groundwater component, the Q95 statistic (the daily mean discharge exceeded on 95% of days in each year) is more strongly correlated with summer (r = 0.71) than spring (r = 0.45) or winter (r = 0.14) effective rainfall. Therefore the IRCPI was derived from the average bias in summer effective rainfall (Table 3). Results for winter (December to February) effective rainfall are also shown for comparison.

Table 3. Weights Assigned to Each GCM on the Basis of the Chosen IR-CPIa
Model Summer Winter
Bias, % Weight Bias, % Weight
CGCM2 52.1 0.138 42.2 0.074
CSIRO 14.3 0.503 6.0 0.522
ECHAM4 49.6 0.145 16.1 0.194
HadCM3 33.6 0.214 14.9 0.210
NCEP 7.6 n/a 4.8 n/a
  • a That is, proportional to the mean bias in effective precipitation across the Thames Basin.

[17] Figure 2a shows the cumulative distribution functions (CDFs) of summer (June to August) effective rainfall produced by each GCM/downscaling pair compared with observations for the period 1961–1990. Downscaling using NCEP predictors yields perfect prognosis results and hence a basis for assessing biases due to SDSM alone. Differences between the CDFs of observed and downscaled NCEP effective rainfall totals (i.e., biases due to SDSM) are largest for the most extreme dry summers. This implies that summer drought severity (and hence low-flow minima) could be underestimated in SDSM scenarios. However, the bias when downscaling from NCEP was always less than the bias when downscaling from GCMs (Table 3).

Details are in the caption following the image
CDFs of combined GCM/SDSM performance compared with observed and perfect prognosis (NCEP) estimates of (a) summer effective precipitation across the Thames basin and (b) winter effective precipitation across the Thames basin.

[18] Visual inspection of the CDFs in Figure 2a suggests that CSIRO totals most closely resemble totals downscaled from NCEP, whereas CGCM2, ECHAM4 and HadCM3 have relatively large positive biases. Differences between GCMs are even more pronounced for effective rainfall in winter (Figure 2b). Downscaling from both CSIRO and CGCM2 yields drier winters than downscaling from NCEP; conversely HadCM3 and ECHAM4 have positive biases that are more pronounced in the wettest winters. These biases are reflected in the weights assigned to each GCM (Table 3). Overall, the distribution of effective rainfall totals downscaled from CSIRO most closely resembles those of NCEP in summer, so the scenarios produced by this GCM are given greatest weight in the following Monte Carlo analysis. Note also that CSIRO best replicated the observed frequency of rain days in summer (not shown).

3.2.3. Downscaling Uncertainty

[19] Although downscaling uncertainty is represented by the SDSM and CF methods, it is acknowledged that substantial differences can arise between future scenarios downscaled from dynamical and empirical methods [Wood et al., 2004] or even when comparing different empirical downscaling methods [Wilby et al., 1998]. Therefore downscaling uncertainty is probably underestimated in the following analyses. By definition, the CF method is based on observed data; therefore it is not a straightforward matter to weight the two downscaling methods using skill measured against observed climatology. Furthermore, the CF and SDSM methods have very different strengths and weaknesses [Diaz-Nieto and Wilby, 2005]. On this basis, the CF and SDSM approaches were given equal weighting in the Monte Carlo analysis.

3.2.4. Hydrologic Model Parameter Uncertainty

[20] Impacts model parameter uncertainty is illustrated using CATCHMOD, a water balance model used by Thames Region Environment Agency (EA) for water resource planning [Davis, 2001; Wilby et al., 1994; Wilby, 2005]. Daily river flows at Kingston were simulated using daily PPT and PET inputs to a three-zone model comprising base flow from the chalk aquifer, runoff from clay areas and effluent returns from urban areas (Figure 3). Parameter definitions and manually calibrated values provided by the EA are shown in Table 4 (R. J. Davis and B. Greenfield, personal communication, 2004).

Details are in the caption following the image
CATCHMOD water balance model of the River Thames basin.
Table 4. Definition of CATCHMOD Parameters Used for Daily Flow Simulation in the River Thames at Kingstona
Parameter and Description Code Zone 1 (Base Flow From Aquifers Mainly Chalk) Zone 2 (Runoff From Clay Areas) Zone 3 (Runoff From Urban Areas)
Area of contributing zone, km2 AREA 3900 4100 600
Direct percolation,b % DP 20 0 0.5
Potential drying constant,c mm PDC 80 100 0
Gradient of the drying curved GDC 0.3 0.3 0.3
Linear storage constant,e days LSC 20 2 0.5
Nonlinear storage constant,f d/km2 NSC 300 2 0.25
Effluent return time series,g ML/d EFF 0 0 900
  • a Parameters in bold must be calibrated for each zone.
  • b A fixed fraction of precipitation that bypasses the soil horizon even during periods of soil moisture deficit.
  • c Value of deficit above which evaporation occurs at a reduced rate.
  • d The reduced rate at which soil moisture is evaporated once the potential drying constant has been exceeded.
  • e Represents temporary storage in the unsaturated zone.
  • f Represents storage in the saturated zone/aquifer.
  • g The net change to river flow due to imported/exported discharges.

[21] In CATCHMOD a direct percolation mechanism allows fixed proportions (DP) of incoming precipitation, which exceeds the potential evaporation rate, to bypass the soil store even during periods of soil moisture deficit (Figure 3). This process represents the observed behavior of fractured soils and macropores during summer rainfall and is only relevant to soils overlying permeable strata. The soil moisture submodel is based on a drying curve such that when the supply of moisture is limited, evaporation occurs at a constant proportion (GDC) of the potential rate. The value of the soil moisture deficit above which evaporation occurs at the reduced rate is termed the potential drying constant (PDC). The “upper” soil horizon therefore has a finite capacity equal to this constant. The “lower” horizon is depleted by the reduced evaporation rate only when the upper horizon is empty, and can accumulate large deficits in droughts.

[22] During recharge, wetting by precipitation fills the upper soil horizon before replenishing the lower horizon. When a contributing zone becomes saturated, excess moisture from the soil store (along with DP) contributes to total percolation. This flow is held temporarily in a linear store (LSC) representing the unsaturated zone. Where a soil is underlain by permeable geological formations, excess water from the overlying soil zone percolates through LSC to the aquifer below and is released at a nonlinear rate (NSC) from the groundwater store. Different parameters are used to represent the conditions in each contributing zone (Table 4) and river flow generated from multiple zones is summed to give the total daily discharge at Kingston.

[23] Despite acknowledged limitations, lumped conceptual models like CATCHMOD are still widely used for climate change impact assessment and water resource planning. Uncertainties in river flow projections arise from the choice of model calibration period, model structure, and nonuniqueness of model parameter sets. Monte Carlo simulations with CATCHMOD show that uncertainty in flow changes due to parameter uncertainty is higher in winter than in summer, and comparable in magnitude to the uncertainty due to future emission scenario [Wilby, 2005]. The present analysis employs the parameter sets of the 100 most skillful model simulations of river flow for 1961–1990 previously identified for the Thames basin [Wilby, 2005]. Each parameter set was weighted by the Nash and Sutcliffe [1970] nondimensional efficiency criterion, noting that all sets yield an NS score ≥ 0.8.

3.2.5. Hydrologic Model Structural Uncertainty

[24] Uncertainty in hydrological model structure is illustrated by comparing low-flow sequences derived from CATCHMOD with a much simpler statistical model. The multiple linear regression model (REGMOD) estimates annual Q95 from effective rainfall in summer (α1) and spring (α2) (Table 5). Despite its simplicity the REGMOD fit to observed low flows was comparable to CATCHMOD (Figure 4), and the two model structures were weighted according to their adjusted correlation coefficients, respectively radj = 0.81 (REGMOD) and radj = 0.88 (CATCHMOD). As with CATCHMOD, uncertainty in REGMOD parameters was also explored by randomly sampling different coefficients given their mean and standard error (Table 5).

Details are in the caption following the image
Comparison of observed, CATCHMOD, and REGMOD annual low-flow (Q95) series for the River Thames.
Table 5. Summary of REGMOD Parametersa
Model radj = 0.81 Coefficients
α0 α1 α2
Mean 35.70 0.0563 0.0432
SE 1.360 0.0095 0.0127
p <0.001 <0.001 0.002
  • a The independent variables are effective rainfall in summer (α1) and spring (α2). The dependent variable is the annual Q95.

3.3. Monte Carlo Simulations

[25] The outcome of any Monte Carlo analysis is conditional on the weighting scheme. In the present case weights were assigned on a component-wise basis, but it is acknowledged that there are other ways of determining weights. For instance, integrated systems of GCM/downscaling/hydrologic model combinations could have been weighted using their ability to reproduce the low-flow metric (Q95). However, this approach has the disadvantage of concealing the influence of individual components on the final CDF. Therefore the former uncertainty assessment framework was applied.

[26] Two experiments were undertaken. First, a fully conditional (COND) experiment in which weights were prescribed within each component of uncertainty. Second a fully unconditional (UNCON) experiment in which no weighting was applied within any component. In both cases Monte Carlo analysis was performed using 2000 runs. For each run, the sampled emission scenario, GCM and downscaling method was first used to select a PPT and PET scenario. Next the climate scenario was used to generate an annual series of the Q95 metric for 1961–2100 given a sampled hydrological model structure and parameter set. Finally the mean Q95 was calculated from the transient series for standard time slices (2020s, 2050s, 2080s) and compared with the 1961–1990 baseline to estimate percent changes. The resulting changes in low flows were then presented as CDFs.

4. Results

[27] The outcomes of the analysis are reported in two sections. The first compares changes in daily PPT downscaled from different forcing scenarios. The second compares CDFs of changing low flows linked to different components of uncertainty.

4.1. Climate Change Scenarios

[28] Fully transient daily PPT and PET totals were produced for the Thames basin by SDSM using CGCM2, CSIRO, ECHAM4 and HadCM3 outputs for the A2 and B2 emission scenarios over the period 1961–2100. Scenarios were also constructed for the 2020s, 2050s and 2080s time slices using the CF method.

4.1.1. Precipitation

[29] All SDSM scenarios show significant increases in winter precipitation ranging between +22% and +80% by the 2080s (Table 6a). Large increases in maximum wet day totals and variance projected by all four GCMs in winter point to increased flood risk. The CF winter scenarios are also in agreement about the direction of change but the range is narrower, just +3% to +27% by the 2080s (Table 6b).

Table 6a. Changes in Seasonal PPT Totals with Respect to 1961–1990a
A2 B2 A2 B2 A2 B2 A2 B2
2020s +9 +24 +23 +14 +33 +31 +1 +2
2050s +45 +13 +55 +33 +50 +47 +11 +18
2080s +69 +41 +80 +42 +80 +88 +30 +22
2020s +18 +26 −9 −4 +10 +1 −7 +2
2050s +30 +21 −10 +17 0 −3 −13 −18
2080s +56 +40 +20 +35 −10 −3 −37 −20
  • a All results are based on means of 20-member ensembles downscaled by SDSM from each GCM. Values are in percent.
Table 6b. As Table 6a but for the CF Method
A2 B2 A2 B2 A2 B2 A2 B2
2020s −6 +5 +5 +3 +3 +19 +7 +2
2050s +3 −8 +16 +5 +17 +13 +13 +18
2080s +5 +3 +26 +17 +19 +26 +27 +25
2020s +8 +8 −6 −5 −17 −9 −6 −5
2050s −2 +3 −7 −1 −33 −24 −26 −22
2080s +1 −7 +9 +7 −52 −26 −50 −36

[30] However, SDSM changes in summer are split between increases of up to +56% projected by CGCM2 and CSIRO, and reductions of up to −37% by HadCM3 and ECHAM4. This division is evident in the CF scenarios, which show changes of +9% by CSIRO, and reductions of around −50% by HadCM3 and ECHAM4. The contrasting behavior of the two groups of GCMs is further confirmed by changes in subseasonal precipitation diagnostics. For example, scenarios downscaled by SDSM from HadCM3 and ECHAM4 show large reductions in the frequency of wet days and maximum 5-day totals in summer by the 2050s and 2080s.

[31] Inter-GCM uncertainty in winter and summer precipitation has been noted previously [Jenkins and Lowe, 2003]. Same sign variations between the precipitation scenarios downscaled from the CF and SDSM methods may be partly explained by differences in spatial scales. Whereas the former employs grid box scale information covering a domain of ∼90,000 km2, the latter downscales to the river basin scale, in this case ∼10% of the GCM grid box area (see Figure 1). Differences in the direction of change (as in the case of summer rainfall under the CGCM2 B2 scenario) are more difficult to interpret and will be discussed later.

4.1.2. Potential Evapotranspiration

[32] Changes in PET projected by SDSM for the 2020s, 2050s and 2080s were also used by the CF method because the constituent daily variables of the Penman-Monteith equation were not available for all GCMs. Changes in PET shown in Table 7 are therefore identical for the SDSM and CF scenarios. All four GCMs project increases in winter PET totals ranging from +5% by HadCM3 to +43% by CSIRO in the 2080s (Table 7), noting that these increases are being applied to small baseline PET rates. Percent changes in summer PET are more modest, spanning +11% in HadCM3 and CGCM2 to +22% in CSIRO by the 2080s. Overall, inter-GCM uncertainty in seasonal changes in PET is small compared with those for precipitation, although larger differences emerge at monthly and daily timescales. For example, HadCM3 suggests higher maximum 15-day PET totals in September and October than other GCMs.

Table 7. As in Table 6a but for Seasonal PET Totals
A2 B2 A2 B2 A2 B2 A2 B2
2020s +10 +15 +17 +21 +9 +11 +3 +3
2050s +23 +19 +25 +33 +11 +13 +8 +5
2080s +31 +25 +43 +42 +14 +12 +12 +5
2020s +4 +6 +7 +10 +5 +5 +3 +2
2050s +11 +7 +14 +15 +11 +9 +7 +7
2080s +13 +11 +22 +20 +18 +13 +16 +11

4.2. Monte Carlo Analysis

[33] The first experiment (COND) yields CDFs of projected changes in low flows conditional on the choice of GCM and hydrological model, recalling that no attempt was made to weight uncertainties in the emissions scenario or downscaling method. Results from the Monte Carlo analysis were stratified by each component to assess their relative significance.

[34] Stratifying the COND results into A2 and B2 subsamples highlights uncertainty in low flows due to the emissions scenario (Figure 5a). Although the uncertainty range for the projected low flows increases with time, uncertainty due to emissions is still minor even by the 2080s. Under A2 emissions there is an 83% likelihood of reduced low flows (i.e., a negative change) compared with 82% under B2 emissions. However, there are larger differences between the two emission scenarios for more extreme low-flow reductions (Figure 5a).

Details are in the caption following the image
CDFs of changes in low flows (Q95) by the 2080s reflecting uncertainty in (a) the emission scenario only, (b) the GCM only, (c) the downscaling technique only, (d) the hydrological model structure only, and (e) the hydrological model parameters (under the A2, HadCM3, CF scenario).

[35] In comparison, stratification by GCM reveals far greater uncertainty (Figure 5b). The likelihood of lower summer minimum flows than baseline conditions ranges from 47% under CGCM2 to 100% under HadCM3 and ECHAM4. The form of the CDF is also very different for CGCM2, reflecting the large projected increases in summer rainfall for this GCM (Table 6a). Stratification by downscaling technique shows large differences between the CF method and SDSM (Figure 5c), but the uncertainty due to choice of empirical downscaling method was not as great as for GCMs. The likelihood of flow reductions now ranges between 66% and 100%.

[36] Stratification by hydrological model structure reveals systematic differences between CATCHMOD and REGMOD (Figure 5d). The likelihood of lower flows ranges from 72% to 92% respectively. The more conservative response of CATCHMOD may reflect the longer “memory” of wet winter conditions due to the soil moisture accounting that is unseen by REGMOD. Uncertainty due to hydrological model parameters was explored for both models given the CF scenario for HadCM3 downscaled from A2 emissions (Figure 5e). Under this dry scenario (Table 6b) projected changes in Q95 vary between −10% to −22% for CATCHMOD, and between −15% to −34% for REGMOD, solely due to the choice of parameter set. If this outcome is representative, it is evident that a greater range of parameter uncertainty has been applied throughout the analysis for REGMOD. The results are also consistent with previous analyses showing that uncertainties in river flow projections due to hydrological model parameter uncertainty can be significant [Wilby, 2005]. This component of uncertainty was found to be robust to the choice of model skill criterion; replacing Nash-Sutcliffe with a measure of absolute mean error had a negligible affect on river flow projections.

[37] In the second experiment (UNCON), uncertainties arising from the choice of emission scenario, GCM, downscaling method, etc. are handled naively. All models and emissions are assigned equal likelihood in the Monte Carlo analysis and the results are compared with COND experiment (Figure 6). As expected, the uncertainty in projected changes in Q95 increases with time from −19% to +74% in the 2020s, through to −35% to +79% by the 2080s for UNCON (not shown). The majority of simulations point to decreased low flows in each time slice. The step change toward increased summer low flows was, as before, linked to the very wet scenarios downscaled by SDSM from CGCM2 under A2 emissions (Table 6a).

Details are in the caption following the image
CDFs of changes in low flows (Q95) by the 2080s for unconditional and conditional experiments.

[38] Despite the added sophistication of COND experiment compared to UNCON, the resulting CDFs are remarkably similar (Figure 6). This reflects the relatively small sample of GCMs included in the analysis, and the fact that some components were unweighted in both experiments. Hence the influence of individual GCMs is still evidenced by abrupt changes in CDFs, even though inclusion of different emissions, downscaling and hydrological model combinations produce a range of outcomes for any given GCM. Greater differences would be expected for the COND and UNCON experiments for changes in flood risk because uncertainty in hydrological model parameterization has a greater influence on high flows than low flows [Wilby, 2005].

[39] In summary, an exploration of uncertainty in future low flows for the River Thames suggests the following order of component significance (greatest to least): GCM > (empirical) downscaling method > hydrological model structure > hydrological model parameters > emission scenario. On the basis of a very limited set of GCMs, the likelihood of a reduction in summer low flows by the 2080s is 76% for the conditional experiment, and 82% for the unweighted treatment of uncertainty.

5. Discussion

[40] The climate change impacts community has traditionally viewed uncertainty in a rather narrow sense, taking into account contributions from individual sources (such as the choice of forcing scenario and/or GCM) but rarely quantifying other uncertainties within a probabilistic framework. However, the emergence of multimodel and multiensemble experiments has paved the way for differential weighting of uncertainties based on reliability criteria and raises the prospect of probabilistic climate change information for impact assessment [e.g., Giorgi and Mearns, 2003; Murphy et al., 2004]. This paper has used changing low flows in the River Thames basin to illustrate a framework for exploring different components of uncertainty such as emissions scenario, GCM, downscaling method and (hydrological) impact model.

[41] The suite of uncertainties considered herein was far from exhaustive. For example, the effects of empirical versus dynamical downscaling uncertainty were not considered [Wilby et al., 2000; Wood et al., 2004], and only a single model run was used for each GCM. The sample of emission scenarios and GCM runs was constrained by availability of daily output needed for the SDSM method: just two pathways and four models. The GCMs represent a relatively narrow range of climate sensitivities (3.1–3.6°C) compared with the larger ensemble of 15 GCMs (2.0°–5.1°C) of Cubasch et al. [2001], and the likelihood-weighted range for the multiensemble (2.4°–5.4°C) reported by Murphy et al. [2004]. Nonetheless, end members of the low-flow CDFs originate from GCMs with greatest (CGCM2) and least (HadCM3 and ECHAM4) climate sensitivity (see Table 1). Had the outcomes been independent of GCM climate sensitivity there would have been less justification for weighting by reliability for the present climate. From a practical perspective, applying the less data intensive CF method enables a broader suite of emission scenarios and GCMs to be considered [e.g., Arnell, 2004]. However, such methods do not deliver changes in the temporal sequencing of weather events that are often of consequence to precipitation sensitive impacts [Diaz-Nieto and Wilby, 2005].

[42] A further caveat of the study was the evaluation of GCM reliability by downscaling from grid boxes overlying the target basin. This was justified given that SDSM employs information from a wider domain (e.g., airflow indices such as VSUR are computed from a nine grid box array). Furthermore, GCM/downscaling was used to assess skill from the perspective of impact relevance which is highly local, as opposed to the conventional approach of evaluating global fields [e.g., Murphy et al., 2004]. Because of its significance to low flows in the Thames, summer effective rainfall was used to weight the performance of the GCM/downscaling.

[43] At this length scale there are large differences among GCMs in the behavior of individual predictor variables. For example, HadCM3 exhibits relatively modest increases in summer SHUM compared with the other GCMs (Figure 7). As a key determinant of summer precipitation, this disparity accounts for the wide divergence in scenarios produced by HadCM3 and CGCM2 in the 2050s and 2080s. Differences in atmospheric moisture are also thought to explain variance between the summer precipitation scenarios of the SDSM and CF methods. The former is sensitive to the quality of the downscaling predictor variables; the latter reflects the realism of precipitation schemes and behavior of atmospheric boundary conditions in the GCM. It may be no coincidence that the two GCMs yielding least summer drying also exhibit the largest biases in relative humidity (RHUM) for the baseline period.

Details are in the caption following the image
Annual series of (top) SHUM and (bottom) RHUM in summer.

[44] The preceding discussion underlines some of the dangers arising from climate change impact assessments based on a single GCM and/or impact model. For example, applying either HadCM3 or CGCM2 scenarios in isolation would yield contrasting river flow scenarios for the Thames basin. Hence the Environment Agency of England and Wales is working with partners in the water industry and government to assess multiple uncertainties in the context of water resource planning and flood risk management [e.g., Reynard et al., 2004]. The forthcoming UK Climate Impacts Programme (UKCIPnext) scenarios will provide new information on extreme events and probabilistic aspects of climate change at regional scales. Therefore new frameworks for handling uncertainty at all levels are urgently needed. The proposed framework shows how some components of uncertainty could be objectively weighted, leading to conditional probabilities for climate change impact assessments - in this case, low flows in the River Thames. Further work will be needed to develop practical guidance for planners and engineers who have the difficult task of translating probabilities into adaptation responses.


[45] The views contained in this paper reflect those of the authors and are not necessarily indicative of the position held by the Environment Agency. The authors appreciate the constructive comments of Mark New and Christoph Frei as well as the excellent reviews of three anonymous referees. The work was supported in part by STARDEX (Statistical and Regional Dynamical downscaling of Extremes for European regions) under the European Community Research Programme (EVK2-CT-20010015) and by Environment Agency Science Research Project X1-045/2B.