Variability in simulated recharge using different GCMs



[1] Variations in the prediction of recharge is addressed by comparing recharge simulated using climate data generated using a state-of-the-art downscaling method, TreeGen, with a range of global climate models (GCMs). The study site is the transnational Abbotsford-Sumas aquifer in coastal British Columbia, Canada and Washington State, USA, and is representative of a wet coastal climate. Sixty-four recharge zones were defined based on combinations of classed soil permeability, vadose zone permeability, and unsaturated zone depth (or depth to water table) mapped in the study area. One-dimensional recharge simulations were conducted for each recharge zone using the HELP hydrologic model, which simulates percolation through a vertical column. The HELP model is driven by mean daily temperature, daily precipitation, and daily solar radiation. For the historical recharge simulations, the climate data series was generated using the LARS-WG stochastic weather generator. Historical recharge was compared to recharge simulated using climate data series derived from the TreeGen downscaling model for three future time periods: 2020s (2010–2039), 2050s (2040–2069), and 2080s (2070–2099) for each of four GCMs (CGCM3.1, ECHAM5, PCM1, and CM2.1). Recharge results are compared on an annual basis for the entire aquifer area. Both increases and decreases relative to historical recharge are simulated depending on time period and model. By the 2080s, the range of model predictions spans −10.5% to +23.2% relative to historical recharge. This variability in recharge predictions suggests that the seasonal performance of the downscaling tool is important and that a range of GCMs should be considered for water management planning.

1. Introduction

[2] In recent years, there have been a growing number of modeling studies that have attempted to use climate change predictions from global climate models (GCMs) or regional climate models (RCMs) as input to groundwater models. Most of these studies have focused on simulating the shifts in direct or diffuse recharge (i.e., from precipitation) that may affect groundwater levels under future climate conditions [e.g., Yusoff et al., 2002; Scibek and Allen, 2006a; Jyrkama and Sykes, 2007; Toews and Allen, 2009a]. A few have also considered shifts in streamflow as it interacts with groundwater [e.g., Rosenberg et al., 1999; van Roosmalen et al., 2007; Scibek et al., 2007], and a smaller number have considered how changes in land use and climate can affect groundwater [e.g., Vaccaro, 1992; Holman, 2006; van Roosmalen et al., 2009].

[3] Determining shifts in recharge is of particular importance to groundwater studies because recharge is highly dependent on climate variables, which are anticipated to change under future climate conditions. The complexity of approaches for obtaining the climate data series needed as input to the recharge models appears to have increased over time, ranging from the use of global averages [e.g., Zektser and Loaiciga, 1993; Loaiciga et al., 1999], to the use of regional bulk projections or shift factors derived from GCMs [e.g., Vaccaro, 1992; Yusoff et al., 2002; Allen et al., 2004; Jyrkama and Sykes, 2007], to the use of shift factors derived from downscaled GCM data [Scibek and Allen, 2006a; Serrat-Capdevila et al., 2007; Toews and Allen, 2009b], to the use of regional climate models [van Roosmalen et al., 2007; Rivard et al., 2008]. In some of these studies, shift factors derived from downscaled GCM data or raw GCM data have been applied to a stochastic weather generator to generate a weather series for input to the recharge model [e.g., Scibek and Allen, 2006a; Toews and Allen, 2009b].

[4] While a number of these studies have considered a range of GCMs or the average projection from several GCMs, only a few have considered a range of emission scenarios or downscaled GCM data using statistical or dynamic methods. This is likely because there are quite a number of GCMs (roughly 25 globally), a range of emission scenarios [IPCC, 2007], and a growing number of downscaling methods to choose from [e.g., Wilby et al., 2002; Zorita and von Storch, 1999; Yates et al., 2003; Schnur and Lettenmaier, 1998; Stahl et al., 2008; Cannon, 2008].

[5] The purpose of this study is to illustrate the range in recharge simulated using climate data that are downscaled using a state-of-the art method for four GCMs. The motivation for this study was the uncertainty associated with the downscaling results obtained previously for the study area, which were used to simulate recharge for input to a groundwater flow model [Scibek and Allen, 2006b]. In that study, downscaled results from a single GCM, namely, the Canadian Global Coupled Model 1 (CGCM1) [Flato et al., 2000], were used to derive shift factors for future climate periods. These shift factors were used to perturb a stochastic weather generator, LARS-WG [Semenov et al., 1998], and the resultant climate series was used to force a recharge model. The resulting recharge was then used in a three-dimensional groundwater flow model to evaluate changes in groundwater level for current and future climate periods.

[6] Here, the TreeGen downscaling model [Stahl et al., 2008; Cannon, 2008] is used with a range of GCMs to generate climate input data for the same recharge model used by Scibek and Allen [2006b]. The objective of the paper is to illustrate the range of results for the same study area, drawing attention to the issue of uncertainty in recharge predictions arising as a result of choice of GCM. Comments are provided on the implications of incorporating uncertainty into water management planning.

2. Background

[7] The Abbotsford-Sumas aquifer is approximately 161 km2 in aerial extent and is bisected by the international boundary between British Columbia and Washington State (Figure 1). The aquifer is situated within the Fraser and Nooksack River Lowlands in the central and eastern Fraser Valley in southwest British Columbia (BC), Canada and northern Washington State (WA), USA. The aquifer is highly productive and provides water supply for nearly 10,000 people in the US (towns of Sumas, Lynden, and farmlands) and 100,000 in Canada, mostly in the City of Abbotsford, but also in the township of Langley [Mitchell et al., 2003]. This surficial aquifer is composed of uncompacted sands and gravels of the Sumas Drift, a glacial outwash deposit. The thickness of Sumas Drift can be up to 65 m and is thickest in the northeast, where glacial terminal moraine deposits are found. The aquifer is mostly unconfined and forms broad outwash plain, which is elevated above the adjacent river floodplains. The outline of the aquifer proper is shown in Figure 1. The outwash terrace slopes southward and terminates in escarpments along the Nooksack River floodplain. The uplands are centered on the City of Abbotsford, BC, and extend westward through Langley, BC, and south to Lynden, WA.

Figure 1.

Map showing the location of the Abbotsford-Sumas aquifer in southwestern British Columbia (BC) and northwestern Washington State (WA) (see inset box for study location in North America). Shown also are the locations of major towns and the surface drainage. The black box surrounding the aquifer shows the map area for Figure 2.

[8] To the west, centering around Langley (Figure 1), there is an extensive glaciomarine unit composed of stony clays that is referred to as the Fort Langley Formation. In this area, the Abbotsford-Sumas aquifer transitions to an aquitard, although pockets of high-permeability sediments are found (Figure 2). The Sumas Valley, extending northeast of Sumas (Figure 1) and containing the Sumas River, is a large, sediment-filled deep bedrock valley. The aquifer is thought to dip beneath this valley, becoming confined at depth (Figure 2). More recent fluvial and lacustrine sediments underlie streambeds and lakes (Figure 2).

Figure 2.

Generalized surficial geology of study area. Map extent shown as box in Figure 1.

[9] The climate in this region derives from its coastal setting and is characterized as humid and temperate, with 1000 to 1500 mm mean annual rainfall over most of the year. Table 1 shows the climate normals (1971–2000) for the Abbotsford International Airport climate station. Recharge to the aquifer is primarily from direct precipitation, mostly as rain, which falls from October to May [Scibek and Allen, 2006b]. Recharge is expected to be spatially variable due to differences in soil cover, surficial materials (as described above) and local topography.

Table 1. Climate Normals (1971–2000) for Abbotsford International Airporta
Mean temperature2.
Standard deviation2.
Max. Temperature (°C)5.88.511.314.517.820.323.423.821159.15.914.7
Min. Temperature (°C)−−0.35.3
Rainfall (mm)17414814212099795049761452341911508
Snowfall (mm w.eq.)23134000000061764
Precipitation (mm)19716114612099795049761452402081572

[10] Regional groundwater flow is generally from north to south (from upland areas to low-lying areas) following major topography, such that the groundwatershed area is approximately coincident with the outer extent of the surface watersheds for the three streams that drain southward to the Nooksack River (see Figure 1). Some drainage occurs to the east toward the Sumas River. To the north is Fraser River floodplain, which drains the watersheds lying to the north of the study area. Thus, at a regional scale, the Nooksak, Sumas, and Fraser Rivers act as major groundwater discharge zones. Locally, groundwater discharges to streams, and springs and seeps are observed at the edges of the raised drift deposit. Locally, groundwater flow is influenced by the complexity of the geology [Scibek and Allen, 2006b].

3. Methodology

3.1. Historical Climate Series

[11] A climate data series is needed as input to the recharge model. Input data include mean daily temperature, total daily precipitation, and daily solar radiation. Scibek and Allen [2006b] generated historical daily weather using the LARS-WG stochastic weather generator [Racsko et al., 1991]. The resulting weather series was used to generate historical recharge for the aquifer; this historical recharge is used here as the base case for comparison with recharge modeled using downscaled GCM data.

[12] LARS-WG utilizes semi-empirical distributions for the lengths of wet and dry day series, daily precipitation, and daily solar radiation. LARS-WG has been shown to compare favorably to other weather generators [Wilks and Wilby, 1999]. For inputs to the stochastic weather generator, the observed mean daily precipitation at the Abbotsford station was converted to mean monthly precipitation and then converted to mean monthly precipitation for all days in the month by multiplying by the percentage of wet days in a month. Average monthly and average daily solar radiation at Abbotsford were modeled using the cloudless-sky irradiance formulations as described by Davies and McKay [1982] and corrected using observed hourly cloud opacity data (daily mean) at Abbotsford for the period of record 1975–1995. Details concerning this approach can be found in Scibek [2005]. Three hundred years of synthetic daily weather were generated. It is noted that although generated weather runs of 1000 years converge better to specified “normals,” there were diminishing returns of performance after 100 years. The length of generated weather time series is not meant to model actual changing climate year to year, but rather to model climate change stepwise for each time period and to generate a long enough weather time series to preserve and properly represent statistical properties for the site and the specified climate (e.g., seasonality of precipitation and its variability).

[13] LARS-WG was calibrated against observed daily precipitation, maximum and minimum temperature, and modeled solar radiation (1975–1995 period). The calibration process uses conditional relationships among the observed parameters. For example, realistic temperature and solar radiation values are generated depending on whether the day is rainy or dry. Calibration performance of LARS-WG for mean, maximum, and minimum temperature was excellent, as was daily minimum and maximum temperature variation (standard deviation) (Figure 3a). In winter months, LARS-WG produced 0.5 to 1.0°C cooler minimum temperatures than observed. Mean monthly rainfall was within 2 mm/month (5%) of observed (40–50 mm/month) for all months (Figure 3b). Seasonal variation in rainfall showed a very good fit to observed rainfall normals, and variability of rainfall (standard deviation of monthly precipitation) was preserved in synthetic weather, but there are small discrepancies between modeled and observed precipitation in May–July and November (not shown). A χ2 test was used to compare observed probability distributions versus simulated histograms for seasonal wet/dry precipitation series. χ2 values were geneally low (below 3), and P values close to 1.0, suggesting a very good fit between observed and simulated precipitation series. Precipitation distribution (histograms by month) also demonstrated a very good fit (P value close to 1.0). Monthly and daily solar radiation were reproduced to within 5% of observed values (Figure 3c), although values were underpredicted. This underprediction in variability might cause small errors in evapotranspiration estimates.

Figure 3.

Monthly mean air temperature and monthly precipitation observed at Abbotsford International Airport over the period of record 1975–1995. Corresponding LARS-WG model results shown here for a 300-year run. (a) Minimum and maximum temperature, averaged monthly from daily temperature data. (b) Precipitation as mean monthly precipitation. (c) Average monthly and average daily solar radiation (represented as mean daily values) at Abbotsford, modeled using a clear sky radiation algorithm corrected for observed hourly cloud opacity data (daily mean) for period of record 1975–1995 and compared to LARS-WG output.

3.2. TreeGen Downscaling

[14] A state-of-the-art downscaling method, TreeGen [Stahl et al., 2008; Cannon, 2008], which is conceptually similar to other nonparametric downscaling algorithms [see, e.g., Buishand and Brandsma, 2001; Yates et al., 2003; Gangopadhyay et al., 2005; Apipattanavis et al., 2007], was used to downscale daily mean temperatures and precipitation totals at the Abbotsford International Airport climate station based on outputs from four GCMs: CGCM3.1 (Meteorological Service of Canada) [Environment Canada, 2008]; PCM1 (U.S. Department of Energy) [Washington et al., 2000]; ECHAM5 (E. Roeckner, et al., The atmospheric general circulation model ECHAM 5. Part I: Model description, Max Plancke Institute for Meteorology, available at; and CM2.1 (U.S. Geophysical Fluid Dynamics Laboratory (GFDL) [Delworth et al., 2006]. Both the A2 and B1 series emissions scenarios were downscaled; however, only the A2 series emission scenario was used. The A2 scenario has higher emissions, higher CO2 concentrations, and higher temperatures by the end of the century compared to the B1 scenario [Nakićenović and Swart, 2000].

[15] The TreeGen algorithm consists of two main parts, model fitting GCM and downscaling, which can be further divided into seven steps. Fitting of the TreeGen component models is conducted first (steps 1–4). Once the component models have been developed, they are used to downscale the GCM scenarios (steps 5–7). Model fitting involves (1) common principal component analysis (PCA) of observed/GCM predictor fields for a baseline historical time period [Imbert and Benestad, 2005]; (2) synoptic map type classification of the common principal components (PCs) via a multivariate regression tree (MRT) [Cannon et al., 2002a, 2002b]; and (3) developing linear regression equations, stratified by map type, for the temperature and precipitation series [Imbert and Benestad, 2005]. Following model fitting, downscaled estimates of surface weather elements are produced by (4) estimating future map types by entering the common PCs into the fitted MRT; (5) applying a nonparametric weather generator based on conditional resampling of temperature and precipitation data from each map type [Buishand and Brandsma, 2001]; (6) entering the common PCs into the linear regression equations [Imbert and Benestad, 2005] and estimating trends with time for the resulting predictions of temperature and precipitation; and (7) adding these within-type trends onto the outputs from the nonparametric weather generator. A conceptual diagram of the steps is shown in Figure 4. Each step is described in turn below.

Figure 4.

Conceptual flowchart of the TreeGen model. A description of each step is provided in section 3.

[16] Step 1: To mitigate potential biases between the NCEP/NCAR Reanalysis [Kalnay et al., 1996] and GCM simulated predictors, common PCA is applied to the two data sets [Imbert and Benestad, 2005]. First, predictors from the reanalysis and GCM, in this case large-scale sea-level pressure, surface temperature, and precipitation fields, are standardized so that the time series for each grid point has zero mean and unit variance during a common 1961–2000 baseline period. Second, standardized data from the reanalysis and the GCM are concatenated to form a single data matrix. Third, PCA is applied to the correlation matrix of the concatenated predictors. Finally, common PC scores from the GCM are rescaled so that their means and variances in the simulated baseline period match observed values from the same period. To ensure consistency of the simulated seasonal cycle, rescaling is performed on each month separately.

[17] Step 2: Following the common PCA, synoptic map types are defined using a multivariate regression tree (MRT) model that recursively split the observed data into groups on the basis of thresholds in the common PC scores [Cannon et al., 2002a, 2002b]. Values of the thresholds are optimized so that the associated daily temperature and precipitation observations at Abbotsford are placed into groups (or map types) that are as homogeneous as possible.

[18] Step 3: As the nonparametric weather generator in step 6 samples cases from the historical data set, future trends in surface climate conditions are due exclusively to changes in the frequency and timing of the synoptic map types simulated by the GCM. Additional processing is thus needed to generate values above/below the highest/lowest records in the historical data set and to accurately reflect trends occurring within map types. To capture both between- and within-type trends, a modified version of the linear regression-based extrapolation algorithm described by Imbert and Benestad [2005] is adopted. For each map type, multiple linear regression (MLR) equations linking the common PC scores and the surface weather elements are created via stepwise regression on the basis of the Bayesian Information Criterion [Schwarz, 1978]. For precipitation, separate models are built for occurrence of precipitation (via regression estimation of event probabilities [REEP]) and log-transformed precipitation amounts on wet days.

[19] Step 4: Once synoptic map types have been defined from the historical record, common PC scores from the GCM scenarios are entered into the MRT, resulting in each day being classified into one of the map types.

[20] Step 5: Next, surface weather conditions on a given day are predicted using a nonparametric weather generator on the basis of conditional resampling from cases assigned to that day's map type [Buishand and Brandsma, 2001]. The probability p(i) of randomly selecting the temperature and precipitation observed on day i as the predicted values on day t is taken to be inversely proportional to the square of the Euclidean distance d(t − 1, i − 1) between the predicted values on the previous day t − 1 and historical values of the weather elements on day i − 1 plus a small offset:

equation image

where i is the set of historical days assigned to the predicted map type occurring on day t. The offset is included to prevent division by zero when the predicted and historical values are the same.

[21] Step 6: Common PCs for the GCM scenarios are entered into the fitted MLR and REEP equations from step 3. This results in a series of predicted temperatures and precipitation amounts.

[22] Step 7: Finally, linear trends in the predicted temperatures and precipitation amounts from the previous step are calculated, and these trends are superimposed onto the time series derived from the nonparametric weather generator from step 4.

3.3. Downscaling Results

[23] Downscaling results obtained using the TreeGen model compare reasonably well to the observed data for each of the GCMs considered (Figures 5 and 6), although summer temperatures (July and August) are underestimated by all four downscaling models and are slightly overestimated in January, September, and December. Downscaling results for the historical period, based on cross-validation using the NCEP/NCAR Reanalysis data set as inputs (not shown), suggests that the over- and underestimation is due partly to uncorrected biases in the GCM circulation fields and partly to the small number of map types (10) selected by the TreeGen algorithm. Choosing a larger but less interpretable number of map types resulted in smaller biases. The improvement was, however, smaller than the overall range in bias over the four GCM scenarios.

Figure 5.

Mean monthly temperature at Abbotsford International Airport, BC, observed and downscaled from various GCM model runs (CGCM3.1, CM2.1, ECHAM5, and PCM1) for current and future climate scenarios using TreeGen downscaling model.

Figure 6.

Mean monthly precipitation at Abbotsford International Airport, BC, observed and downscaled from various GCM model runs (CGCM3.1, CM2.1, ECHAM5, and PCM1) for current and future climate scenarios using TreeGen downscaling model.

[24] The ECHAM5 model predicts the smallest increases in mean monthly temperature over the next century, CGCM3.1 and CM2.1 predict moderate increases, and the PCM1 model predicts the largest increases in temperature.

[25] All models appear to slightly underestimate precipitation in the summer months, particularly CM2.1, and all four models underestimate precipitation in November and December. Overall, the downscaling results are encouraging, but the seasonal mismatch between observed and downscaled temperature and precipitation may have implications for recharge simulation as discussed later.

3.4. Recharge Modeling

[26] Direct recharge via precipitation was modeled for both the historical period and for each future time period using the generated weather series as input to a hydrologic (recharge) model. Spatially distributed recharge estimates were generated using a unique combination of physical parameters that influence recharge. These included soil (and its depth layering), vegetation, slope, and water table depth. Details concerning the recharge modeling approach are given below.

[27] The HELP model [Schroeder et al., 1994] was used to simulate recharge. HELP simulates surface and near-surface hydrologic processes critical for estimating recharge, including accumulation of solid precipitation (snow and ice) on the surface, surface runoff/infiltration, estimated and potential evapotranspiration, transpiration in relation to the growth and decay of vegetation, soil freeze/thaw from air temperature, and subsurface water flow through discrete layers of variably saturated soil. The model uses vertical (one-dimensional) soil profiles and simulates the leakage at the base of the profile. If the base of the soil column is set equal to the water table depth, the leakage across this boundary is effectively the groundwater recharge. HELP has been used in many groundwater recharge studies [e.g., Gogolev, 2002; Allen et al., 2004; Jyrkama and Sykes, 2007; Scibek and Allen, 2006a, 2006b; Toews and Allen, 2009b].

[28] The main limitation of HELP is that it employs a storage-routing unsaturated flow process. HELP uses the following critical assumptions: (1) water may only escape upward (as evapotranspiration) if it is within the evaporative depth zone, which is specified by the user to coincide with the rooting depth; and (2) water drained from the base of the evaporative depth zone will eventually be routed to the base of the model. A limitation of using a 1-D recharge model is that surface (and subsurface, if important) water routing between adjacent grid cells is not accounted for by the model. Thus, this excess water is not routed horizontally to adjacent cells. For this reason, use of a 1-D recharge model is not particularly well suited to areas with steep topography. However, in this particular study area, the aquifer surface is generally flat, and the soils are well drained, so that there will be little surface runoff and little lateral flow within the vadose zone, only vertical flow.

[29] Despite these limitations, the HELP model was selected for several reasons: (1) its simplicity and speed, a soil profile with 200 years of climate data can be simulated within seconds, rather than hours; (2) it utilizes daily climate data; (3) it simultaneously models multiple hydrologic processes, including soil freeze/thaw, and (4) it was used originally by Scibek and Allen [2006b] to simulate the historical recharge for this study area and therefore facilitated a comparison of recharge results for different climate data sets. Scanlon et al. [2002] compared HELP to several similar hydrologic codes and rated HELP poorer than others on the basis of a comparison with field data collected in a semi-arid area, due to an underestimation of evapotranspiration in this water-limited region. The Abbotsford-Sumas aquifer, however, is situated in a wet climate region, and HELP is thought to provide reasonable estimates of recharge [Scibek and Allen, 2006b]. Notwithstanding this, if the absolute values of recharge are in error (too high or too low), relative changes due to different input climate data sets are likely accurately simulated.

[30] As mentioned previously, HELP uses daily mean temperature, total daily precipitation, and total daily solar radiation as weather inputs. In this study, station data from Abbotsford International Airport were assumed to be representative of the area, although a precipitation gradient does exist [Scibek and Allen, 2006b]. Scibek and Allen [2006b] used Abbotsford as an index station and linked the precipitation gradient over the valley to daily precipitation records at Abbotsford, thereby deriving percentage differences in precipitation relative to Abbotsford. They then adjusted all recharge estimates proportionally by the same percentage difference, assuming that recharge is directly proportional to precipitation for any given recharge zone. This correction was not done in this study as the focus was to compare GCM results, but the final results could be corrected in a similar fashion.

[31] Other climate parameters obtained from Abbotsford were also assumed to be constant for the modeling domain and include a wind speed of 9.1 km/hr and quarterly relative humidity of 75%, 69%, 70%, and 79% for 3-month periods starting in January. These parameters are used by HELP to estimate evapotranspiration using a modified Penman equation [Schroeder et al., 1994]. For evapotranspiration calculations, additional parameters were needed, including type of vegetation, wilting point, field capacity, initial moisture content, soil thickness, soil type, porosity and saturated hydraulic conductivity of the vadose zone (Ksat), and vadose zone depth (or water table depth). A sensitivity analysis showed that there was no noticeable or very small (<5% change) effect on recharge when (1) stand of grass type, (2) wilting point, (3) field capacity, and (4) initial moisture content were varied. There was a moderate effect on recharge when (1) soil thickness and (2) the porosity of vadose zone were varied. There was a strong effect on recharge when (1) depth of vadose zone (percolation layer), (2) soil type (permeability), and (3) Ksat of vadose zone were varied.

[32] To keep the simulations relatively simple, only the three most sensitive parameters were varied spatially as discussed below. A uniform value of 4.5 was used for leaf area index (LAI) to reflect a fair stand of grass, the dominant vegetation type throughout the area. Evaporative zone depth was set to 20 cm to reflect this vegetation type and a dominantly sandy soil. Growing season start and end dates were set at 126 and 287, respectively. From 2164 well lithology logs, a histogram showed that soil thickness is generally 0.4 to 1.6 m thick, with a median thickness of 0.92. Therefore, a uniform soil thickness of 1.0 m was assumed for all percolation columns. All percolation columns were constructed with two layers. The top layer represented the soil and the second layer, which extended down to the water table, represented the vadose zone.

[33] The rainfall-runoff processes HELP are modeled using the U.S. Department of Agriculture curve-number method [Soil Conservation Service, 1985], which is widely accepted and allows the user to adjust the runoff calculation to a variety of soil types and land management practices. The curve number (CN) is defined with respect to the runoff retention parameter (S), which is a measure of the maximum retention of rainwater after runoff starts (in inches):

equation image

The maximum value of CN, which is 100, occurs when there is no infiltration. The smaller the CN, the more rainwater will infiltrate the soil. The minimum realistic value for CN can be assumed to be appropriately equal to 50. HELP adjusts the value of CN to account for variations in surface slope, soil texture, and vegetation class. For runoff calculations, zero slope, 100% runoff area, and a fair stand of grass were used for all simulations and soil texture was varied as described below. The topography of the aquifer is slightly undulating or sloping, but over small areas, the surface is approximately horizontal. Steep escarpments are exceptions, but the relative area of these features is very small compared to the aerial aquifer extent.

[34] Because HELP is a 1-D model, spatially distributed recharge estimates require simulations for each unique combination of physical parameters that were shown in the sensitivity analysis to influence recharge. These included soil permeability, vadose zone permeability, and water table depth. Consequently, spatial maps for each of these parameters were created using four classes for each as described below. These spatial maps represent the various parameters for a model domain used to simulate groundwater flow by Scibek and Allen [2006b]. The model domain does not encompass the exact extent of the aquifer, but rather extends beyond the aquifer to the northeast and terminates near the Nooksak River to the south, representing the approximate groundwatershed. The spatial maps discussed below show the outline of the Abbotsford-Sumas aquifer as illustrated in Figure 1, and extent of the various parameters maps used in this study.

3.5. Soil Permeability

[35] Soil maps for the Abbotsford-Sumas aquifer were obtained from published soil maps for the Fraser Valley (BC). Whatcom County soil data were obtained from National SSURGO Data, U.S. Dept of Agriculture, Natural Resources Conservation Service. Since most of these soils are rapidly drained, the unconfined surficial aquifer is directly connected to the ground surface such that rainfall and snowmelt water are expected to rapidly infiltrate and recharge the aquifer. A small portion of the aquifer is occupied by the City of Abbotsford and other smaller communities (Sumas, Lynden, Aldergrove), with associated transportation network and built-up areas. In these areas, a large proportion of the ground surface is paved, compacted, or covered by structures, such that most of the rainfall and snowmelt water is redirected to stormflow network and removed. Infiltration to unconfined aquifer is limited in those areas. Thus, paved areas were assigned zero permeability, regardless of underlying soil types. The various soils throughout the region were first grouped according to the drainage code in the respective soil attribute tables provided with the GIS coverages. The soil map was converted to raster format with 20 m resolution and then reclassified into five soil permeability classes (Table 2). Vertical saturated hydraulic conductivity (Ksoil) values were assigned on the basis of default values in the soil texture database within HELP, but are thought to be reasonably representative (Figure 7).

Figure 7.

Classified distribution of soil drainage for the Abbotsford-Sumas aquifer. Drainage classification was described on the basis of soil permeability (Table 2).

Table 2. Soil Types in HELP Model, Soil Saturated Hydraulic Conductivity, and Permeability Class for Recharge Modelinga
Soil LayerSaturated Ksoil (cm/s)Permeability Class
  • a

    The spatial distribution of permeability classes is shown in Figure 7.

Pavement0Very low
Silty loam1.9 × 10−4Low
Loam3.7 × 10−4 
Fine sandy loam5.2 × 10−4 
Sandy loam7.2 × 10−4Moderate
Loamy fine sand1.0 × 10−3 
Loamy sand1.7 × 10−3High
Sandy gravelly soils5.8 × 10−3Very high

3.6. Vadose Zone Ksat

[36] The saturated hydraulic conductivity, Ksat, of the vadose zone was estimated on the basis of well lithology data. Lithology data for well log were partitioned at the water table so that only those materials in the vadose zone would be accounted for. For each layer, an average value for K was assigned on the basis of the dominant and subdominant material types. To calculate the vertical saturated K value, the equivalent Kz (harmonic mean) was used:

equation image

where mi is the thickness of layer i having equivalent hydraulic conductivity Ki. The averaged units mi are, by default, homogeneous and isotropic as represented by equivalent Ki. There are no data for the aquifer on microscale isotropy.

[37] The Ksat (or vertical Kz as shown in equation 3 above) in the vadose zone were interpolated using inverse distance weighed interpolator (power 2, number of points = 5, output cell size 100 m), and computed on representative vertically averaged log Ksat values at all available point locations where lithologs exist. After interpolation, 10^(log Ksat) of the interpolated raster was computed. Ksat values were then converted to units of meters/day. A histogram of Ksat values for each grid cell (over 1 million pixels) showed a range from 0.1 to 105 m/day, median of 50.91 m/day, mean 46.3 m/day, and quartile values of 0.51 and 89.84 m/day. Four Ksat classes were chosen as 1 × 10−6 to 40 m/day (low), 41 to 60 m/day moderate), 61 to 80 m/day (high), and 81 to 120 m/day (very high). Representative material Ksat in HELP soil columns were according set to 0.5, 51, 75, and 105 m/day (midvalue in each class).

[38] Overall, the Ksat distribution was found to be very heterogeneous (Figure 8). Low Ksat values occur over Fort Langley Formation (stony clays) sediments in the northwestern part of area, and in Sumas Valley in the former location of a lake (lacustrine silts). There are also low Ksat values along river channels where there are mapped silts and other low-K deposits (slack water deposits). Moderate Ksat values occur in Sumas Valley due to floodplain silty sands cover and over southern parts of the aquifer system. High Ksat values are found in Abbotsford City area, in the uplands associated with highly permeable Sumas Drift consisting of gravels and sands, interspersed with till deposits.

Figure 8.

Classified distribution of Ksat of the vadose zone for Abbotsford-Sumas aquifer. Four Ksat classes were chosen as 1 × 10−6 to 40 m/day (low), 41 to 60 m/day (moderate), 61 to 80 m/day (high), and 81 to 120 m/day (very high).

3.7. Depth to Water Table

[39] The depth to water is the distance from the ground surface to the water table. It determines the depth of material through which water must travel before reaching the water table. Depth to water was estimated for wells in the Abbotsford-Sumas aquifer directly from the historical static water levels recorded in drillers' logs. Static water levels provide a one-time measure of the depth of water in the well. Normally, these measurements are made immediately following drilling and therefore can result in lower values that would be measured some time following drilling when the well has reequilibrated with the surrounding aquifer water levels. The Abbotsford-Sumas aquifer is a highly permeable aquifer; consequently, the hydraulic disturbance during drilling activities can be expected to dissipate fairly quickly. In this respect, it is reasonable to assume that postdrilling measurements of water level may be similar to those of the surrounding undisturbed aquifer. In addition to drilling disturbance, water levels vary throughout the year in an aquifer according to seasonal factors (e.g., changes in recharge and changes in storage). Because wells are drilled at different times of the year, the static water elevations recorded following drilling might be expected to vary depending on season. Notwithstanding this, static water level measurements are assumed to be representative of groundwater levels in the aquifer and act as a surrogate for ambient groundwater conditions in the aquifer.

[40] Values of static water level (recorded as depth to water in a well), were imported into the GIS as point values that are representative of the water level at each well. The median depth was 5.0 m, the mean was 8.0 m, and the standard deviation was 9.3 m (the histogram is skewed by many small depths to water table and a few large ones near scarps), but range was from 0 to 78 m. A composite water surface was calculated using a geostatistical analysis involving interpolation between points and extrapolation to the boundary of the aquifer. By subtracting the water table surface from the ground surface (using digital elevation model), a map of depth to water table was produced in 20 m raster format. Depth to water table determines the total thickness of HELP soil column for recharge computation. Five depths were selected using quartiles of the distribution of the depths (min and max bounding values). The depth classes were chosen as 0 to 2 m (class 1), 2.1 to 5 m (class 2), 5.1 to 13.0 m (class 3), 13.1 to 78 m (class 4), with roughly 25% of aquifer area in each category (Figure 9).

Figure 9.

Classified depth to water table (from ground surface) in Abbotsford-Sumas aquifer. The depth classes were chosen as 0 to 2 m (class 1), 2.1 to 5 m (class 2), 5.1 to 13.0 m (class 3), and 13.1 to 78 m (class 4).

3.8. Recharge Zones

[41] The spatial maps for classes of soil permeability, vadose zone Ksat, and depth to water table were combined using a GIS. This resulted in 4 × 4 × 4 = 64 unique recharge zones (Figure 10); the HELP model run for each zone. Recharge was computed for the two-layer percolation columns using daily weather data series corresponding to the stochastically generated historical period (hereafter referred to as the base case) and both the historical and future time periods for each of the four downscaled GCMs. One hundred years of weather data were used as input for all cases, thus allowing for calculation of monthly and long-term mean recharge for each column. Raster calculations were done to compute differences in spatially distributed recharge. Soil moisture was first initiated using the field capacity, but model spin-up was extended by running 200 years of climate data (back-to-back 100-year periods), and only keeping the past 100 years for analysis. A longer spin-up time was needed to initialize soil moisture, as simulations using tall soil columns generally underestimated recharge during the early time series.

Figure 10.

Recharge zones for the Abbotsford-Sumas Aquifer. The recharge model was run for each of the 64 zones.

4. Results

[42] To facilitate a comparison of the recharge results, we first present the spatially distributed recharge for the historical period on the basis of the stochastically generated weather series using LARS-WG. We make an assumption that this base recharge map is representative of observed historical recharge and thus consider it the base case. However, it is noted that measured recharge values, particularly spatial recharge values, are lacking for this aquifer, which was one reason why recharge was simulated by Scibek and Allen [2006b] for input to the groundwater model. The simulated historical recharge produced a well-calibrated (both steady-state and transient) groundwater flow model for appropriate ranges of measured aquifer hydraulic conductivity values [Scibek and Allen, 2006b] lending support for these simulated values. The success of LARS-WG for reproducing the historical climate also justifies selecting this simulation as the most representative of observed historical climate.

[43] Figure 11 shows the spatially distributed mean annual historical recharge to the Abbotsford aquifer (mm/yr). Values range from near <980 to >1040 mm/yr. Mean annual precipitation from LARS-WG is 1613 mm/yr (Table 3) compared to the observed normal (1572 mm/yr) (Table 1). Precipitation is highest from November to January and lowest in July and August (Table 3). Evapotranspiration is highest during the summer months (May to September) and nearly balances precipitation; July has a moisture deficit (Table 3). Runoff is close to zero in all months. Simulated recharge represents roughly 61 to 64% of the generated mean annual precipitation and is highest from November through March–April. Overall, recharge is quite high in this coastal aquifer due to the high permeability of the soil and aquifer materials and high precipitation rates, particularly during months of the year when evapotranspiration rates are at their lowest (fall and winter).

Figure 11.

Mean annual recharge for the Abbotsford study area simulated for the historical or base case. Weather data were generated using the LARS-WG weather generator.

Table 3. Average Monthly Precipitation, Evapotranspiration, Runoff and Recharge Calculated Using HELP for the Base Case Model (historical period)a
 JanFebMarAprMayJunJulAugSepOctNovDecTotal Annual
  • a

    Runoff and recharge are the spatial averages across all recharge zones.

Precipitation (mm)211.0166.4153.4120.799.472.149.054.888.5151.6221.4224.41612.7
Evapotranspiration (mm)17.424.351.782.082.570.458.747.260.133.820.416.6565.1
Runoff (mm)
Recharge (mm)173.3155.4138.699.668.345.234.826.820.922.578.1145.11008.6

[44] For comparison, Chesnaux and Allen [2008] estimated 1280 mm/yr potential recharge to the aquifer over the period 2000–2003, based on the difference between total annual precipitation and evapotranspiration (excluding runoff). However, runoff is very low in this aquifer due to the high permeability of the soils and sediments and is not expected to result in major differences between modeled recharge and potential recharge (see Table 3). Rather, the lower estimate obtained by Chesnaux and Allen [2008] is thought to be the result of lower annual precipitation during these years compared to the normals due to El Niño Southern Oscillation (ENSO) cycles. Nonetheless, the two studies yield values that are of a similar magnitude.

[45] Figure 12 compares the recharge results for the base case as derived from LARS-WG (Figure 11) and each of the TreeGen-downscaled results for CGCM3.1, ECHAM5, CM2.1, and PCM1. The results show roughly 25 to 100 mm less recharge on an average annual basis compared to the base case model. These results are consistent with the downscaling results for precipitation (Figure 6), which showed that TreeGen slightly underestimates November and December precipitation (Figure 6). Monthly recharge is correspondingly lower (data not shown) for these months. These results have important implications with respect to considering the seasonal performance of downscaling algorithms. In this particular study area, the rate of recharge is highest during the fall to early winter months, following the dry summer. Therefore, underestimation of precipitation during the fall to early winter can have significant effects on mean annual recharge as illustrated in Figure 12.

Figure 12.

Maps showing the difference in annual historical recharge for the TreeGen downscaled climate data sets from four GCMs (CGCM3.1, CM2.1, ECHAM5, and PCM1) relative to the historical base case for recharge shown in Figure 11.

[46] Figure 13 shows a comparison of changes of mean annual recharge for the Abbotsford-Sumas aquifer on the basis of downscaled results from the four CGMs relative to the current base case (Figure 11). For example, in the 2050s, a moderate projection time period, the various models predict a range of +9.4% increase in recharge to a −1.6% decrease in recharge. In the 2080s, the various models predict a range of +23.2% increase in recharge to a −1.5% decrease in recharge. Notably, there does not appear to be any consistency from one model to the next with respect to whether it predicts increases or decreases in recharge relative to the base case. For example, CGCM3.1 predicts an increase in recharge for all future time periods, which can be attributed to its increased winter precipitation, particularly during the 2080s (Figure 6). In contrast, PCM1 predicts little change in recharge (although the 2050s time period is slightly positive). CM21 and ECHAM5 predict both increases and decreases, but results are inconsistent. There is an overall tendency, however, for these models to predict positive changes in recharge in the study area.

Figure 13.

A comparison of relative changes in mean annual recharge for the Abbotsford-Sumas aquifer based on downscaled results from four global climate models (CGCM3.1, ECHAM5, CM2.1, and PCM1). The base case to which all others were compared was historical recharge shown in Figure 11. Data for the A2 emission scenario were used.

5. Discussion

[47] Although the results are not shown here, Scibek and Allen [2006b] carried through to a groundwater flow model the recharge predictions based on CGCM1 projections. CGCM1 predicted consistently negative changes in mean annual recharge relative to the base case in the 2020s (average of −5.9%, with a range of −5.6 to −6.3%) and the 2050s (average of −13.4%, with a range of −12.7 to −14.6%). This gradual lowering of recharge resulted in a corresponding lowering of the groundwater level throughout most of the aquifer. In the upland areas, the groundwater levels were predicted to decrease by between −0.05 m to more than −0.25 m by the 2020s. The decrease in groundwater levels was greater in the 2050s; between −0.10 and −0.25 m in most areas. In places with suspected perched water tables, which tended to be areas of poor model calibration, the changes were between −0.5 and −3.0 m. Thus, overall, relatively small declines in groundwater level were predicted, which would likely not cause significant impact to water wells; that is, wells would not have to be deepened, and already high aquifer productivity would largely remain unchanged. However, Scibek and Allen [2006b] suggested that there would be a corresponding lowering of the discharge (as base flow) to the many streams that drain the aquifer, although the effects were not quantified due to limitations of the flow model for accurately simulating stream-aquifer interactions.

[48] On the basis of the results of the current analysis, base flow might be expected to either increase or decrease, and the range of variation is considerable, considering the differences in recharge predicted by the different GCMs. For example, by the 2080s, our study showed that the mean annual recharge to the Abbotsford aquifer could change by −1.5% to +25% (or −15 mm/yr to 250 mm/yr, assuming a rough mean annual spatial average of 1000 mm/yr) (Figure 13). A 25% increase in recharge could be a positive outcome, resulting in base flow being maintained at levels higher than present. This increased recharge could, however, result in flooding in low-lying areas, particularly during the wet season. A lowering of mean annual recharge by 1.5%, albeit by a small amount, would likely have the effect of lowering groundwater levels across the aquifer, and discharge to streams would decrease by a corresponding amount. The area of the aquifer is 161 km2, and assuming that the entire aquifer drains to local streams, a −15 mm annual decline in recharge could result in approximately 2.4 × 106 m3/yr decline in potential base flow discharge. If we assume that base flow contribution to streamflow is uniform throughout the year, a reasonable assumption, then one quarter of this amount (6 × 105 m3/yr) would discharge during the summer low flow period (July, August, and September). This equates to a loss of base flow, averaged over the summer to the three main streams, of roughly 0.02 m3/sec total (or 0.006 m3/sec per stream); summer low flows in these streams range from 0.001 to 0.2 m3/sec as measured at the international border [Berg and Allen, 2007]. Thus, during the critical low flow period in the summer, this loss of base flow could have negative implications for aquatic habitat. Of course, only four GCMs were considered in this study, and the range of outcomes is likely greater, spanning both negative and positive changes in recharge.

[49] The results of this study demonstrate the variability in simulated recharge, when only the GCM is varied, and the downscaling method and all other model recharge parameters remain the same. The approach used to quantify this variability considered the GCMs individually and explicitly simulated recharge for each. Alternative approaches (ensemble mean, weighted average) are other ways of communicating GCM outputs, but these tend to represent the greatest likelihood rather than the variability among the model output. In this paper, we specifically aimed to demonstrate the range of recharge outcomes owing to the choice of GCM. From a water management perspective or for protection of aquatic habitat, one could argue that it is better to know a possible range of outcomes and plan accordingly, rather than plan for the average likelihood.

6. Conclusions

[50] The uncertainty in GCM predictions and the consequence of this uncertainty on groundwater processes is very relevant to decision makers who are trying to manage water resources. It is important for scientists to quantify uncertainty in all stages of the modeling process, and it is also incumbent upon scientists to communicate this uncertainty to water managers so that they can plan accordingly. Uncertainty arises from the GCM itself (i.e., how well a particular model predicts climate both in the past and into the future) and at each step throughout the modeling process: the downscaling process; the recharge modeling; and the groundwater flow modeling.

[51] In this study, we only considered the variability of simulated groundwater recharge arising from choice of GCM, and the range of outcomes was significant. In this particular study area, not all GCMs predicted the same magnitude or direction of change for precipitation (and consequently recharge). As an extreme example, in the 2080s, the difference in recharge relative to the base case was −1.5% to +23.2%. Many studies have indeed demonstrated similar ranges of variability in other areas. However, in most cases the models at least all agree on the direction of change (all predicting more, or all predicting less precipitation). Coastal British Columbia and the Pacific Northwest of the USA is a region of close to zero precipitation change, based on numerous GCM runs [IPCC, 2007]; temperature, however, consistently increases in GCM model runs. Therefore, in this (and similar) regions, it is particularly important to consider that changes in groundwater recharge may be positive or negative, and that these may translate into positive or negative changes in groundwater level.

[52] Despite using a state-of-the-art downscaling method, downscaled temperature and precipitation for the four GCMs were not ideal. In particular, lower downscaled precipitation in the fall and early winter resulted in less annual groundwater recharge compared to simulated historical recharge generated using a stochastic weather series that is considered representative of observed data. These results point to the difficulties of downscaling climate data, and highlight the importance of obtaining accurate seasonal results for recharge modeling, particularly during the months when recharge is highest.

[53] Perhaps more important is the absolute change in recharge and what impact this might have on broader water management issues. During the critical low flow period (here in the summer), a loss of base flow under reduced recharge conditions could have negative implications for aquatic habitat. Therefore, in addition to planning for potential changes in groundwater supply, understanding and planning for potential changes in groundwater recharge due to climate change can have important implications for fisheries management.

[54] At this time, the various recharge results obtained from this study have not been input to the existing groundwater model for this study area. Because the streams are groundwater-fed, it is critical to use a model that does not unduly constrain the boundary conditions that represent them. Therefore, a coupled groundwater-surface water model is being developed to better explore the interconnections along the streams; the recharge results will be applied to this new model.


[55] We thank Christina Hendry for her assistance with graphing climate data output.