Water Resources Research

Modeled impacts of predicted climate change on recharge and groundwater levels

Authors


Abstract

[1] A methodology is developed for linking climate models and groundwater models to investigate future impacts of climate change on groundwater resources. An unconfined aquifer, situated near Grand Forks in south central British Columbia, Canada, is used to test the methodology. Climate change scenarios from the Canadian Global Coupled Model 1 (CGCM1) model runs are downscaled to local conditions using Statistical Downscaling Model (SDSM), and the change factors are extracted and applied in LARS-WG stochastic weather generator and then input to the recharge model. The recharge model simulated the direct recharge to the aquifer from infiltration of precipitation and consisted of spatially distributed recharge zones, represented in the Hydrologic Evaluation of Landfill Performance (HELP) hydrologic model linked to a geographic information system (GIS). A three-dimensional transient groundwater flow model, implemented in MODFLOW, is then used to simulate four climate scenarios in 1-year runs (1961–1999 present, 2010–2039, 2040–2069, and 2070–2099) and compare groundwater levels to present. The effect of spatial distribution of recharge on groundwater levels, compared to that of a single uniform recharge zone, is much larger than that of temporal variation in recharge, compared to a mean annual recharge representation. The predicted future climate for the Grand Forks area from the downscaled CGCM1 model will result in more recharge to the unconfined aquifer from spring to the summer season. However, the overall effect of recharge on the water balance is small because of dominant river-aquifer interactions and river water recharge.

1. Introduction

[2] In studies of global climate change, the impacts on water resources and the interactions between unconfined aquifers and the atmosphere are studied and modeled to determine impacts on water table levels. It is expected that predicted global changes in temperature and precipitation will alter groundwater recharge to aquifers, causing shifts in water table levels in unconfined aquifers as a first response to climate trends [Changnon et al., 1988; Zektser and Loaiciga, 1993]. Most research to date has been directed at forecasting the potential impacts to surface water hydrology, while for groundwater hydrology, typically only large, regional and coarse-resolution models have been undertaken to determine the sensitivity of groundwater systems to changes in critical input parameters, such as precipitation and runoff [e.g., York et al., 2002; Yusoff et al., 2002], with few exceptions of small aquifers and detailed investigations of potential impacts of climate change (scenarios) on unconfined aquifer water levels [e.g., Malcolm and Soulsby, 2000].

[3] Aquifer recharge from infiltration of precipitation, referred to as “recharge” in this paper, has traditionally been difficult to estimate for large areas, but a variety of methods have been used [Simmers, 1998]; from statistical empirical models linking precipitation trends to aquifer recharge and groundwater levels [Chen et al., 2002], to spatially distributed recharge applied to three-dimensional groundwater flow models [Jyrkama et al., 2002]. For the purposes of climate change impacts modeling, relative changes in recharge rates are of interest, and particularly how these relative changes are translated to the groundwater levels. Ideally, recharge rates should be as accurate as possible to represent the small shift from present to future climatic conditions. In practice, however, data limitations preclude detailed and highly accurate estimates that can be applied over a large area.

[4] This study was motivated by the Canadian government's initiative to assess the impacts of climate change and develop adaptive strategies for climate change under the auspices of Natural Resources Canada's Climate Change Action Fund. The study expands on a previous modeling study by Allen et al. [2004a] within the Grand Forks aquifer, in south central British Columbia, Canada (Figure 1), which used a sensitivity approach to model the potential impacts of climate change for published ranges of temperature and precipitation derived from Global Climate Models (GCMs) for the south central interior of BC (i.e., no downscaling). The approach used in that study was to assume uniform recharge for the entire surface of the aquifer, and to model recharge (one value) for the different scenarios by applying change factors to the monthly climate normals, which reflected extreme climate conditions (i.e., wet and warm, wet and cold, dry and warm, dry and cold). Ultimately, two extreme recharge conditions were used as boundary conditions to a steady state flow model. An independent sensitivity analysis was conducted to explore the effect of river stage elevation (using lower than base flow and higher than peak flow stage).

Figure 1.

Mountainous topography of the Grand Forks valley showing the unconfined valley aquifer (shaded gray) and drainage (white). The Kettle River (east flowing) is shown to meander through the valley and eventually discharges into the Columbia River. The valley widens near town of Grand Forks, where the Granby River flows into the Kettle River. Inset map shows the location map of the study area in British Columbia, Canada.

[5] An outcome of that work was that a fully integrated (recharge and river stage), high-resolution, transient approach may be needed to understand the dynamics of the groundwater system under future predicted climate conditions. In particular, employing a more rigorous approach to recharge modeling was anticipated to yield greater resolution on the subtleties of climate change on seasonal groundwater recharge. Thus the research objectives of the current study were twofold; to evaluate the importance of spatial distribution of recharge on groundwater modeling results, and to identify qualitatively (and, where possible, quantitatively) the uncertainties in the climate-to-model process.

[6] This paper presents a methodology for linking GCM predictions (via downscaling) to a recharge model, and then to a groundwater flow model. We discuss the recharge modeling results within the context of historic and future climate conditions, the sensitivity of the groundwater levels to recharge distribution (both spatial and temporal), and the uncertainties in the process. We also briefly compare the results to those of Allen et al. [2004a] in order to comment on the benefits of the added resolution to the modeling process.

[7] Because climate change is anticipated to impact both the timing and amplitude of flow in the Kettle River (snowmelt dominated system), consideration of impacts of climate change must necessarily consider both surface water and groundwater in this aquifer. In addition to refining the groundwater flow model to take into consideration the complexities of the aquifer architecture and consequent interconnection with surface water, modeled river stages for current and future climate scenarios were used as boundary conditions for each climate period modeled. The methodology and results of linking climate change through surface water interactions in this aquifer are treated in a separate paper [Scibek et al., 2006]. In this paper we focus on the effects of climate change on recharge from precipitation, touching on certain aspects of the hydrology where appropriate, to differentiate the effects where possible.

2. Recharge Modeling

2.1. Downscaling GCM Predictions

[8] Spatial downscaling techniques [Hewitson and Crane, 1996; Wilby and Wigley, 1997] are used to derive finer resolution climate information from coarser resolution GCM output, assuming that the statistical relationships, linking observed time series to GCM variables, will remain valid under future climate conditions. GCMs do not accurately predict local climate, but the internal consistency of these physically based climate models provides most likely estimates of ratios and differences (scaling factors) from historical (base case) to predicted scenarios [Loaiciga et al., 1996] for climatic variables, such as precipitation and temperature.

[9] Climate scenarios for modeled present and future conditions were taken from the Canadian Global Coupled Model (CGCM1) [Flato et al., 2000] for the IPCC IS92a greenhouse gas plus aerosol (GHG + A1) transient simulation. CGCM1 predictions are valid for Canada and fall in the average of other GCMs. These include absolute and relative changes in precipitation and temperature. Temperature statistics were: mean, median, minimum, maximum, variance, and interquartile range. Precipitation variables were: mean, median, maximum, variance, dry/wet spell length, and % wet days in the month. Five daily data sets for CGCM1 were obtained from the CCIS Canadian Climate Impacts Scenarios (Web site access to climate scenarios and downscaling tools, including Statistical Downscaling Model (SDSM), http://www.cics.uvic.ca/scenarios/) for a grid location nearest to Grand Forks (Y = 11 latitude: 50.09°N and X = 16 longitude: 120°W; Grand Forks is at 49.1°N and 118.2°W). Four were CGCM1 scenarios, each with data for a number of potential predictor variables. The “current climate” scenario was generated by CGCM1 for the period 1961–2000. The subsequent “future climate” experiments using CGCM1 with GHG + A1 were for 2020s, 2050s, and 2070s. The fifth data set was a calibration data set for the downscaling model. The calibration data set contains observed daily data for 1961–2000, derived from the National Center for Environmental Prediction (NCEP) reanalysis data set [Kalnay et al., 1996] for the period 1961–2000. Monthly means and other statistics were calculated from mean daily values, and the NCEP data set had 10% or smaller bias to observed precipitation at Grand Forks (compared monthly means), thus we have high confidence in using NCEP data for calibration of downscaling model.

[10] The downscaling of CGCM1 results was accomplished using two independent methods: (1) Statistical Downscaling Model (SDSM) [Wilby et al., 2002] and (2) principal component K-nn [e.g., Zorita and von Storch, 1999; Yates et al., 2003] computed by Environment Canada [Whitfield and Cannon, 2000]. Four climate scenarios (30 years of daily weather) were generated using each calibrated downscaling model: current climate (1960–1999), 2020s climate (2010–2039), 2050s climate (2040–2069), and 2080s climate (2070–2099).

[11] Downscaled daily temperature time series were analyzed for (1) mean daily temperature and (2) standard deviation. The two downscaling methods yield comparable estimates of mean monthly temperature, and calibration bias is small (Figure 2). Downscaled daily precipitation time series were analyzed for (1) mean monthly precipitation, (2) standard deviation in daily precipitation, (3) percent wet days, (4) dry spell length, and (5) wet spell length. Precipitation has variable seasonal/monthly predicted changes, and results vary somewhat between downscaling methods. Figure 3 shows the calibration results for mean monthly precipitation as calculated from daily values. The SDSM downscaled precipitation series are too low in the late spring to summer months, especially June, but fit the observed normals reasonably well in other months. This is surprising because the SDSM model was well calibrated for monthly precipitation means, and calibration bias from NCEP data set to observed was less than a 10% difference for most months. The problem lies in inability of CGCM1 to adequately model precipitation for Grand Forks, especially in the summer months, likely due to local convective precipitation and valley-mountain-rain-shadow effects, which have a strong influence on local precipitation. Precipitation was underestimated by roughly 40% compared to observed during the summer, even after downscaling with an adequately calibrated model.

Figure 2.

Mean monthly temperature at Grand Forks, British Columbia: observed and downscaled from CGCM1 model runs for current and future climate scenarios using (left) SDSM and (right) K-nn.

Figure 3.

Mean monthly precipitation at Grand Forks, British Columbia: observed and downscaled from CGCM1 model runs for current and future climate scenarios using (left) SDSM and (right) K-nn.

[12] PCA k-nn downscaling of the same data set gave worse results than SDSM downscaling. The only exception is late summer when PCA k-nn performed better than SDSM model, in terms of fitting to observed data. Other variables used in calibration, such as precipitation variability (standard deviation), percent wet days, wet spell length, dry spell length gave similar results for both methods, although standard deviation was notably better for SDSM. Both downscaling methods underestimated percent wet days by about 30% in May and June, underestimated wet spell length by about one day, and underestimated dry spell length by about 30% in all months except June and July.

[13] Ideally, several downscaling methods could be used and compared, and either the average of all computed, or the best method selected, noting the uncertainties involved. On the basis of all the variables considered during calibration, SDSM is better calibrated than PCA k-nn; however, neither preformed very well. This represents a fundamental limitation of CGCM1 predictions, and is considered as “model bias”. Such large model bias precluded direct use of daily downscaled precipitation at this site, because the important seasonality of precipitation would not be represented correctly. To overcome this problem, we had to assume that at least the relative and/or absolute changes in precipitation and temperature between present and future climate scenarios have strong physical basis and meaning. Thus our approach was to compute change factors from SDSM (relative for precipitation, and absolute for temperature), and redistribute them to daily time series using a stochastic series weather generator, LARS-WG [Racsko et al., 1991; Semenov et al., 1998].

2.2. Applying Change Factors to LARS-WG

[14] Change factors for temperature indicate an increase of approximately 1°C per 30 years for all months (Figure 4). The change factors for precipitation are more complex, and indicate an increase in July and August, variable changes (increase or decrease) in other months, and corresponding changes in percent wet days for those months (Figure 5); relative changes in precipitation are shown separately for the dry and wet months of the year. The climate at Grand Forks is predicted to become slightly wetter in the “dry months,” February–March and July–August (Figure 5a), increasing to greater and greater amounts from present time to 2099. From present to the 2010–2039 period, there is predicted increase in precipitation by factors of < 1.2, 1.1 to 1.4 by 2040–2069, and eventually, 1.2 to 1.9 by 2070–2099. However, in the early autumn (September–October), the precipitation will decrease from present to 2040–2069 by a factor of approximately 0.9. In the “wet months” (Figure 5b), there will be a very small increase in winter precipitation (factor of 1.1 or less), but a decrease in precipitation in May and June by factor of 0.9 to 0.82 (probably as a result of shift in spring to an earlier date and a shorter winter). Although not shown here, a comparison of the results from SDSM and PCA K-nn reveals different magnitudes and directions of change, mostly for precipitation, in future climate scenarios, demonstrating the uncertainty associated with the downscaling process [Allen et al., 2004b]. Although this uncertainty limits the predictive aspect of this (and similar) studies, it does not detract from the study's usefulness as a realistic sensitivity analysis to potential climate change, whatever the actual climate changes in each month will be in the future.

Figure 4.

Absolute change in seasonal temperature predicted by CGCM1 model runs, after downscaling with SDSM for Grand Forks, British Columbia.

Figure 5.

Relative change in monthly and seasonal precipitation predicted by CGCM1 model runs, after downscaling with SDSM for Grand Forks, British Columbia, comparing four seasons and months within each season.

[15] Daily weather was generated using LARS-WG (calibrated) to reproduce accurately both the magnitude and variability of seasonal precipitation and temperature, and to generate daily weather for future climate scenarios based on those change factors. LARS-WG utilizes semiempirical distributions for the lengths of wet and dry day series, daily precipitation and daily solar radiation, and compares favorably to other weather generators according to Wilks and Wilby [1999]. Calibration performance of LARS-WG for mean, maximum and minimum temperature was excellent, as was daily minimum and maximum temperature variation (standard deviation). In winter months, LARS-WG produced 0.5 to 1.0°C cooler minimum temperatures than observed, when comparing variability in monthly values. Mean monthly rainfall was within 2 mm/month (within 5%) or closer for all months (4–50 mm/month precipitation values). 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. A χ2 test gave very good results for wet/dry precipitation series.

[16] The base case for stochastic weather generation is defined as the average of the entire historical period, assuming that it is representative of preclimate change conditions. Then, climate change scenarios are generated by perturbing the generated weather using the change factors to modify the base case (Table 1). Each scenario consists of 100 years of generated weather, noting that although generated weather runs of 1000 years converge better to specified “normals”, there are diminishing returns of performance after 100 years (Figure 6). 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 for the scenario (e.g., seasonality of precipitation and its variability).

Figure 6.

Comparing mean monthly weather parameters for the 2010–2039 climate scenario and the 30 and 100 year synthetic weather runs. Data are expressed as relative or absolute changes in monthly precipitation, temperature, and solar radiation relative to the observed value, where observed represents the 1970–1999 historical monthly mean value.

Table 1. Climate Scenario Input (Scenario File Example) From SDSM to LARS-WG Stochastic Weather Generator for Grand Forks, British Columbiaa
 GF_CGCM1_2010_2039GF_CGCM1_2040_2069GF_CGCM1_2070_2099
RainWetDryTstdev TSRadRainWetDryTstdev TSRadRainWetDryTstdev TSRad
  • a

    Rain, precipitation relative change (future/base) or (base/base), mm/mm; Wet, wet spell length relative change, mm/mm; Dry, dry spell length relative change, mm/mm; T, temperature absolute change, °C; stdev T, standard deviation of temperature relative change, °C; SRad, solar radiation absolute change, MJ/m2/d. Base case (1970–1999) for all months have Rain = 1.00, Wet = 1.00, Dry = 1.00, stdev T = 1.00, and T = 0.00 and SRad = 0.00.

Jan1.401.330.810.740.940.141.441.490.872.181.270.191.711.420.815.111.120.10
Feb1.141.310.950.620.980.031.281.380.912.301.050.051.511.480.904.991.020.04
Mar1.121.211.070.941.090.171.371.300.902.131.220.271.671.580.764.871.370.25
Apr1.131.061.121.531.220.261.451.140.943.481.190.401.491.141.076.081.010.34
May0.900.971.212.190.910.391.000.831.164.070.960.401.190.951.236.570.94−0.31
Jun1.040.921.301.710.86−0.080.860.991.313.630.82−0.190.650.841.375.820.65−0.36
Jul0.831.101.011.000.780.150.731.001.151.900.660.260.570.891.302.510.600.31
Aug1.191.270.950.410.85−0.021.111.210.961.100.71−0.171.101.320.840.970.61−0.56
Sep1.030.870.750.910.84−0.240.921.230.802.740.78−0.441.271.530.713.930.55−0.58
Oct0.971.171.241.270.940.111.031.011.283.800.970.250.880.921.496.630.880.50
Nov1.041.171.291.330.840.001.151.391.444.030.840.051.221.211.247.190.880.20
Dec1.211.020.821.150.92−0.031.391.110.603.880.78−0.071.461.170.716.190.99−0.12

[17] Thus daily data were generated for each climate period: current climate (1960–1999), 2020s climate (2010–2039) and 2050s climate (2040–2069), and 2080s climate (2070–2099). These data were used as input to a one dimensional recharge model. Spatially distributed recharge zonation was accomplished, as discussed in the following section, by running this model for different combinations of soil and unsaturated zone properties. Precipitation was assumed to be uniform over the aquifer in this valley.

2.3. Recharge Zonation

[18] Recharge models are used to estimate recharge where there is lack of direct measurements, which would adequately represent a large aquifer area. In recharge modeling, GIS data-handling capabilities allow raster map algebra with classed maps of unsaturated zone properties, such as soil permeability (from soil type and land cover), to generate spatial distribution of recharge [Fayer et al., 1996]. There are many methods for recharge modeling [York et al., 2002], but the methodology presented here generates spatially distributed and temporally varying recharge zones, using a GIS linked to the one-dimensional U.S. Environmental Protection Agency's Hydrologic Evaluation of Landfill Performance (HELP) model [Schroeder et al., 1994]. The program WHI UnSat Suite [Waterloo Hydrogeologic Inc., 2000], which includes the subcode Visual HELP, is used to estimate direct recharge from precipitation to the Grand Forks aquifer.

[19] HELP is a versatile quasi-two-dimensional layer model that was originally designed for conducting water balance analyses and predicting hydrologic processes at landfills. However, in recent years it has been used effectively for estimating of groundwater recharge [Jyrkama et al., 2002]. While the performance of the HELP model has been found to be adequate compared to other similar models for most conditions [Scanlon et al., 2002] and therefore was selected for this study, any suitable infiltration model could be used, or several models could be compared in cases where local recharge is the controlling variable on groundwater levels.

[20] HELP uses numerical solution techniques that account for the effects of surface storage, snowmelt, runoff, infiltration, evapotranspiration, vegetative growth, soil moisture storage, and various engineering parameters (e.g., lateral subsurface drainage). The natural water balance components that the program simulates include precipitation, interception of rainwater by leaves, evaporation by leaves, surface runoff, evaporation from soil, plant transpiration, snow accumulation and melting, and percolation of water through the soil profile. The rainfall-runoff processes in HELP are modeled using the U.S. Department of Agriculture (USDA) Soil Conservation Service curve number (CN) method [USDA, 1986], and allows the user to adjust the runoff calculation to a variety of soil types and land management practices. HELP uses different procedures to adjust the value of CN to surface slope (here assumed zero slope), soil texture, and vegetation class (soil classes were used). The initial water moisture storage of layers were estimated, applied as initial values, and then recharge was simulated over a 1-year period.

[21] The approach used for spatially distributed recharge modeling is similar to that of Jyrkama et al. [2002], in which a methodology was developed for estimating temporally varying and physically based recharge using HELP for any MODFLOW grid cell. Our approach also depends on high-resolution GIS maps (20 m grid) for defining recharge zones, and links these zones to MODFLOW model grids, although we developed a distinct methodology and code that links Visual MODFLOW version 3.1.84 [Waterloo Hydrogeologic Inc., 2004] to Arc GIS version 8.13 [Environmental Systems Research Institute, 2004] for input and output of MODFLOW simulations. Our method differs from previous distributed recharge methods in that we also estimate the distribution of vertical saturated hydraulic conductivity in the unsaturated zone and the thickness of the unsaturated zone, and incorporate irrigation return flow.

[22] There are many physical properties of the subsurface that affect recharge to an unconfined aquifer and these have three-dimensional distribution; some change with time as well, such as soil moisture and depth to water table. The available data constrain the choice of some parameters through relatively good ground truthing, while other parameter values must be inferred from other information, and essentially estimated (Table 2). The parameters in Table 2 are listed in order of importance in each group. Usually, the type of local climate and, more specifically, seasonal distribution of precipitation will have dominant control on aerial recharge. The aquifer properties will control the actual amount of recharge into the aquifer, and are assumed constant in time, except unsaturated zone thickness, which will fluctuate seasonally.

Table 2. Variables Considered for Recharge Modeling
Available VariablesEstimated Variables
Climate
Precipitation (daily to hourly),Evapotranspiration (daily)
 Surface runoff (in low-permeability soils)
 
Aquifer Media Properties
Unsaturated zone thickness (depth to water table)Unsaturated zone hydraulic properties from lithology at point locations (equivalent saturated hydraulic conductivity)
Soil type (permeability)Soil thickness distribution (assume uniform 1 m due to lack of data of adequate spatial resolution)
 
Ground Surface Properties
Vegetation coverEffect of vegetation on recharge
Irrigation rates and areas affectedReturn flow to recharge
Elevation and slope of ground surface (valley floor topography) 

[23] Soil thickness was interpolated from 55 well lithologs (only 55 contained soil thickness data out of a total of 150) and dozens of soil pits. However, soils are expected to vary in thickness over microtopography of the valley, thus any valley-wide interpolation of thickness would have very large error locally. With the exception of a few anomalous locations, the soil thickness is rather similar over the valley; the mean interpolated soil thickness is 0.92 ± 0.21 m, and thus the soil thickness was assumed to be simply 1.0 m in all percolation columns for recharge modeling (Figure 7a).

Figure 7.

(a) Soil thickness, (b) soil permeability classes (see Table 3), (c) reclassed Kz map of unsaturated zone above water table in Grand Forks aquifer, (d) depth to water table classes, and (e) resulting recharge zones.

[24] Soil permeability maps were modified by land use to account for less permeable areas. Four representative permeability classes were created for very high, high, medium, and low permeability (see Table 3). Similar soils were combined into one category to reduce the number of categories to four, based on the spatial extent of each permeability class, to preserve the most representative soil types over the aquifer extent (Figure 7b).

Table 3. Permeability Classes for Soils in Grand Forks Aquifer Area
Vertical Percolation Layer in HELPVertical Kzsat,a cm/sVertical Kzsat, m/dPermeability
  • a

    Read 1.90E-04 as 1.90 × 10−4.

Silty loam1.90E-040.164low
Loam3.70E-040.320 
Fine sandy loam5.20E-040.449 
Sandy loam7.20E-040.622moderate
Loamy fine sand1.00E-030.864 
Loamy sand1.70E-031.469high
Sandy gravelly soils5.80E-035.011very high

[25] Saturated hydraulic conductivity was estimated for geologic units encountered above the water table. Well lithology data were standardized and classified to simplify the data. Up to three material descriptions were retained for each depth interval. Saturated (assumed vertical) hydraulic conductivity (Kz), specific storage (Ss) and specific yield (Sy) for each material type were assigned based on representative values in the published literature, and were constrained by parameters estimated from pumping test data. Geometric means of the Kz values were calculated for each layer in each well where more than one material type was recorded. A manual examination of the output data was carried out in order to ensure that the calculated hydraulic conductivities were consistent with the dominant sediment description in the original well log. In only a few cases (<10) were modifications made as a result of the standardization scheme not correctly identifying the dominant material types. Equivalent Kz was computed for each well location, assuming homogeneous and isotropic “units”. Kz values in 285 wells ranged from a minimum of 1 × 10−6 m/d to a maximum of 1000 m/d, had a geometric mean of 7.5 m/d, a median of 21 m/d, and quartile values of 1.6 and 316 m/d. The Kz values in the unsaturated zone were interpolated using Inverse Distance Weighed interpolator, and computed on representative vertically averaged log Kz values at all available point locations where lithologs exist. After interpolation, the inverse logarithm (i.e., 10^[Log Kz]) of the interpolated raster was computed, and converted to units of m/d. Four classes were chosen as 1 × 10−6 to 9.9, 10 to 49.9 m/d, 50 to 100 m/d, and 100 to 1000 m/d (Figure 7c). The representative Kz values for each material in the HELP soil columns were 5, 30, 75, and 500 m/d (mid value in each class).

[26] Depth to water table was estimated from the difference between ground surface and a numerically derived static groundwater table [Allen et al., 2004a]. This assumption is reasonable as the depth to the water table is usually much larger than the variation in groundwater level, except in the low-lying river floodplain region where river effects dominate the water levels, not recharge [Scibek et al., 2006]. Depths to the water table in 285 wells ranged from 1.5 m to 46.8 m, with a mean of 11.4 m, a median of 10.1 m, and quartile values of 6.1 m and 12.9 m. The depth classes were based on quartiles of distribution: 0 to 6 m, 6.1 to 10 m, 10.1 to 12.9 m, 13.0 to 47.0 m, with roughly 25% of aquifer area in each of four categories (Figure 7d). Representative sediment columns in HELP were assigned representative midclass depths: 3, 8, 11, and 25 m.

[27] Recharge zones were defined for a 50 m raster grid through cross classification of maps of all important variable distributions (Figure 7e), resulting in 65 zones (zone 1 was the default zone in MODFLOW with no recharge). A recharge zone is any unique combination of soil permeability class, hydraulic conductivity class, and depth to water table class, thus the number of combinations of soil permeability, Kz, and water depth, was (4 × 4 × 4 = ) 64 soil columns. As stated earlier, soil thickness was assumed the same for all columns. There is a degree of uncertainty in each of these properties because data come from various sources and formats. In this study, the limiting variable in terms of number of classes possible is soil type (originally soil polygons), and the most uncertain is Kz, as representative values were estimated, whereas depth to water table could be represented at 20 m grid or smaller with reasonable accuracy. Higher spatial resolution would have required more subclasses in each variable and resulted in many more recharge zones.

[28] An important element of this study was to evaluate the sensitivity of modeled recharge to HELP input parameters. The HELP model results showed a very small effect (<5% change) for: type of stand of grass, wilting point, field capacity, and initial moisture content (results not shown). A moderate effect was found in soil thickness (Figure 8e) and porosity of percolation layer (Figure 8f). As soil thickness increased, the modeled recharge decreased, but only very strongly from May to July (wet months at this site). Recharge was found to be most sensitive to depth of unsaturated zone (depth to water table) for either a low (Figure 8c) or high soil permeability (Figure 8d). Similar sensitivity was observed for different soil permeability (Figure 8b) and for saturated vertical hydraulic conductivity of unsaturated zone (Figure 8a). The effects were seasonal and most pronounced in early summer, due to the combination of precipitation and temperatures, which control evapotranspiration and infiltration rates (together with unsaturated zone properties). Lag times were observed for recharge reaching the aquifer following precipitation events (also seen on monthly means presented in Figure 8), especially where low-permeability soil was used. The high sensitivity of recharge models to unsaturated zone properties suggests that spatial distribution of such properties must be accounted for in recharge modeling for climate change impacts assessment in surficial aquifers.

Figure 8.

Sensitivity of recharge estimates modeled with HELP to (a) soil permeability, (b) saturated vertical hydraulic conductivity of unsaturated zone, (c) depth of unsaturated zone (depth to water table) for with a high-permeability soil, (d) depth of unsaturated zone (depth to water table) with a low-permeability soil, (e) soil thickness, and (f) porosity of unsaturated zone material.

[29] A better representation of recharge in an agricultural area takes into account the amount of water that is returned to the aquifer when the land is irrigated. This is commonly referred to as irrigation return flow (from the aquifer perspective). In all pumping model scenarios, the recharge zones were modified by superimposing estimated irrigation return flow to the aquifer. Generalized estimates of return flow were obtained through consultation with experts in irrigation practices (roughly 25% of the amount of irrigation for the types of crops present). We assumed constant irrigation return flow in irrigated fields for each month (June to August only) in all present and future climate scenarios. As future irrigation predictions would require estimates of population change, land use change, technology change, and climate change and associated feedbacks, we were unable to predict changes in irrigation demand, and thus changes in future irrigation return flow. Nonetheless, this is certainly an aspect that should be considered if such information could be predicted. In the MODFLOW model, described later, additional recharge zones were created to represent the modified recharge after addition of return flow from irrigation.

2.4. Historical Recharge Results

[30] Previous recharge modeling [Allen et al., 2004a] used a uniform annual recharge value for the Grand Forks aquifer of 135.5 mm/year, or approximately 27% of precipitation. According to this study, mean annual recharge varies considerably across the 64 recharge zones (Figure 9), ranging from slightly less than 30 mm/year to over 120 mm/year (6% and 24% of mean annual precipitation). The low recharge values are associated with areas of thick gravels in the terraces, which can absorb a large amount of rainfall before reaching saturation, and because this region is relatively dry, the modeled recharge is low in areas with large depth to water table. It would be expected that infiltrated water below the root zone would continue to water table and, after some lag time, contribute to recharge, so this result is somewhat surprising, and perhaps points to a limitation of the recharge model used. The initial moisture contents of these deep layers are unknown, but were estimated in HELP by first running one year of weather, and using the moisture content at the end of that year as the initial value. Nonetheless, the recharge model was not greatly sensitive to initial moisture content in a one year simulation conducted at daily time steps.

Figure 9.

Historical mean annual recharge to the Grand Forks aquifer for the historical climate scenario (1961–1999), modeled in HELP and assigned to recharge zones.

[31] Mean monthly recharge (to the inset area shown in Figure 9) is shown in Figure 10. Recharge follows the annual distribution of precipitation. In winter the ground is frozen and snowmelt infiltration does not occur. Most of the recharge is received in late spring and summer through snowmelt and summer rainstorm events. The autumn season has moderate recharge, less than in early summer. Monthly recharge varies from <2 mm/month to >12 mm/month (Figure 10). The range in percentages (i.e., 6% to 24%) of mean annual precipitation is smaller than the range in percentages (10% to 80%) of monthly precipitation, due to seasonal variation in precipitation and averaging on annual timescales. For example, during late summer the aquifer receives 60% to 90% of precipitation, but the overall recharge amount is small because rainstorms are infrequent. The high percent value is explained by higher rain intensities in that season, as the LARS-WG preserves the intensities of rain events; we observe that if a high-intensity event occurs during the late summer (such as a thunderstorm), it rains heavily and most of the water infiltrates the aquifer. If it were to rain slowly and over a longer time, much more of it would evaporate. This type of relation may be very different in other climate regions and in other aquifers where high-intensity rainfall events may lead to increased runoff and less infiltration.

Figure 10.

Historical mean monthly recharge maps for inset area (central portion of valley) as shown in Figure 9.

2.5. Predicted Recharge for Future Climates

[32] The predicted changes in mean annual recharge were converted to percentage differences: (future − historical)/historical. The 2010–2039 climate scenario has a predicted 2 to 7% increase from historical mean annual recharge and there are no predicted decreases (Figure 11, top). The 2040–2069 climate scenario has a predicted 11 to 25% increase from historical mean annual recharge, also without any predicted decreases (Figure 11, bottom). Although not shown, changes in seasonal recharge follow a pattern consistent with what is predicted annually, with April to June showing the greatest increases in monthly recharge.

Figure 11.

Percent change in mean annual recharge to the Grand Forks aquifer modeled in HELP and assigned to recharge zones between (a) 2010–2039 and historical and (b) 2040–2069 and historical. Historical climate scenario is 1961–1999.

[33] We did not specifically model the potential changes to snowmelt timing and resulting earlier spring thaw in the soil, and also a potentially longer growing season. However, by the time the spring rains come at Grand Forks the ground is usually thawed and rain on snow is not significant. Changes in the amount of snow on the ground prior to snowmelt, caused by climate modification, were not modeled as there is no reliable information suggesting any such changes and neither in which direction these might occur.

[34] To quantify the effect of changes in recharge on groundwater levels in the aquifer, recharge values for each climate time period were used as a surface boundary condition in a transient groundwater flow model, implemented in Visual MODFLOW, as discussed in the following section.

3. Groundwater Modeling

3.1. Flow Model Construction

[35] The Grand Forks aquifer (34 km2 in area) is contained within the mountainous valley of the Kettle River near the City of Grand Forks (see Figure 1). Bedrock is composed of metamorphic rocks, and valley fill consists of glaciolacustrine and glaciofluvial sediments. Within the Grand Forks valley, the Kettle River is a meandering gravel bed river incised into the glacial outwash sediments. Previous interpretations of the hydrostratigraphy [Wei et al., 1994; Allen et al., 2004a] assumed a uniformly layered paradigm, whereby the unconsolidated sediments have presumed horizontal layered stratigraphy. Reinterpretation of the valley hydrostratigraphy as part of this study was aided by considering other similar valleys in BC, where the basal units are commonly silt, clay and gravel, overlain by thick glaciolacustrine silts [Fulton and Smith, 1978; Fulton, 1984; Clague, 1981; Ryder et al., 1991], and capped by Holocene sandy and gravelly outwash and floodplain deposits and paraglacial alluvial fans.

[36] The hydrostratigraphic model (Figure 12) was developed using GIS in combination with GMS version 4.0 [Brigham Young University, 2002]. First, the bedrock surface was modeled using known bedrock depths (in wells) and extrapolating the valley sides along numerous transects [Allen et al., 2004b]. The valley attains a maximum depth of approximately 250 m below ground surface, but typical sediment thickness is about 50 to 100 m (Figure 12). The valley fill was then modeled from 150 standardized, reclassified, and interpreted well borehole lithologs [Allen et al., 2004b]. Hydrostratigraphic units from the bottom up consist of a clay, silt, sand and gravel. The deep stratigraphy is relatively poorly understood as most wells are shallow; however, based on the low-permeability nature of these deeper sediments, the lack of deep information is not anticipated to affect the results for the upper aquifer horizons. Solid models, representing different hydrostratigraphic units, were constructed and converted to layers in Visual MODFLOW [Waterloo Hydrogeologic Inc., 2004], as is typically done with complex multilayer aquifer systems [Herzog et al., 2003].

Figure 12.

Fence diagram of hydrostratigraphic units in the Grand Forks valley.

[37] The bedrock is considered impermeable and forms a zero flux boundary condition. Spatially distributed recharge was applied as a surface specified flux boundary condition, with each MODFLOW cell having an independent schedule for recharge. River stage for the historic and future climate periods was modeled using BRANCH [Schaffranek et al., 1981], based on current and future predicted discharge at several gauging stations. The details of river modeling are discussed by Scibek et al. [2006]. As river bottom sediments consist of mostly gravels, with very little fine sediments, the river nodes in the model were treated as specified head boundaries, with head schedules representing the modeled river stage in the transient flow model. Major production wells, used largely for irrigation in the summer, were also included in the model.

3.2. Modeling Approach

[38] The steady state model was calibrated to historic static water levels in approximately 300 wells, while the transient model was verified against one high-quality provincial observation well with monthly records and the anticipated transient behavior of the aquifer. The normalized RMS (root mean squared) error for residuals between calculated and observed head was roughly 8%, and most data fall within the 95% confidence interval. A detailed model description is beyond the scope of this paper, but can be found elsewhere [Allen et al., 2004b].

[39] Before modeling the different climate periods, we investigated the flow model sensitivity to the type of recharge distribution over the aquifer area. We compared the results of the base case (scenario 1A representing spatially distributed and temporally varying historical recharge under nonpumping conditions) to each of spatially distributed mean annual recharge (scenario 5A), and nonspatially distributed monthly varying recharge (scenario 5B). This analysis quantifies the uncertainties associated with representation of recharge as a spatial distribution in the model.

[40] The model was then run for the climate change periods using the base case and appropriate river boundary conditions for each climate scenario. Results for each future climate period (scenarios 1C, 1D, and 1E) were also compared to active pumping conditions (scenario 2A, 2B, 2C and 2D). These model runs form the estimates of climate change impacts on groundwater levels in this aquifer, and the practical “result” of the entire analysis based on our methodology.

[41] The model was also run to test a different representation of aquifer properties (homogeneous K versus heterogeneous K distribution) for each climate change period. In this paper, we describe the results of the homogenous aquifer properties model. The reader is referred to Allen et al. [2004b] for a detailed discussion of these results. The homogeneous case was selected because it is simple, errors in the hydraulic conductivity data are reduced by averaging over the aquifer area, and the least assumptions are made about local geology. Homogeneous and isotropic values of hydraulic conductivity were assigned to each layer, based on average values determined from pump test data.

[42] Finally, a water balance for each model was computed using Zone Budget (ZBUD) in MODFLOW [Harbaugh, 1990]. The zones represent different irrigation districts within the unconfined aquifer, the river floodplain, and deeper model layers. Temporal changes in mass balance components were graphed to show relations between pumping, storage, recharge, and flow for each climate scenario.

3.3. Modeling Results

3.3.1. Irrigation Return Flow

[43] On the basis of groundwater flow model simulation results for historical climate (scenario 1A), volumetric recharge accounts for only 1 to 7% of other flow components, such as flow between zones and storage. The solid black line in Figure 13 graphs historical recharge for zones 4 and 5, which represent the large Sion (two zones) and Big Y (one zone) irrigation districts in the valley. Recharge flow volumes are small for such large zones compared to other flow terms (e.g., recharge is 2% of other flow rates, such as flow to/from other zones). Irrigation return flow increases recharge by 10 to 20%, and the importance of return flow depends on percent irritated area. In the late time steps of the model year, the recharge rates for the pumping model are higher than for the nonpumping model, possibly as a result of drawdown in some areas. Drawdown creates more “dry” cells in overlying aquifer layers in the MODFLOW model, and redirects more recharge to the silt layer below.

Figure 13.

Volumetric recharge to zones 4 and 5 (corresponding to Sion and Big Y irrigation districts, respectively), comparing nonpumping to pumping for all climate scenarios. Symbol legend applies to both graphs. Note the different vertical scale on graphs.

3.3.2. Sensitivity to Recharge Distribution

[44] Two sensitivity scenarios were run to investigate the effects of representation of distribution of recharge on model results. The first scenario (scenario 5A) had spatially distributed mean annual recharge as the recharge input. The second scenario (scenario 5B) had temporally variable recharge rates with uniform spatial distribution (one recharge zone). Each scenario (5A and 5B) is compared to the historical base case (scenario 1A). Thus the difference between scenario 5A and 1A highlights the advantage of using temporally varying recharge, and the difference between scenario 5B and 1A reflect the advantage of using spatially dependent soil permeability and water table depth. Only the nonpumping historical climate scenarios were used to test the sensitivity of the model to recharge distribution.

[45] In the Grand Forks aquifer, the water levels are sensitive to recharge only away from river floodplain where river levels dominantly control groundwater levels [Allen et al., 2004b; Scibek et al., 2006]. The maximum difference in water table elevation between scenario 5B and 1A is between 10 and 50 cm, but typically about 20 cm (Figure 14). For example, at days 101, 131 and 265, the difference in water levels away from the river is less than 10 cm (Figure 14). In this sensitivity analysis, the single recharge zone in 5B was chosen arbitrarily to represent “high” recharge, or a shallow depth to water table and high hydraulic conductivity of the unsaturated zone, consistent with previous studies [Allen et al., 2004a]. Thus applying recharge zonation reduces recharge from a uniform “high” value to a range of values depending on recharge zone.

Figure 14.

Water level differences calculated from the difference between model scenario outputs (5B − 1A) to illustrate the significance of spatial distribution of recharge on results. Scenario 5B represents spatially nondistributed recharge (i.e., mean monthly recharge applied to a single zone). Scenario 1A is the historical base case using spatially and temporally varying recharge. Maps are by time step in days 101 to 265. Contours are shown at 0.1 m interval. Zero contour is dashed line.

[46] The components of flow were also considered in the comparison of the different scenarios. Both storage (Figure 15a) and flow from/to other zones (Figure 15b) as a percentage of flow volume show a greater percent difference for scenario 5B − 1A as compared to scenario 5A − 1A. Thus the spatial distribution of recharge representation in the model has a more significant influence on the water balance than the temporal representation of recharge. Consequently, if the mean annual recharge value is applied to the model, as would be the case when climate data are lacking or when temporal recharge estimates are unavailable, then modeled water levels would likely be within 10 cm of that modeled with temporally variable recharge.

Figure 15.

Effect of both spatial and temporal distribution of recharge on (a) storage and (b) flow from/to other zones in nonpumping groundwater flow models. Values are calculated from the difference between model scenario outputs (5A – 1A) and (5B − 1A), where 1A is historical base case using spatially and temporally variable recharge, 5A is temporally constant recharge (i.e., annual recharge applied in a distributed fashion), and 5B is spatially nondistributed recharge (i.e., mean monthly recharge applied to a single zone). Percentage is calculated from (OUT − IN)/average (OUT + IN).

3.3.3. Surface Water Interactions With Kettle River

[47] High quality monthly water level records in an observation well located in the river floodplain indicate that the groundwater levels fluctuate predictably and regularly with rising and falling Kettle River stage over each annual hydrologic cycle [Allen et al., 2004b]. The river carries an order of magnitude more flow (per unit time) than exchanges between the river and the aquifer (i.e., a maximum 41 m3/s, which translates to 15% of river flow during spring freshet), as determined using our flow model, so effects of groundwater inflow on river discharge and stage are presumed negligible. In areas distal from the river, the effect is relatively small, but significant, and it varies over the year. A detailed water balance has been calculated [Scibek et al., 2006].

[48] During spring freshet on the Kettle River, the rise in river stage causes inflow of water to various ZBUD zones (after passing through the floodplain area). This excess water is stored in the aquifer. Mass balance calculations indicate that storage rates are less than 50% of interzonal groundwater flux, and 15 to 20% of river-aquifer flux. As river stage drops, the hydraulic gradient is reversed; water is released from storage and enters the floodplain zone where it eventually returns to river as base flow. As most of the pumping water is lost to evapotranspiration on irrigated fields, there is a small reduction in the base flow component to the Kettle River during the pumping period.

[49] Currently, the peak of the hydrograph for the Kettle River occurs between day 100 and 150. In future climate scenarios the hydrograph peak is predicted to shift to an earlier date [Cannon and Whitfield, 2002], although the peak flow will remain the same and the summer base flow period is extended and at lower levels [Scibek et al., 2006]. The shifts in the river hydrograph are predicted to be much greater in 2040–2069 than 2010–2039, both compared to the modeled historical 1960–1999 time period.

3.3.4. Climate Change Impacts on Groundwater Levels

[50] Within an annual cycle and between climate scenarios the results show different spatial and temporal distributions in groundwater conditions. Figure 13 shows graphically the temporal results for both pumping and nonpumping conditions for each climate change period for two major irrigation districts. Recharge is slightly lower in the late winter, and increases significantly in the late spring and summer for future climate periods. Pumping has the effect of increasing recharge during the summer, when the wells are active and irrigation return flow is added to recharge.

[51] Head difference maps show differences in water levels between two future climate periods relative to historic (Figure 16). Although pumping was included in all climate change simulations, the effects of pumping are effectively subtracted out on these maps because drawdown was identical in all climate scenarios (pumping rates were constant in all models for the pumping time period). At present-day, the flow patterns are influenced by river channel profile, and generally follow valley floor topography. In this aquifer, the effect of changing recharge on groundwater levels is very small compared to changes in timing of basin-scale snowmelt events in the Kettle River, and the subsequent shift in the hydrograph as discussed above, although this may not be the case in other aquifers that are less strongly affected by their interaction with surface water. In the 2010–2039 scenario, water levels rise and fall with the river hydrograph at different times because of a shift in river hydrograph peak flow to an earlier date. The maximum aquifer water levels associated with the peak hydrograph are very similar to present climate because the peak discharge is not predicted to change, only the timing of it. Elevated water levels up to 30 to 40 cm persist along the channel and drain within a month. From late summer to the end of the year, water levels are similar to present conditions, with small increases observed due to the increase in recharge in areas away from the river channel. In the 2040–2069 climate scenario, the hydrograph shift is larger than in the 2010–2039 climate scenario, resulting in up to 50 cm change in groundwater levels.

Figure 16.

Water level differences (measured as head in layer 2 of the unconfined aquifer) between (left) future (2010–2039) and present climate and (right) future (2040–2069) and present climate. Maps are by time step in days 131 to 235. Contours shown are at a 0.1 m interval. Darkest blue colors indicate values <−0.5 m (along rivers only). At day 101, the difference map (not shown) has values within 0.1 m of zero.

4. Discussion

[52] In all climate change impacts studies, there is a considerable amount of uncertainty in the results. This uncertainty is most directly related to the predictions from global climate models. In groundwater modeling, there are many steps required to reach a prediction on the anticipated change in groundwater levels, and each step in the modeling process introduces uncertainty. In this study, the downscaling model bias for precipitation was significant. This uncertainty was avoided by not using the downscaled daily series directly, but rather by assuming relative changes in monthly means and applying these shifts to stochastically generated weather. The use of a calibrated weather generator ensured correct representation of precipitation seasonality, but as noted before, the choice of downscaling method nevertheless has strong effect on the directions of future changes in precipitation (i.e., the error can be the same as the magnitude of predicted change).

[53] The uncertainty of recharge estimated using any recharge model exists for most sites, largely due to uncertainty in input parameters to the recharge model (recharge was found to be most sensitive to soil and aquifer permeability, and water table depth). Nonetheless, the recharge results, when combined with realistic boundary conditions for the rivers, yielded acceptable model calibration error, with no spatial distribution in calibration bias (i.e., the results were no better and no worse for areas near and far from the river).

[54] In the groundwater flow model, uncertainty is tied primarily to aquifer heterogeneity. As mentioned earlier, climate change simulations were also carried out for a heterogeneous model, whereby the upper aquifer layer was represented not by a uniform K value, but rather a heterogeneous distribution of K values. The effects of aquifer heterogeneity on climate change impacts modeling are discussed in detail by Allen et al. [2004b]; however, we note that valley-wide results are entirely consistent with those presented here in that the same general trends are observed, except that in the heterogeneous model, there is more local variation in the predicted changes in water level.

[55] A final point of discussion is the value added by undertaking a rigorous approach to the modeling exercise. As discussed in the introduction, this study was motivated by previous work by Allen et al. [2004a], in which the sensitivity of the aquifer to future predicted climate change was assessed using a steady state flow model, with simplistic aquifer and river representation, for two extreme scenarios of climate change. High-recharge and low-recharge simulations resulted in approximately a +0.05 m increase and a −0.025 m decrease, respectively, in water table elevations throughout the aquifer. Simulated separately, were changes in river stage elevation to reflect higher than peak flow levels (20% and 50%) and lower than base flow (20% and 50%). These resulted in the most significant changes to the water table elevation (about +0.02 and −0.04 m, respectively). Overall, the results of that study are consistent with what we observed in this study despite the simple approach used; with the exception that transient effects in groundwater storage were not accounted for. The consistency of the results also suggests that even simple approaches to quantifying the impacts of climate change on groundwater have value where detailed data are lacking. Of course, this particular aquifer is highly impacted by its interaction with surface water, and the overall impacts of climate change must consider both changes in surface water hydrology and direct changes in groundwater recharge.

5. Conclusions

[56] In undertaking climate change impacts modeling, we rely on estimations of future climate as determined from imperfectly calibrated downscaling models, which themselves are from similarly uncertain results of CGCM1 climate model. For precipitation predictions at this location, the choice of downscaling method was shown to be very important for interpreting predictions of GCM models, as GCMs do not directly model precipitation locally. There are many local controls on local precipitation (elevation, rain shadow effects, distance from ocean coast, etc.), which are not captured by a GCM. Two different downscaling methods were used and compared, the SDSM model and the K-nn model. Neither adequately modeled precipitation, and differed in goodness of fit to observed precipitation in different months of a year; the SDSM model results were ultimately selected for recharge modeling. To overcome the limitations of the downscaling results, relative changes in climate parameters were applied as shifts to historical climate data within a stochastic weather generator.

[57] The HELP hydrologic model used in this study was found to be sensitive to depth of water table (percolation layer depth), soil type, and saturated hydraulic conductivity (Ksat) of the unsaturated zone. Therefore, in order to achieve accurate results for recharge, it is important to capture the spatial variability of these key variables. Results indicate that Grand Forks receives between 10% and 80% of recharge from monthly precipitation, depending on location within the valley. The spatial distribution of recharge was found to have a more significant effect on groundwater levels and flow components than temporal distribution of recharge. In this particular aquifer, the model is most sensitive to recharge away from river floodplain.

[58] The predicted future climate for the Grand Forks area from the downscaled CGCM1 model will result in more recharge to the unconfined aquifer from spring to the summer season. That said, the river water level perturbation is much more important here than recharge perturbation over the aquifer area, because of the vastly different flow rates and volumes involved (river versus precipitation recharge), but nonetheless, the overall methodology presented will allow for similar studies to be undertaken on aquifers elsewhere.

Acknowledgments

[59] Financial support for this research was provided by the Natural Sciences and Engineering Council of Canada (NSERC) in the form of an undergraduate research award to J. Scibek, Natural Resources Canada under the Climate Change Action Fund (CCAF) Program, and the BC Ministry of Water, Land and Air Protection (BC WLAP). Technical support has been provided by Environment Canada (climate change predictions for the Kettle River) and BC WLAP (water well database).

Ancillary