Significance of surface water in the terrestrial water budget: A case study in the Prairie Coteau using GRACE, GLDAS, Landsat, and groundwater well data

Authors


Abstract

[1] The terrestrial water budgets of the Prairie Coteau (PC; 38,000 km2) and Northern Glaciated Plains (NGP; 66,000 km2) regions of South Dakota, USA were characterized using a combination of in situ observations of groundwater and surface water, remote sensing estimates of terrestrial water storage changes from the Gravity Recovery and Climate Experiment (GRACE) and surface water changes from Landsat, and modeled changes in soil moisture, snow water equivalent, and total canopy water storage. In response to increased wetness in 2010 and 2011 over the region, prevalent surface water bodies accounted for a significant fraction of the terrestrial water budget of the PC, whereas the NGP, an area with sparse surface water, exhibited a greater increase in groundwater storage concomitant with enhanced seasonal variability. Over the study period from 2003 to 2011, GRACE-based estimates of terrestrial water storage agreed well with combined groundwater, soil moisture, snow water equivalent, and total canopy water storage over the NGP, as surface water is not a significant component in this area. However, closure was improved over the PC if surface water changes were included in the water budget.

1. Introduction

[2] Terrestrial water storage (TWS) defines all forms of water stored above and beneath the land surfaces of Earth, including soil moisture (SM), snow and ice, groundwater (GW), surface water (SW), and water contained in biomass [Famiglietti, 2004]. Variations in the components of TWS, individually or collectively, reflect changes in the hydrological cycle and have an immediate impact on sediment and nutrient transport and greenhouse gas emissions to the atmosphere [Famiglietti, 2004; Syed et al., 2008].

[3] Characterization of TWS and its components has been investigated in many locations worldwide using a combination of Gravity Recovery and Climate Experiment (GRACE) data, in situ measurements, and/or model simulations. Based on precise measurements of the Earth's gravity field, GRACE provides monthly variations in TWS since April 2002 [Syed et al., 2008]. Studies conducted in Illinois (∼280,000 km2) [Swenson et al., 2006] and in the Hai River basin (∼320,000 km2) in northern China [Moiwo et al., 2009] showed that GRACE TWS estimates agreed well with in situ measurements of GW and SM. Over the High Plains Aquifer (450,000 km2) in the USA, Strassberg et al. [2007] found there was high correlation between GRACE-derived TWS estimates and estimates derived from combined in situ GW data and SM data from the National Centers for Environmental Prediction/Oregon State University/Air Force/Hydrologic Research Lab Model (NOAH) of the North American Land Data Assimilation System. However, all of these studies were conducted in areas where TWS is dominated by changes in GW and SM. Other than studies in riverine systems and floodplains [Frappart et al., 2008; Han et al., 2009; Kim et al., 2009], the authors are unaware of any study conducted in an area where SW (lakes, ponds, potholes, and sloughs) variability may have a significant contribution to the TWS balance.

[4] Determining the quantity and quality of SW and GW necessary to sustain US human populations and ecosystem resilience has been identified as a top research priority for science to inform US conservation and management policy [Fleishman et al., 2011]. Despite the dominant appearance of large lakes, over 99% of global SW bodies exist as millions of scattered lakes, ponds, and impoundments with areas smaller than 1 km2, representing ∼40% of global SW area [Downing et al., 2006]. These small lakes and ponds, together with wetlands, play an important role in freshwater ecosystems and provide contributions toward water storage and water quality improvement that are disproportionately larger than their small sizes would imply [Downing, 2010; Zedler and Kercher, 2005]. However, although these small SW bodies are an integral part of the terrestrial water cycle, little is known about their contribution to the total terrestrial water budget. The Prairie Coteau (PC), a 38,000 km2 highland plateau characterized by glacial-moraine topography [Beissel and Gilbertson, 1987] and spanning 120 km wide across east-central South Dakota and 320 km in length from southern North Dakota to southeastern South Dakota [U.S. Environmental Protection Agency, 2011], represents an area with an abundance of small SW bodies (Figure 1a). Thousands of lakes and ponds (both perennial and intermittent), potholes, and sloughs cover nearly 7% of the PC land area and constitute 99.9% of its SW area (with the remaining 0.1% found in rivers, streams, and managed reservoirs) [U.S. Geological Survey (USGS) and U.S. Environmental Protection Agency (EPA), 2004]. Furthermore, 98% of these SW bodies have areal extents smaller than 1 km2, altogether constituting ∼50% of total SW area within the PC. In contrast, just 0.3% of the land area is covered with SW bodies in the surrounding Northern Glaciated Plains (NGP) [USGS and U.S. EPA, 2004], which is characterized by gently undulating to rolling glacial till plains [McNab and Avers, 1994].

Figure 1.

(a) The PC, with prevalent surface water bodies, and its surrounding NGP. (b) The aquifers and observation wells (solid circle) of the study area. The 1° × 1° cells are shown in red (over PC) and in tan (over NGP).

[5] In systematic freshwater conservation planning, accounting for connectivity between SW and GW is of critical importance [Nel et al., 2009]; accordingly, SW-GW connectivity represents an active area of research [see Bertrand et al., 2012; Dahl et al., 2007; Fleckenstein et al., 2010; Kalbus et al., 2006; Sophocleous, 2002]. In addition to containing prevalent SW bodies, the PC represents an area where recharge and discharge dynamics are hydraulically connected between SW and shallow GW aquifers. For example, lakes Kampeska and Pelican are in hydrologic connection with the Big Sioux aquifer [Hansen, 1990], Pickerel and Enemy Swim lakes serve as aquifer recharge areas [Leap, 1988], while other lakes in the vicinity of Pickerel and Enemy Swim act as aquifer discharge areas. Hydrological connections between GW and SW occur not only in larger lakes; numerous potholes and ephemeral wetlands also serve as sinks for both groundwater and SW runoff [Hamilton, 1986; Leap, 1988]. Interchange of GW and SW throughout the PC takes place through surface streams, surficial outwash, and likely through underlying deeper tills [Leap, 1988], as well as in areas of hummocky surface terrain that favor vertical flow from SW storage downward to the permeable basal aquifers [Hamilton, 1986].

[6] The GW hydrogeology of the PC and NGP (Figure 1b) is characterized largely by shallow unconsolidated-deposit aquifers of Quaternary age, consisting primarily of sand and gravel deposited as glacial outwash or stream-valley alluvium, and these are the most productive aquifers within this area [Whitehead, 1996]. Intermixed with the unconsolidated-deposit aquifers are fine-grained lake deposits and poorly sorted glacial till; these deposits have minimal permeability and commonly form local confining units. Lower Cretaceous aquifers, consisting of consolidated sandstone exposed at the land surface or buried only to a shallow depth beneath glacial deposits, are found in a narrow, discontinuous band parallel to the eastern state line of North and South Dakota.

[7] The objectives of this study were to (i) evaluate the dynamics of TWS over the PC region where GW and SW are hydrologically connected [Kume, 1985] and (ii) contrast the TWS dynamics of the PC, which is characterized by thousands of small SW bodies, to the TWS dynamics of the surrounding NGP, an area where SW is far less prevalent. Remotely sensed changes in SW areal extent were combined with topographic information to estimate changes in SW [Bjerklie et al., 2005; Brakenridge et al., 1998; Matgen et al., 2007; Smith and Pavelsky, 2009]. Volumetric changes in SW were combined with land surface model simulations and GW well observations to represent and quantify all components of TWS, with subsequent comparisons to TWS estimates from GRACE.

2. Data and Methods

[8] In general, changes in TWS can be computed as

display math(1)

[9] where SM is soil moisture, SNWE is snow water equivalent, TCWS is total canopy water storage, GW is groundwater, and SW is surface water. In this study, ΔTWS and its individual components are presented as monthly equivalent water thickness (EWT) anomalies in a time series spanning from March 2003 through December 2011; for each month, the anomaly value of a variable represents its deviation from the mean value of the respective variable over the reference period (January 2004 through December 2009). Our study area is defined as three 1° × 1° cells overlying the PC and twelve 1° × 1° cells overlying the surrounding NGP (Figure 1). Monthly values across the entire PC and surrounding NGP were calculated by taking the mean of values from these 3 and 12 cells, respectively.

2.1. GRACE ΔTWS

[10] GRACE land data were processed by Sean Swenson, supported by the NASA MEaSUREs Program, and are available at http://grace.jpl.nasa.gov. Data preprocessing included application of a destriping filter, a 300 km Gaussian smoothing filter, and a spherical harmonic cutoff filter at degree 60, with subsequent rescaling to restore much of the energy removed by these filtering processes [Swenson and Wahr, 2006]. GRACE ΔTWS has a 1° × 1° spatial resolution and was provided along with associated measurement and residual errors [Landerer and Swenson, 2012]. Uncertainties in GRACE ΔTWS are estimated to be ±6.4 cm EWT for the PC and ±4.5 cm EWT for the NGP.

2.2. ΔSM, ΔSNWE, and ΔTCWS

[11] The Global Land Data Assimilation System (GLDAS) [Rodell et al., 2004] provides monthly estimates of SNWE, TCWS, and SM at a 1° × 1° spatial resolution; the data used in this effort were acquired as part of the activities of NASA's Science Mission Directorate and are archived and distributed by the Goddard Earth Sciences (GES) Data and Information Services Center (DISC). GLDAS provides output from the Community Land Model (CLM) [Bonan et al., 2002], Mosaic [Koster and Suarez, 1996], NOAH [Ek et al., 2003], and variable infiltration capacity (VIC) [Liang et al., 1994] land surface models. The average results from the Mosaic, NOAH, and VIC simulations was used to represent ΔSNWE, ΔTCWS, and ΔSM over our study area, with uncertainty in the GLDAS simulations computed as the standard deviation of results from the three contributing simulations [Kato et al., 2007]. CLM simulations were excluded on the basis of a preliminary analysis indicating that, for our study area, simulations from CLM deviated significantly from the other three models and correlated poorly with GRACE TWS. For the purposes of this investigation, ΔSM, ΔSNWE, and ΔTCWS were summed, with the sum hereafter referred to as soil, snow, and canopy water anomaly (ΔSSCW). Total soil-column depths in the three models ranged from 200 to 340 cm; thus, SM contributions from deeper soil layers (subroot zone) were not accounted for. However, the moisture content of subroot-zone soil, varying only through gravitational drainage or diffusion, is relatively stable as compared to root-zone soil moisture. Therefore, contribution of subroot-zone soil moisture to ΔSSCW throughout our study period is limited.

2.3. ΔGW

[12] Monthly observation well data were obtained from the South Dakota Department of Environment and Natural Resources, the Minnesota Department of Natural Resources, and the Nebraska Water Center at the University of Nebraska—Lincoln. As shown in Figure 1b, observation wells spread over the entire study area, with each of four aquifer categories well represented. Well reading is related to GW volume through specific yield, a parameter that varies for different aquifers. Within each 1° × 1° cell, for a particular month, ΔGW was calculated as

display math(2)

[13] where i represents one of the four aquifer categories in the region (Figure 1b) and ΔWi, SYi, and FCi are, respectively, the median value of the well level anomaly within, the specific yield of, and the fractional coverage of, the aquifer category i. Most of the well data used in this study had missing observations in at least one of the winter months; in these cases, ΔWi was assigned a null value if there was only one available well reading. While monthly well anomalies ranged from a minimum of −305 m to a maximum of 369 m, 99% of the anomalies fell within the range from −10 to 10 m. Therefore, monthly observation well anomalies outside the range of ±10 m were classified as outliers and excluded from calculations of ΔWi. Accompanying each estimate of ΔWi, the standard deviation of ΔWi (σWi) was also estimated.

[14] Throughout the study area, there are 173 observation wells within the High Plains Aquifer, 232 observation wells within Lower Cretaceous sandstone aquifers, and 1290 observation wells within 58 unconsolidated-deposit aquifers or the surrounding glacial till that envelopes these aquifers. Gutentag et al. [1984] reported specific yields ranging from 0.12 to 0.17 for the High Plains Aquifer, with an average value of 0.15. Previous efforts to model the Big Sioux aquifer, representing 16% of the unconsolidated-deposit aquifer within our study area, used specific yields of 0.10 [Koch, 1980], 0.14 [Putnam and Thompson, 1996], 0.15 [Koch, 1980], and 0.20 [Hansen, 1988; Koch, 1982; Niehus and Thompson, 1998]. For our investigation, a value of 0.15 ± 0.03 is assumed for the specific yield for the High Plains Aquifer and 0.17 ± 0.03 for unconsolidated-deposit aquifers. Study of the Elm-Middle James-Deep James aquifer system [Marini et al., 2012], representing 1270 km2 within the northwestern portion of the NGP, indicated a specific yield of 0.10 for glacial till intermixed within the aquifer system. Since we do not know how specific yield varies within glacial till, we assumed a variability of 20% based on the estimates for the High Plains Aquifer and unconsolidated-deposit aquifers. For Lower Cretaceous sandstone aquifers, the specific yield is assumed to vary from 0.21 to 0.27, as reported in Johnson [1967] for fine-grained to medium-grained consolidated deposits; this represents a variability of 12.5% about the average value of 0.24. The uncertainty for monthly ΔGW (σGW) within each cell was estimated, similar to equation (2), as the weighted average of the standard error [Taylor, 1982] of the well readings and the variability of the specific yields:

display math(3)

[15] where σWi is the standard deviation of the well readings, σSYi is the uncertainty of the specific yield, and Ni is the number of well readings within each aquifer category i.

2.4. ΔSW

[16] Among the numerous lakes, pot holes, and sloughs that comprise SW within the PC, only lakes have water depth that is regularly monitored. The South Dakota Department of Environment and Natural Resources (DENR; http://denr.sd.gov/des/wr/dblakesearch.aspx) measures lake elevations twice per year, spring (April, May, or June) and fall (September, October, or November), as these time periods approximately correspond to the seasonal maximum and minimum, and presents monthly time series by linear interpolation between measurement points. Time series of in situ elevation data for 37 lakes from 2003 to 2011 were used to estimate ΔSW (explained below).

[17] Imagery from the Landsat 5 TM sensor (30 m resolution), obtained from U.S. Geological Survey (USGS), was used to estimate total SW changes over the PC and three adjacent 1° × 1° cells. Unfortunately, cloud coverage prevented us from obtaining corresponding imagery for all months of available in situ data. Therefore, we focused on 9 months (April 2006, October 2006, May 2007, May 2008, October 2008, May 2009, April 2010, October 2010, and October 2011) for which minimally cloud-contaminated images over the PC were available, hereafter referred to as “Landsat months.” SW bodies in the Landsat images were delineated via an unsupervised classification of mosaicked Landsat imagery, with SW elevation, SW volume, and SW EWT subsequently estimated using a combination of the classified imagery and a 30 m resolution Digital Elevation Model (DEM) obtained from USGS, similar to approaches taken by Brakenridge et al. [1998], Bjerklie et al. [2005], and Matgen et al. [2007]. Since the DEM does not represent complete SW bathymetry (e.g., Figure 2a) due to the inability to interpolate bathymetric elevations from aerial photographs used to create the DEM [Christensen and Bergman, 2005], absolute changes in SW volume and SW EWT cannot be estimated using this methodology. Therefore, changes in SW EWT are defined as deviations from the baseline month of October 2006 and are represented by δSW. October 2006 was chosen as a baseline because it represents a minimum in the seasonal water cycle; coincidentally, it is also the Landsat month when ΔSSCW + ΔGW and GRACE ΔTWS were most similar (e.g., see Figure 4). By overlaying SW areas onto the DEM, for each area classified as a SW body i, for each month t, the elevation at each SW pixel p, math formula, was assigned the median of DEM values of all shoreline pixels. Note that the median shoreline elevation values were selected in favor of means to reduce the effect of outliers, e.g., a high point on the shore or a man-made structure on the shore. The change in SW elevation from time 0 (the baseline month) to time t ( math formula; Figure 2b) was computed as:

display math(4)
Figure 2.

Example changes in SW elevation in relation to the USGS DEM from (a) a cross section view and (b) a planar view. The 30 m pixel resolution DEM is represented by the grid in the planar view. This diagram illustrates a case where SW elevation increases, but it is straightforward to apply the same method for opposite cases.

[18] where Dp is the pixel DEM value. The resultant change in SW volume for each pixel was calculated as the product of math formula and the pixel area (900 m2). To represent δSW for an individual 1° × 1° cell, the sum of SW volume change for all pixels within the cell was divided by the cell area.

[19] As illustrated in Figure 3a, our analysis indicated that the largest changes in SW elevation over the study area are concentrated in the three 1° × 1° cells overlaying the PC, with minimal changes in the adjacent cells. The only exception to this pattern is a decrease in SW elevation of more than 10 cm in the cell northeast of the PC due to changes in two managed reservoirs (Lake Traverse and Big Stone Lake). To verify the results of our analysis, in situ SW elevation changes for the 37 lakes described above (totaling 179 observations over all observed time periods) were converted into changes relative to October 2006 and the results were compared to our imagery-derived SW elevation change (Figure 3b). An ordinary least-squares (OLS) regression indicated a significant relationship between the two data sets (p < 0.0001, R2 = 0.47) and a slope of 0.94 with a 95% confidence interval of ±0.17. However, examination of residual plots for regression diagnostics revealed nine suspected outliers that were points of influence in the OLS analysis; five of these outliers belonged to a single lake (School Lake) while an additional three outliers belonged to a single time period (April 2006). A subsequent iterated reweighted least squares (IRLS) regression with a bisquare weighting function [Li, 2006], which is robust against the influence of outliers, indicated a significant relationship between the two data sets (p < 0.0001, R2 = 0.46) and a slope of 0.77 with a 95% confidence interval of ±0.19. Since neither regression analysis provides strong evidence of a slope that is significantly different than one, these analyses confirm that our calculated SW elevation changes generally agree with in situ observations with no systematic biases. Residual error within the regression models can be primarily attributed to errors associated with substitution of a DEM for dry land topography, while other compounding factors include image resolution, shoreline slopes, and height accuracy of the DEM [Alsdorf et al., 2007]. The uncertainty associated with our estimation of SW elevation change was defined as the root-mean-square of the residuals from the IRLS regression described above; the product of this uncertainty and the maximum SW areal extent observed among our nine Landsat months yielded our uncertainty estimate for SW volume change. Uncertainty in δSW was calculated as the uncertainty in SW volume change divided by the total area of the three 1° × 1° cells overlaying the PC and was estimated to be ±1.0 cm.

Figure 3.

(a) Distribution of SW elevation change over the study area, October 2006 to October 2010. Nearly the entire change is concentrated in a distinctive area of the PC. (b) For 37 lakes located within the PC, comparison of in situ SW elevation changes from the October 2006 baseline to calculated SW elevation changes.

Figure 4.

Comparison of monthly estimates of different water storage components for the study area: ΔSSCW, ΔGW, and cumulative precipitation for the (a) PC and (b) NGP; GRACE ΔTWS and combined ΔSSCW and ΔGW for the (c) PC and (d) NGP. For a given month, cumulative precipitation is the sum of that month's precipitation and the precipitation received over the prior 11 months. Shaded regions represent the uncertainties of the data sets used to derive each water storage component.

[20] Due to limited availability of cloud-free Landsat imagery, it is impossible to derive a time series of ΔSW that accounts for all the surface water bodies in the study area. Instead, the time series of ΔSW was estimated based on in situ elevation data for 37 lakes with DENR monitoring dates that fell within our Landsat months. Elevation anomaly for each of the 37 lakes was converted to a volumetric basis, summed, and divided by the total area of the PC, with the lake area estimates required for SW volume computations derived from analysis of Landsat TM5 imagery. For each of the Landsat months, surface water areal extents over the 37 lakes were directly estimated from the images. For all other months (where corresponding Landsat imagery was unavailable), each lake was assigned the area calculated for a Landsat month for which the in situ elevation was closest to the Landsat-based SW elevation. While there are uncertainties associated with the areal extents assigned by this procedure, the impact on estimates of ΔSW is negligible. For example, when ΔSW was instead calculated using the minimum and the maximum Landsat-estimated areas for the 37 lakes, respectively, the difference in EWT anomaly between the two ΔSW time series was no greater than 3 mm in any 1 month and was less than 1 mm when averaged across all months. Also note that, since areal extent of these 37 lakes represents just ∼10% of total SW area within the PC, ΔSW thus estimated would underestimate the actual changes. Nonetheless, we believe the trend in ΔSW is valid.

3. Results

[21] Figures 4a and 4b show, for the PC and NGP, monthly ΔGW, ΔSSCW, and cumulative precipitation from March 2003 to December 2011. Cumulative precipitation, representing the total precipitation of the previous 12 months and calculated using rainfall and snowfall rates from NOAH output, shows the long-term precipitation trend. Over the study period, both ΔGW and ΔSSCW exhibited clear seasonality; however, their dynamics were distinctly different between the two areas. Over the PC, ΔSSCW (Figure 4a) displayed a stronger seasonal variability (standard deviation σ = 4.3 cm) than did ΔGW (σ = 2.9 cm) from 2003 to 2009, in contrast to the NGP (Figure 4b) where the seasonal variability of the two components had a similar magnitude (SSCW σ = 4.4 cm, GW σ = 4.2 cm). In addition to the observed seasonal variabilities, both ΔGW and ΔSSCW increased in both areas since 2010, seemingly concomitant with increased precipitation. From 2003 to 2009, no clear trend was spotted for either ΔGW or ΔSSCW within their respective uncertainties; for example, the mean monthly ΔSSCW for the period was −0.2 cm for the PC and −0.3 cm for the NGP, while the mean monthly ΔGW was −0.4 cm for the PC and −0.1 cm for the NGP, all suggesting a nearly balanced seasonal cycle in water storage. From 2010 onward, the mean monthly ΔSSCW increased to 3.0 cm over the PC and 2.2 cm over the NGP, indicating a much wetter soil condition for the region. During the same time period, a more dramatic increase was observed in ΔGW, with mean monthly values of 9.6 cm over the PC and 11.6 cm over the NGP.

[22] If increased precipitation was the driver for the observed increases in ΔSSCW and ΔGW, then it is apparent that their responses were different between the two areas. The PC saw a greater increase in cumulative precipitation (10 versus 5.5 cm) than the NGP between the two time periods; in response, the increase in ΔSSCW was 24% greater than in the NGP (3.1 versus 2.4 cm). However, the increase in ΔGW was 14% less (9.9 versus 11.5 cm). We believe this suppressed response in ΔGW is largely due to prevalent SW bodies receiving and storing increased precipitation that would otherwise percolate from soil downward into GW. Furthermore, the presence of these SW bodies also explains the suppressed seasonal variability in ΔGW that was observed within the PC from 2003 to 2009 (see above). In contrast, similar seasonal variabilities in ΔSSCW and ΔGW within the NGP over this timeframe are consistent with a hydrological system with few SW bodies acting as a reservoir for increased precipitation, allowing soil water to percolate readily into shallowly underlying unconsolidated-deposit aquifers. Altogether, these results suggest that the interconnectedness of SSCW, GW, and SW within the hydrologic cycle of the PC [Hamilton, 1986; Hansen, 1990; Kume, 1985; Leap, 1988] has a measurable impact on the balance between the individual components of TWS.

[23] To further evaluate the role played by SW in the TWS budget, ΔTWS and ΔSSCW + ΔGW over the PC and NGP are compared in Figures 4c and 4d, respectively. The time series of GRACE ΔTWS and ΔSSCW + ΔGW over two time periods (before and after January 2010) were compared using two-way analysis of variance (ANOVA) models considering EWT anomaly as the response variable and data set (GRACE ΔTWS versus ΔSSCW + ΔGW), month, and year as categorical variables. All possible interactions among variables were included in the ANOVA models, and three values were used to define EWT anomaly for each month × year: our estimate of ΔEWT and the upper and lower bounds of the uncertainty for ΔEWT. From 2003 to 2009, there was no significant difference between ΔTWS and ΔSSCW + ΔGW over the PC (p = 0.41) and the NGP (p = 0.78). In contrast, from 2010 to 2011, ΔTWS was significantly greater than ΔSSCW + ΔGW over the PC (p < 0.01) but not significantly different than ΔSSCW + ΔGW over the NGP (p = 0.28). ΔSW over the PC (Figure 5), based on data from 37 lakes (see section 2.4), increased from −0.1 cm for 2003 through 2008 to 0.4 cm in 2009, 1.2 cm in 2010, and 1.7 cm in 2011. Adding the contributions from ΔSW with those from ΔSSCW + ΔGW (Figure 4c) helps to close the ΔTWS budget of the PC, but the mean value of this combined data set from 2010 to 2011 is still significantly lower than the corresponding value for GRACE ΔTWS (14.5 versus 15.9 cm, p = 0.07); this is expected, as the 37 lakes used to derive ΔSW only comprise approximately 10% of the total SW areal extent over the PC. Nonetheless, these analyses collectively suggest that SW was a significant TWS component over the PC, especially from 2010 onward.

Figure 5.

Monthly estimates of ΔSW for the PC (left axis) and SW elevation anomaly for each of the 37 lakes used to compute ΔSW (right axis).

[24] To fully account for all SW within the PC, Figure 6 compares δSW, δGW, δSSCW, and their sum (δTWS′) with GRACE δTWS for each of eight Landsat months over the PC. Same as the estimate of δSW, δ values for other variables were calculated as deviations of their values from the baseline month of October 2006. Estimates of δTWS′ were within the respective uncertainties of GRACE δTWS (i.e., the error bars for the two data sets overlap) for all but one of the eight Landsat months (May 2007). δSW over the PC accounted for −6.7%, 10%, −4.9%, and 9.5% of δTWS′ for April 2006, May 2007, May 2008, and October 2008, respectively. In contrast, corresponding values for May 2009, April 2010, October 2010, and October 2011 were 29%, 14%, 19%, and 35%, respectively, reaffirming the significance of SW within the TWS budget of the PC over the latter portion of our study timeframe. Furthermore, the inclusion of δSW reduced the root-mean-square difference between δTWS′ and GRACE δTWS from 6.5 to 6.1 cm, and from 5.4 to 3.8 cm if the Landsat month of May 2007 was excluded. Altogether, Figures 4-6 clearly illustrate the importance of SW to the TWS budget of the PC.

Figure 6.

For the PC, deviations of GRACE δTWS and of individual water storage components (δSSCW, δGW, and δSW) from the baseline month of October 2006. Note the negative δSW trends for April 2006 and May 2008; for each of these months, the sum of individual water storage components is indicated by a data point (solid circle). Error bars represent the uncertainties of the data sets used to derive each water storage component.

4. Discussion and Conclusions

[25] This study presents a comparison of TWS dynamics over the PC and the surrounding NGP, two areas with extensive networks of shallow GW aquifers but vastly differing SW hydrology. In the PC, an area containing thousands of small SW bodies, SW was a significant contributor to TWS from 2010 to 2011, serving as a reservoir for increased precipitation that would have otherwise percolated into shallow GW aquifers. Conversely, ΔGW displayed greater seasonality within the surrounding NGP, an area with sparse SW bodies, and increased greatly in response to increased precipitation in the latter portion of the study timeframe.

[26] One challenge in applying GRACE data in our study is that the areal extents of the PC (38,000 km2) and NGP (66,000 km2) are relatively small, leading to elevated uncertainty in GRACE-based estimates of ΔTWS. For example, uncertainties in GRACE ΔTWS are estimated to be ±6.4 cm EWT for the PC and ±4.5 cm EWT for the NGP, compared to reported uncertainties of ±1.8 cm and ±1.1 cm, respectively, for similar studies conducted in Illinois (∼280,000 km2) [Swenson et al., 2006] and the High Plains Aquifer (450,000 km2) in the USA [Strassberg et al., 2007]. Furthermore, increased noisiness of GRACE data may explain relatively poorer agreement between GRACE ΔTWS and combined ΔSSCW and ΔGW over the PC (Figure 4c; R2 = 0.75, RMSE = 5.4 cm) and the NGP (Figure 4d; R2 = 0.64, RMSE = 6.0 cm) when compared to reports from Swenson et al. [2006] (R2 = 0.90, RMSE = 2.0 cm), from Strassberg et al. [2007] (R2 = 0.67, RMSE = 3.3 cm), and in the Hai River basin (∼320,000 km2) in northern China [Moiwo et al., 2009] (R2 = 0.31, RMSE = 3.3 cm). Nonetheless, even with these scale-induced issues, our study has clearly shown that SW contributions are statistically significant for TWS in the PC region, where nearly 7% of the land surface is covered by small water bodies. In addition, the observed poorer agreement may also be partially due to the longer time period in our investigation, which spans approximately 9 years, compared to time periods of ≤4 years in earlier studies. For example, if considering the period from 2003 to 2005 only, the agreement between GRACE ΔTWS and combined ΔSSCW and ΔGW would improve for both the PC (R2 = 0.78, RMSE = 3.1 cm) and the NGP (R2 = 0.63, RMSE = 4.0 cm).

[27] The results presented herein also demonstrate the difficulty in fully accounting for all components of TWS over a continuous time series. Estimating SW is particularly challenging for a region, such as the PC, where SW is comprised of numerous lakes, sloughs, and potholes. The TWS closure improved when (partial) ΔSW based on 37 lakes in the region was included in the comparison shown in Figure 4c (R2 = 0.77, RMSE = 5.2 cm). The TWS closure was also improved when SW change over the entire region was estimated using Landsat images and a DEM, but cloud contamination of Landsat imagery limited the analysis to only 9 months. A possible alternative is to use Moderate Resolution Imaging Spectroradiometer (MODIS) data, which have daily coverage versus Landsat's 16 day interval of revisit. However, at 500 m, the spatial resolution of MODIS is too coarse for our study area that is characterized by a prevalence of small water bodies.

[28] In conclusion, our study demonstrates an effective methodology that combines satellite observations, model simulations, and ground observations to quantify GW and SW contributions to the regional water balance in an area where SW is dominated by thousands of small lakes, ponds, potholes, and sloughs, with many of these SW bodies in hydrologic connection with shallowly underlying GW aquifers. However, although SW contribution toward the TWS budget of the PC was successfully examined via two complementary methodologies, future investigations of SW contribution to overall TWS would be aided by accurate remote sensing of SW storage without interference from cloud cover.

Acknowledgments

[29] We thank Sergey Molodtsov, Jiexia Wu, Tatiana Molodtsova, and Eric Castle for their assistance with data acquisition and processing. We thank Felix Landerer and Victor Zlotnicki from NASA JPL for helping us to better understand the GRACE TWS products. We appreciate the three anonymous reviewers for their valuable comments that have greatly improved the manuscript.

Ancillary