1. Introduction
Natural recharge to aquifers can be broadly categorized as diffuse recharge and focused recharge (Alley et al. 2002; Healy 2010). Diffuse recharge refers to slow downward moisture fluxes over large areas engendered by precipitation, snowmelt and surface ponding. Focused or localized recharge refers to fast water percolation from discrete surface water bodies, drainage through preferential flow paths, or lateral flows from high mountains and their fractured geological formations which can absorb and store water from precipitation and snowmelt (Wilson and Guan 2004; Markovich et al. 2019). This study focuses on diffuse recharge, a major recharge mechanism for aquifers in humid to semiarid regions of the United States (Reilly et al. 2008). Such recharge mainly occurs in unconfined and semiconfined aquifers that are in direct interaction with the atmosphere (i.e., they respond to precipitation and evapotranspiration).
Information on recharge is important for sustainable groundwater development, defined as current and future groundwater use that will not cause undesirable environmental consequences (Alley et al. 1999). Over the past several decades, overexploitation of groundwater resources has led to depleted aquifers, land subsidence, groundwater contamination, decreased low flow, and seawater intrusion in many regions across the world (Robins 1998; Konikow and Kendy 2005; Kuhn et al. 2007; Rodell et al. 2009; Wada et al. 2010; Famiglietti et al. 2011; Gleeson et al. 2012; Feng et al. 2013; Döll et al. 2014; Richey et al. 2015; Jasechko and Perrone 2020). These issues have prompted certain nations and U.S. states to consider implementing groundwater management policies, such as California’s Sustainable Groundwater Management Act (https://ca.water.usgs.gov/sustainable-groundwater-management/). To comply with these laws and to develop sustainable groundwater use strategies, water managers need information on recharge including its temporal and spatial variability. In addition, climate change affects groundwater through its impacts on recharge (e.g., Green et al. 2011; Taylor et al. 2013), prompting studies to examine recharge under future climate change scenarios (Allen et al. 2004; Crosbie et al. 2013; Meixner et al. 2016; Condon et al. 2020). Diffuse recharge is also critical for quantifying submarine groundwater discharge and transport of nutrients from lands to coastal environments (Befus et al. 2017). Such efforts and applications would benefit from better understanding of the environmental controls on recharge, which to date have not been as closely examined as other land surface processes.
Measuring recharge in the field is challenging and often relies on simplified assumptions. For instance, techniques using tracers to determine residence time assume steady state and piston flow in the subsurface (Edmunds and Gaye 1994; Flint et al. 2002), which is unrealistic, especially in humid regions with frequent precipitation. Recharge estimated from groundwater dating methods represents average values over tens to thousand years, owing to the long half-life of isotopes (>5000 years for 14C; McMahon et al. 2011). Further, in situ measurements are limited in space and time, making it difficult to use them for assessing regional- to continental-scale groundwater storage changes and for understanding the impacts of climate change on recharge (Scanlon et al. 2006; Crosbie et al. 2010; Mohan et al. 2018; Moeck et al. 2020).
Land surface modeling provides a useful alternative for estimating diffuse recharge. Although they do not simulate complex subsurface processes such as lateral flow, modern land surface models (LSMs), which were originally designed for providing land surface boundary conditions for general circulation models, have complex physics for simulating evapotranspiration (ET) which, along with precipitation, is the main driver for the seasonal to interannual variations in recharge and groundwater storage (Alley et al. 2002; Eltahir and Yeh 1999; Li et al. 2015; Moeck et al. 2020). ET from LSMs has been extensively evaluated at regional to global scales and found to be effective in capturing large-scale patterns of climate variability (Mitchell et al. 2004; Mueller et al. 2011; Jiménez et al. 2011; Xia et al. 2012a). Other factors such as land cover, soil texture and topography, which affect the spatial variability of recharge (Alley et al. 2002; Moeck et al. 2020), are also represented in the models. Along with providing spatially and temporally continuous estimates, LSMs have the advantage of ensuring that the simulated water and energy budgets balance.
The objective of this study is to assess a suite of recharge estimates, derived from LSMs embedded in the North America Land Data Assimilation System (NLDAS-2, Xia et al. 2012a), the NLDAS-2 Testbed (Xia et al. 2017), and the Western Land Data Assimilation System (WLDAS; Erlingis et al. 2021), for suitability in climate analysis and sustainability studies. The ability of these LSMs to simulate soil moisture, groundwater, ET, and runoff has been extensively evaluated (Xia et al. 2012a,b, 2014, 2015, 2017; Koch et al. 2016; Nearing et al. 2016; Kumar et al. 2017, 2018; Zhang et al. 2020). Results indicate that the models perform similarly with respect to annual ET and runoff, averaged over the NLDAS domain, but they differ in certain climate regions such as the northeast, the Great Lakes region, and the western high mountains, where model formulations of snow and frozen soil processes vary significantly (Xia et al. 2012a, 2017; Zhang et al. 2020). Xia et al. (2012b) further showed that ensemble mean ET and runoff compare better with reference datasets than estimates from individual models.
LSM-simulated recharge has not been extensively evaluated, in part, because, until recently, most LSMs did not simulate groundwater storage. In this study, we consider drainage from the deepest soil layer (termed “subsurface runoff” or “base flow”) of the models that lack a groundwater formulation to represent recharge. Applying the same logic, Niraula et al. (2017) appraised recharge estimates from three earlier versions of NLDAS-2 models for the western United States and found they were consistent in identifying high and low recharge zones but differed in terms of the magnitude of recharge. Driven by a common forcing dataset, this suite of models enables examination of how model physics affect recharge, which is not well understood over a diverse set of climates. In addition, the long data record (1979–present) allows us to study how recharge responds to climate change such as changes in precipitation and temperature (Taylor et al. 2013; Meixner et al. 2016). In recent years, hydrological models have been employed to project changes in recharge under future climate change scenarios (e.g., Crosbie et al. 2013; Condon et al. 2020). In addition, recharge estimated from LSMs has been used to drive dynamic groundwater models to simulate lateral groundwater flows (Maxwell and Miller 2005; Hein et al. 2019). Understanding the limitations and uncertainties of LSM-simulated recharge should be useful to such studies.
Herein we assess uncertainties in simulated monthly and annual recharge through model intercomparison and evaluation against independently derived recharge information. We examine the timing of seasonal maximum soil moisture and groundwater storage using in situ soil moisture and well data to infer the day of seasonal maximum recharge and its irregularity. Long-term trends were calculated to explore the relationships between recharge and climate variables (precipitation and temperature). We describe the data and the methods for deriving recharge from LSMs in section 2 and present the results in section 3. Section 4 contains discussions of model deficiencies in simulating soil moisture and groundwater and the implications for climate studies, as well as suggestions for future research and model improvements. Conclusions and recommendations for using these recharge estimates for hydrological applications and climate analyses are provided in section 5.
2. Data and methods
a. Models
1) NLDAS-2 models
NLDAS-2 comprises four LSMs: Noah, Variable Infiltration Capacity (VIC), Mosaic, and Sacramento (SAC). All four models simulate land surface states and fluxes (e.g., soil moisture and snow; ET and runoff) but with different parameterizations (e.g., soil layering) and physics (Table 1). Noah simulates dynamic soil moisture in four soil layers with thicknesses of 10, 30, 60, and 100 cm, respectively, based on the Richards’ equation (Chen et al. 1996). Pedotransfer functions (Cosby et al. 1984) were used to obtain soil hydraulic parameters, such as hydraulic conductivity, based on soil textures from the State Soil Geographic (STATSGO) dataset (Mitchell et al. 2004).
Model configurations and algorithms for soil moisture, groundwater, snow, and runoff (Mitchell et al. 2004; Xia et al. 2012a, 2017).
Unlike Noah, VIC and Mosaic simulate soil moisture using empirically derived parameters (different for each model) (Table 1; Liang et al. 1994; Koster and Suarez 1996). The depths of the two spatially varied deep soil layers in VIC were determined through calibration using streamflow data (Liang et al. 1994). For the version of VIC used in this study, calibration was performed using NLDAS-2 forcing data and streamflow data from more 1000 small basins in the United States (Xia et al. 2012a). Noah, VIC, and Mosaic account for subgrid heterogeneity in soil moisture when partitioning precipitation into surface runoff and infiltration, but using different approaches. All three models simulate the components of ET—vegetation transpiration, bare soil evaporation, and direct evaporation from canopy interception and snow—but with algorithms that differ, including in their dependency on soil moisture and the root-zone profile. For detailed descriptions of these ET and runoff algorithms, we refer to Mitchell el al. (2004), Xia et al. (2012a,b), and Zhang et al. (2020), which also present evaluations and intercomparisons of simulated fluxes and states.
SAC simulates water storage changes in six conceptual water storage components, representing upper and lower zones and areas near streams. When implemented in NLDAS-2, these storage components were mapped to Noah’s soil depths and Noah’s ET algorithms were implemented in SAC to calculate ET internally (Xia et al. 2012a). Although the SAC model used by the National Weather Service River Forecast Centers (RFC) is calibrated for each region, the version used in this study was not calibrated and instead used standard parameters (Xia et al. 2012a).
NLDAS-2 models simulate snow using simple schemes, mostly with a single snow layer (Table 1), yet they employ significantly different physics. For instance, VIC contains subgrid elevation banding which helps snow persist longer in complex terrains (Xia et al. 2012a). SAC uses a temperature index-based model to simulate snow independently and to bypass the energy balance equations used in other models (Xia et al. 2012a). Unlike the other models, SAC does not simulate snow sublimation. Snow simulation in Noah was improved significantly in NLDAS-2, with representation of snow density changes and fractional snow cover and the use of time-dependent maximum snow albedo (Xia et al. 2012a).
The four models also differ in how they simulate surface and subsurface runoff (Table 1). Noah calculates surface runoff use the Simple Water Balance model, which estimates infiltration (and thus surface runoff) based on point-scale infiltration theory with consideration of spatial variability of precipitation rates and soil hydraulic parameters (Schaake et al. 1996). Noah estimates subsurface runoff based on the free drainage condition, which is represented by hydraulic conductivity and assumes unlimited drainage capacity from the lowest soil layer. While VIC and Mosaic simulate runoff using topography-oriented schemes, SAC, as a rainfall–runoff model, simulates runoff based on theories of point processes with more complex parameters (Schaake et al. 1996).
NLDAS-2 models do not simulate groundwater, but their subsurface runoff or baseflow runoff output variables (Fig. 1a), representing drainage from the lowest soil layer, are essentially recharge, assuming there are no significant water storage changes in the zone from the lowest soil layer to the water table (Fig. 1b). Capillary rise from the water table is also neglected.
Schematic diagrams for recharge simulated by land surface models (a) without (Noah, VIC, Mosaic, and SAC) and (b) with (Noah-MP and CLSM) a representation of groundwater storage, and (c) recharge estimated by the USGS water balance approach. S1–S3 represent soil layers in LSMs in general (see Table 1 for soil layer configuration of each model). Dashed lines and gray text indicate processes and zones that are not represented by the models.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
Although not calibrated, some of the parameters used for Noah, Mosaic and SAC were adjusted or tuned based on previous evaluation and comparisons (Table 1; Xia et al. 2012a). For instance, several parameters used in Noah’s ET schemes were adjusted in order to reduce biases in simulated states and fluxes observed in NLDAS-1 Noah (see details in Xia et al. 2012a).
2) NLDAS-2 Testbed models
The NLDAS-2 Testbed was designed to facilitate intercomparison of model processes including groundwater (Xia et al. 2017). It consists of three models, two of which, the Noah LSM with multiparameterization options (Noah-MP) and the Catchment land surface model (CLSM), were used in this study. Noah-MP uses a bucket-type linear reservoir to simulate groundwater storage changes based on the influx (recharge) and outflux (capillary rise and base flow, Fig. 1b), without accounting for hydraulic properties such as hydraulic conductivity, which is used in soil moisture (and therefore recharge) simulation. Water table variations are estimated by dividing groundwater storage changes by a uniformly prescribed specific yield, 0.2 (Niu et al. 2007). This highly simplified assumption is employed due to the unavailability of reliable, large-scale maps of specific yield. In reality, specific yield varies with geologic material and depth (Li and Rodell 2015). The depth to the water table is then used to determine base flow (groundwater discharge to rivers) using the TOPMODEL formulation (Table 1, Fig. 1b). Because the model does not simulate groundwater dynamics (and hydraulic head), capillary rise is estimated based on the hydraulic gradient across the lowest soil layer, rather than that at the water table (Niu et al. 2007).
Noah-MP employs more sophisticated physics for simulating snow and frozen soils than other models. Snow accumulates in up to three layers (Table 1), and snow temperature is computed with soil temperature using a unified heat-conductance equation (Niu et al. 2011). The unified temperature profile is then used to determine heat transfers associated with snowmelt or freezing soils. The module also simulates snow compaction, snow interception, and loading and unloading of snowfall based on wind and canopy temperature. Other improvements include a dynamic (as opposed to climatological mean) vegetation scheme for calculating the leaf area index (Niu et al. 2011) and a semi-tile approach to calculate energy fluxes (latent, sensible, and ground heats) separately on vegetated and bare soil fractions of a grid cell to avoid overexposing grass or snow-covered lands to solar radiation (Niu et al. 2011).
Similarly, CLSM was developed to improve upon Mosaic by representing the full soil moisture profile, including shallow groundwater, and applying the TOPMODEL concept for simulating runoff (Koster et al. 2000). Water storage changes are simulated in a 2-cm surface layer, a root zone (1 m), and a soil profile whose depth depends on a bedrock depth parameter. Following Houborg et al. (2012), who showed that the original bedrock depths were unable to accommodate large water storage losses during severe droughts in the southwestern United States, the bedrock depth used in this study was increased uniformly by 2 m (to preserve the spatial variability of the original data). The decision to add 2 m (not more) was rooted in the desire to minimize unintended impacts on simulated fluxes (Houborg et al. 2012). Because the root-zone depth is fixed at 100 cm (Table 1), this change effectively increased the dynamic range of simulated groundwater storage. Moisture transfers among the three water stores (surface and root-zone soil moisture and soil profile water storage) were derived empirically based on average profile behaviors. CLSM retained the ET algorithms of Mosaic. CLSM does not explicitly simulate water table variations. Instead, water table depths are determined based on profile moisture and other model parameters for calculating base flow using the TOPMODEL concept (Koster et al. 2000).
CLSM simulates stronger capillary rise than Noah-MP in warm months which leads to large seasonal variations in its simulated groundwater storage (Xia et al. 2017). Neither model incorporates physically based subsurface properties (such as hydraulic conductivity) and in situ water table depths. Initial conditions were derived by spinning up the model with repeated loops through a multiyear forcing dataset until equilibrium is reached for the modeled states (Xia et al. 2017), a commonly adopted approach for land surface modeling (Rodell et al. 2005; Niu et al. 2007).
Because the groundwater schemes are so simplified, these models are more suited for simulating regional groundwater storage changes associated with large-scale surface meteorology than point-scale variations. Previous studies have shown strong correlations between Noah-MP and CLSM simulated groundwater storage and in situ observations at regional scales in the central and eastern United States (mostly R2 > 0.6; Li and Rodell 2015; Xia et al. 2017) and globally for CLSM (mostly R2 > 0.5; Li et al. 2019b). The same studies reported low correlations at point scales, due to the strong influence of subsurface properties and the limitation of coarse-scale forcing data.
Because the two models do not simulate moisture changes in the transition zone, the estimates based on Eq. (1) represent drainage from the bottom of the lowest soil layer (200 cm for Noah-MP and 100 cm for CLSM, Table 1), instead of influx to the water table (Fig. 1b). Hence, these are similar to the recharge estimates from the NLDAS-2 models except that they may be influenced by capillary rise (Figs. 1a,b). Since the versions of LSMs do not simulate surface water and lateral water exchanges, focused recharge from surface water bodies, preferential flow paths, and groundwater lateral flows is neglected (Fig. 1b).
3) WLDAS
WLDAS was recently developed to provide hydrological information for the western United States, where such information is highly valued due to the limited availability of both water and water data (Erlingis et al. 2021). WLDAS is an instance of NASA’s Land Information System (LIS; Kumar et al. 2006) software driving the Noah-MP model at 1-km spatial resolution, encompassing the area west of the Mississippi River. The high spatial resolution is necessary to accommodate complex terrains where both atmospheric forcing fields and land surface processes such as snowmelt are sensitive to changes in elevation. It also makes the WLDAS output more useful to operational stakeholders and enables integration of high-resolution remote sensing observations including those from NASA’s Airborne Snow Observatory (https://www.jpl.nasa.gov/missions/airborne-snow-observatory-aso/; Painter et al. 2016). WLDAS uses the 1-km University of Maryland (UMD) land cover and State Soil Geographic (STATSGO) datasets for land cover and soil texture (Mitchell et al. 2004), respectively, which are also used by other models, and is referred to as Noah-MP_1km in this study.
b. Forcing data
NLDAS-2 forcing data were used to drive NLDAS-2 and NLDAS-2 Testbed models from 1979 to 2017. NLDAS-2 precipitation is derived by temporally downscaling daily gauge observations using Stage II hourly radar observations (Xia et al. 2012a). The shortwave radiation field is based on NOAA National Centers for Environmental Prediction North American Regional Reanalysis (NARR) with bias correction using the UMD Surface Radiation Budget (SRB) dataset. Other fields (pressure, air temperature, wind, longwave radiation, and humidity) are also derived from NARR with adjustments using the NLDAS elevation data (Cosgrove et al. 2003). NLDAS-2 forcing fields, provided at hourly and 0.125° spatial resolution, cover the conterminous United States, southern Canada, and northern Mexico.
The Noah-MP_1km (WLDAS) simulation was driven by downscaled NLDAS-2 forcing fields. Precipitation was spatially downscaled using the 800-m PRISM precipitation climatology (Daly et al. 2008). Surface pressure, air temperature and relative humidity were downscaled using a uniform lapse rate, 6.5 K km−1, while the radiation fields were downscaled using slope and aspect ratios derived from a 1-km elevation dataset (Kumar et al. 2013). These downscaling procedures and all model simulations were performed using LIS and its preprocessing tools (Kumar et al. 2006; Arsenault et al. 2018).
c. In situ soil moisture
In situ soil moisture measurements from 2007 to 2017 were obtained from the Soil Climate Analysis Network (SCAN, https://www.wcc.nrcs.usda.gov/scan/) at 83 locations (Fig. 2) where soil moisture is measured at 5-, 10-, 20-, 50-, and 100-cm depths. Because diffuse recharge, driven by gravity, is closely related to soil wetness, we used these measurements to evaluate the day of mean maximum recharge. Model simulated soil moisture was interpolated to the SCAN soil moisture depths for this evaluation.
SCAN soil moisture sites (dots), color-coded based on site longitudes, in the western, central, and eastern regions (separated by vertical blue dashed lines) and USGS wells (triangles). Also shown are the 12 RFC regions: Northwest (NW), Missouri Basin (MO), North Central (NC), Ohio (OH), Northeast (NE), Mid-Atlantic (MA), California–Nevada (CN), Colorado Basin (CO), Arkansas Basin (AR), West Gulf (WG), Lower Mississippi Basin (LM), and Southeast (SE).
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
d. In situ groundwater
Figure 2 shows 181 wells located in the four Mississippi River subbasins and four northeastern U.S. regions (Long Island, New Jersey, Pennsylvania, and Massachusetts). These wells were selected from the U.S. Geological Survey (USGS) Groundwater Watch website for use in previous studies based on three criteria (Rodell et al. 2007; Li and Rodell 2015). First, the water level measurements must show seasonal variability, indicating the wells are located in unconfined or semiconfined aquifers where groundwater responds to atmospheric processes (precipitation and ET). Second, time series that show obvious signs of influences of pumping or injection are excluded. Third, the wells must have been continuously monitored for 10–30 years with at least one measurement per month most of the time (Li and Rodell 2015). Depth to water table measurements were converted to water levels relative to a reference height which were further converted to groundwater storage using specific yield estimated for each well individually (Li and Rodell 2015). Temporal variability of these groundwater observations and its relationship with climate and other environmental factors were examined by Li et al. (2015) and the same observational dataset has been used for evaluating the interannual variability of Noah-MP and CLSM simulated groundwater storage in previous studies (Zaitchik et al. 2008; Li et al. 2015; Kumar et al. 2016; Xia et al. 2017; Li et al. 2019a,b). This study builds on those by examining seasonal variations of simulated and observed groundwater storage.
e. USGS recharge
The USGS also provides a monthly recharge dataset (Reitz and Sanford 2019), developed similarly using the water balance approach of Eq. (2) but with monthly water budget inputs. Some of the input data are different from those used in the annual recharge dataset. Monthly ET estimates are derived by combining ET estimates from the water balance approach with remote sensing-based ET estimates (Reitz et al. 2017b). In addition, snow water equivalent (SWE) estimated from the Snow Data Assimilation System (SNODAS) is used for estimating snow and snowmelt in the monthly recharge dataset. Although this is still an experimental product (Reitz and Sanford 2019), it is useful for evaluating seasonal variations of simulated recharge from the LSMs. Effective recharge from 2004 to 2014, when full years of data are available, is used in this study.
One major difference between recharge estimates from the LSMs and from the USGS mass balance approach is that the latter [Eq. (2)] neglects soil water storage changes. Therefore, the USGS estimates actually represent net recharge to the entire soil profile (Fig. 1c). Nevertheless, we consider the USGS data to be an objective reference, as they incorporate county level irrigation data and surface runoff estimated from streamflow gauge data. Note that VIC is the only model that has been calibrated (Table 1) using gauged runoff data. In addition, irrigation, which LIS has the ability to simulate (Ozdogan et al. 2010; Lawston et al. 2015; Nie et al. 2018), was not represented by the LSMs in this study. Focused recharge may be a component of the USGS recharge estimates in regions where groundwater discharges to river systems measured by USGS gauges, considering this discharge depends in part on groundwater levels, which are affected by all types of recharge, including focused recharge. For instance, hydrograph analysis may estimate low surface runoff in a humid region with karst conditions if base flow dominates the observed streamflow.
f. Calculating differences in the timing of seasonal maxima
To determine the timing of seasonal maximum soil moisture, daily soil moisture (smoothed with a 30-day running average) at each SCAN soil moisture depth (5, 10, 20, 50, and 100 cm) from 2007 to 2017 (when most SCAN data are available) was used to derive mean, daily seasonal cycles of soil moisture at each location, from which the days of maximum soil moisture, ranging from 1 to 366, were determined.
3. Results
a. Annual recharge
Mean annual recharge from the LSMs exhibits spatial patterns that are similar to those of the USGS recharge map, derived from the annual recharge dataset, during 2000–13, with the largest rates of recharge in the eastern United States, the Rockies, and the Pacific Northwest (Fig. 3). However, recharge estimates in these areas vary significantly among the models, which can be clearly attributed to differences in simulated ET, runoff and snowmelt. For instance, Noah and VIC estimate more recharge than Mosaic and SAC (and the USGS) in the eastern United States, where they simulate less ET than the latter two models (Xia et al. 2012a). Mosaic and SAC generate the lowest recharge across the United States, which is consistent with their relatively high ET (Xia et al. 2012b) and SAC’s copious surface runoff (Lohmann et al. 2004), both of which decrease the amount of water available for recharge. CLSM produces less recharge in the Rockies and Pacific Northwest, likely because it simulates less snow (Xia et al. 2017) and hence generates less snowmelt. Annual recharge from Noah-MP_1km is similar to that from Noah-MP except in the lower Mississippi River where Noah-MP_1km annual recharge is lower. Noah-MP_1km simulates more vegetation transpiration (and hence less recharge) than Noah-MP in the lower Mississippi where ET is high overall and sensitive to model physics and parameters (Zhang et al. 2020). Note that the maps for Noah-MP, Noah-MP_1km, and Noah (Fig. 3) include an area of greater recharge in the Sand Hills region of western Nebraska, where the sandy soils enhance the simulated drainage rates.
Mean annual recharge (mm), correlation, RMSE (mm), and KGE (the color bar reflects the fact that −0.41 represents the skill of using mean observations for prediction and the arrow indicates that values < −1.4 are included) of the LSMs and the ensemble mean recharge with respect to (top) USGS annual recharge estimates for 2000–13. Numbers in parentheses represent domain averages.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
Annual recharge from the LSMs generally correlates well (R2 > 0.5) with that from USGS, particularly in moist regions (Fig. 3). This can be attributed in part to NLDAS-2 and PRISM precipitation data (used to force the LSMs and derive the USGS recharge, respectively) both relying heavily on gauge data and thus being temporally consistent. Further, NLDAS-2 precipitation was bias corrected using PRISM’s precipitation climatology (Xia et al. 2012a). The correlation between modeled and USGS estimated recharge is low (R2 < 0.1) in the dry western United States, where in simulated runoff and snow are more error prone (Xia et al. 2012a,b). Recharge of irrigation water is included in the USGS estimates but not reckoned by the LSMs, hence the correlation is especially low in California’s Central Valley and the southern Great Plains, where withdrawals for irrigation have led to groundwater depletion (Famiglietti et al. 2011; Scanlon et al. 2012). In the Great Plains, focused recharge from playas is considerable (Gurdak and Roe 2010), but it is not simulated by the LSMs, nor is it likely captured by the USGS approach which relies on streamflow to obtain information on groundwater storage changes. Groundwater, which is typically deep in the Great Plains, is likely a minor contributor to streamflow. Therefore, the low correlation in the Great Plains is mainly due to lack of irrigation in the LSMs.
Uncertainty in the NLDAS-2 and PRISM precipitation data may also contribute to the low correlation between simulated recharge and USGS recharge estimates, especially in the dry western United States where convective storms with high spatial and temporal variability are difficult to observe with limited gauge coverage. Radar quantitative precipitation analyses, which are used for disaggregating daily precipitation in NLDAS-2, are also subject to higher uncertainties for convective storms than stratified systems (Ochoa-Rodriguez et al. 2019).
Root-mean-square error (RMSE) of simulated recharge exhibits spatial patterns similar to those of mean annual recharge because RMSE is proportional to the magnitude of recharge (Fig. 3). That is, RMSE would be doubled if both the simulated and USGS recharge time series were multiplied by 2. This explains the low RMSE in dry climates and makes RMSE less useful for comparing uncertainty in recharge in dry climates with that in other climates.
The Kling–Gupta efficiency (KGE), which incorporates metrics of correlation, bias and variability, with 1 representing a perfect prediction by the model (Gupta et al. 2009), is low for all models except SAC in the dry interior west where correlation between simulated and USGS estimated recharge is low. Note that KGE = −0.41 represents the skill of using mean observations for prediction (Knoben et al. 2019). The KGE scores for SAC recharge are at or slightly above −0.41 across most of the United States. This is because SAC recharge estimates are close to mean annual recharge each year and insensitive to changes in precipitation. Among the models, CLSM and Noah-MP performed best in terms of KGE in large areas of eastern and north-central United States. However, CLSM and Noah-MP earned low KGE scores in parts of the central and southwestern United States. Note that Noah-MP_1KM KGE scores were nearly identical to those of Noah-MP in the domain of the former.
The ensemble mean recharge scored the most positive KGE values, especially in the high mountains of the western United States where individual model skill was typically low. This result is consistent with the conclusions of Xia et al. (2012b), who showed that ensemble mean runoff and ET compared better with reference datasets than the estimates from individual NLDAS-2 models.
As with other statistical metrics, the ensemble mean recharge has the lowest biases in most RFC regions (Table 2). Analysis of the biases does not reveal dependency on climate or model physics, rather, it implies that there are multiple controls on recharge and reflects the fact that discrepancies in other simulated processes also vary spatially. VIC simulated recharge compares well with USGS estimates in the Northwest and Colorado RFCs, where surface runoff from complex terrains may be better simulated by its topography-based runoff algorithm, and also owing to its streamflow calibration. Based on the ensemble mean recharge, the percentage of precipitation that becomes recharge ranges from about 10% in dry climates such as in the Missouri (MO) and West Gulf regions to greater than 30% in wet climates such as in the Northwest and Northeast regions.
Mean annual recharge (mm), unsigned relative bias of annual recharge with respect to the USGS annual recharge dataset (in parentheses), and mean annual NLDAS-2 precipitation in the 12 RFC regions (see Fig. 2 for region abbreviations) during 2000–13.
Given that the two USGS recharge datasets were developed using different input data, we also computed annual recharge from the monthly UGSS recharge dataset for comparison with simulated recharge (Fig. S1 in the online supplemental material). The mean annual recharge derived from the monthly USGS dataset is higher than that from the annual dataset across the entire domain. Consequently, RMSE and KGE, which are sensitive to the magnitude of recharge, differ significantly from those in Fig. 3. The correlation is generally lower across the United States, more so in California’s Central Valley and the southern Great Plains. The remote sensing infused ET data used in the USGS monthly recharge estimates are sensitive to irrigation (Reitz et al. 2017b), further accentuating the LSMs’ lack of irrigation simulation in those areas. Overall, evaluation based on the monthly USGS recharge dataset does not support the ensemble mean as the most skillful dataset in estimating recharge. This outcome underscores the strong impact of ET and snowmelt estimates, which differ between the two USGS datasets, on recharge and uncertainties in the USGS recharge estimates.
b. Seasonality of recharge
Mean monthly recharge during 1979–2017 from Noah-MP, CLSM, and Noah shows that regions with substantial annual recharge, namely, the eastern United States, the Rockies, and the Pacific Northwest, receive most recharge in cold months (Fig. 4), exemplifying the control of seasonal variations of ET on recharge. Compared to Noah-MP, Noah simulates more recharge in eastern Canada and the northeastern United States from June to October. This may be attributed to the free drainage condition (unlimited drainage from the lowest soil layer) used by Noah (Mitchell et al. 2004). Noah-MP simulates a shallow water table (not shown) in this area, which limits the drainage capacity. CLSM recharge shows weak seasonality with less recharge than Noah-MP and Noah in the eastern United States and Pacific Northwest between January and May (Fig. 4), likely due to its underestimated snowpack which leads to less snowmelt in spring (Xia et al. 2017).
Mean monthly precipitation (mm) and mean monthly recharge (mm) from Noah-MP, CLSM, and Noah for 1979–2017.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
Simulated recharge shows strong seasonal variation even in RFC regions where precipitation does not have strong seasonality such as in the Northeast (Fig. 5). Seasonal maximum recharge mainly occurs between February and June, but the timing varies greatly across the models in the northern RFCs such as the Northeast and Northwest, where large uncertainties in simulated snow have been reported (Xia et al. 2012a, 2017). Seasonal recharge in Noah-MP_1km lags that in Noah-MP in the North West. This can be attributed to high-resolution temperature data used in Noah-MP_1km that causes less snowmelt in early spring and hence more snowmelt later in the season. The amplitude of seasonal variation also varies substantially across the models, with the smallest seasonal variations observed in SAC and Mosaic, which simulate less recharge than other models.
Time series of mean monthly precipitation (mm, top bars) and monthly recharge (mm) from the LSMs and USGS in the 12 RFC regions.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
The monthly USGS recharge estimates exhibit similar seasonal variations in most RFCs, but the USGS winter and early spring maxima are often larger than those of the LSMs. This is especially true in the northern RFCs and those containing high mountains, such as Northwest, the Missouri Basin, California–Nevada, and the Colorado Basin. This discrepancy stems from differences in how snow and snowmelt are represented in the LSMs and in the USGS method. The higher spring recharge by USGS may be caused by overestimation of snow by SNODAS which has been reported in the Sierra Nevada and western Canada (Bair et al. 2016; Lv et al. 2019). In addition, not accounting for soil moisture storage causes USGS recharge to peak earlier than simulated. These results again suggest uncertainties in the USGS recharge estimates.
c. Seasonal maximum recharge
1) Evaluation using in situ soil moisture
The strong seasonal variations in recharge are consistent with those observed both in situ and simulated soil moisture and groundwater (e.g., Eltahir and Yeh 1999; Wu et al. 2002; Xia et al. 2014, 2015; Li et al. 2015). Studies have also shown that shallow hydrological processes generally peak earlier in the year than deep processes. For example, Eltahir and Yeh (1999) used in situ data in Illinois to show that root-zone soil moisture leads groundwater storage, which in turn leads streamflow. Li et al. (2015) showed that maximum seasonal recharge (estimated by VIC) leads in situ groundwater storage changes in the four subbasins of the Mississippi River and in the four northeastern U.S. regions shown in Fig. 2. These results suggest that we can examine the timing of seasonal maximum soil moisture and its relationship with depth to infer the timing of seasonal maximum recharge. For this purpose, we calculated the day of maximum SCAN and simulated seasonal soil moisture and their differences as described in section 2f, such that positive/negative differences indicate SCAN lagging/leading modeled soil moisture.
At shallow depths (5–20 cm), simulated seasonal maximum soil moisture may lead or lag that of SCAN soil moisture, depending on the model (Fig. 6) and its seasonal variations in ET (Xia et al. 2012a; Zhang et al. 2020). For instance, SAC seasonal maximum soil moisture leads that of other models at 5–20-cm depth, consistent with its seasonal ET which, on average, leads that of other models (Xia et al. 2012a). At 50 and 100 cm, simulated seasonal soil moisture leads that of SCAN for all LSMs, suggesting improperly modeled controls on deep soil moisture. Note that soil depth in VIC varies spatially and does not always reach 100 cm. This is why fewer data points are shown in the 100-cm panel of Fig. 6.
Differences (y axis) between the day of seasonal maximum soil moisture from SCAN and the LSMs at the five depths of SCAN stations (x axis). Averaged differences in days across SCAN sites are also provided. Colors indicate longitudes of SCAN sites as shown in Fig. 2.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
Because recharge occurs at the water table, which is usually deeper than 100 cm (Li et al. 2019b), we averaged the differences shown in Fig. 6 for each model at each depth and extrapolated them through linear regression with respect to depth to 2 m, where shallow groundwater is common in a wet climate like in Illinois (Eltahir and Yeh 1999). The result suggests that simulated recharge, on average, peaks 1–3 months earlier than in reality, depending on the model (top-left panel of Fig. 7). Similar regression analysis was also performed in the western, central and eastern regions (see Fig. 2 for region boundaries). The lines, representing the inferred linear relationship, are more consistent in the east region than in the other two regions (top-right three panels of Fig. 7). This reflects the fact that precipitation is the dominant control on recharge in wet climates (Moeck et al. 2020) and that these models were forced by a common precipitation dataset. In the drier west and central regions, ET exerts more control on recharge and thus, different ET estimates can lead to large spread among the regression lines. Note that regression analysis was not performed for Noah-MP_1km and VIC in the eastern region where the former was not run and the latter lacks deep soil moisture. Regression analysis based on data at the two lowest soil depths, where all simulated soil moisture peaks earlier than SCAN soil moisture, yielded different slopes for each model (bottom row of Fig. 7). While the spread of regression lines is generally smaller, the average difference at 2 m remains similar to that based on data of all depths.
Regression lines for averaged differences in the day of maximum mean daily soil moisture (top) as a function of five soil depths and (bottom) as a function of the lower two soil depths. Numbers in parentheses represent the slopes of the regression lines. Statistically insignificant slops are indicated by “ins.”
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
2) Evaluation using in situ groundwater
Groundwater seasonality provides additional insight on the uncertainty of simulated recharge. Previous studies have shown that region-averaged groundwater time series from Noah-MP and CLSM generally correlate well (R2 > 0.5) with in situ groundwater data at the interannual scale (Xia et al. 2017; Li et al. 2019b). Here we examine the timing of seasonal maximum groundwater storage simulated by those two models at individual well locations.
The mean depth to the water table has been shown to exert control over groundwater temporal variability and recharge (Bloomfield et al. 2015; Cao et al. 2016; Moeck et al. 2020). As Noah-MP and CLSM recharge is estimated at the lowest soil layer, 200 cm, and 100 cm (Fig. 1b), respectively, it is reasonable to speculate that simulated groundwater may peak earlier in the year than in situ groundwater, particularly when in situ groundwater is deep. However, seasonal groundwater storage from Noah-MP, on average, peaks during the same month as in situ observations (Fig. 8) in Massachusetts, where the mean water table is the shallowest (4 m), and in Pennsylvania, where the average water table depth is much deeper (10 m) (Li and Rodell 2015). One explanation is that subsurface properties and their heterogeneity also influence the response time of in situ groundwater to atmospheric effects. Most wells in Pennsylvania are located in fractured sandstone, shale and crystalline rock aquifers [based on metadata of these wells; the USGS well numbers are provided in Li et al. (2019b)] where fractures allow infiltrated water to reach the water table faster than in other types of media. These subsurface properties are not reflected in the hydraulic parameters used in Noah-MP, which were derived based on surface soil texture. The bedrock depth in CLSM was derived based loosely on geological conditions, but it was mainly tuned for simulating climate impacts such as spatial variability of annual ET (Koster et al. 2000; Li et al. 2019b).
Differences in the month of seasonal maximum groundwater from the wells and the LSMs as a function of in situ groundwater depth in the four northeast regions and the four Mississippi subbasins. Regionally averaged differences are also provided.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
In Ohio, simulated groundwater, on average, lags in situ groundwater by one month, consistent with seasonal snow simulated by the two models which also lags an independent dataset by 1 month in the region (Xia et al. 2017). In general, the differences in the timing of seasonal maximum groundwater do not correlate with depth to groundwater as suggested by the analysis on seasonal maximum soil moisture (Fig. 7).
To further investigate these results, we examine the timing of maximum SCAN 100-cm soil moisture and the timing of maximum USGS groundwater (Fig. 9). In situ groundwater generally peaks between March and June, except at a few wells in New Jersey and Massachusetts where groundwater peaks later in the year. Shallow groundwater in New Jersey and Massachusetts may respond to increased precipitation in winter (Fig. 5). SCAN 100-cm soil moisture peaks at different months of the year than USGS groundwater, even where the wells and SCAN sites are nearly collocated such as in the lower Mississippi River basin. In the central United States where both in situ soil moisture and groundwater measurements are found, seasonal maximum soil moisture at 100 cm often lags maximum groundwater storage. These results suggest that groundwater, at the point scale, and deep soil moisture may be more strongly influenced by other environmental factors than by meteorological conditions. In particular, aquifers may be recharged through other mechanisms, such as preferential flows and from surface water, that do not occur at SCAN sites which are mostly located in agricultural areas with unconsolidated soils favoring diffuse recharge. Further, SCAN sites and UGSS wells are not collocated and they may be influenced by different subsurface hydrogeological conditions.
The month of seasonal maximum SCAN soil moisture at 100 cm (dots) and that of seasonal maximum groundwater at the USGS wells (triangles) with the inset for the northeastern United States.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
d. Interannual variability of recharge
The Mann–Kendall trends in simulated annual recharge during 1979–2017 exhibit spatial patterns that are largely consistent across the seven models (Fig. 10). These include significant increasing trends in southern Canada and decreasing trends in the lower Mississippi, the Tennessee Valley, and large areas of the western United States, in responses to precipitation trends in these areas (Fig. 10). However, the magnitudes of the recharge trends vary substantially among the models, with generally smaller trends in models that simulate less annual recharge. For instance, SAC, which simulates relatively low mean annual recharge (Fig. 3), also estimated smaller magnitudes of trends in recharge than other models. In the lower Mississippi and Tennessee Valley, where precipitation shows insignificant decreasing trends, most models simulated significant decreasing trends in recharge. This can be contributed to the low-pass-filtering effect of soil moisture and groundwater, which makes long term variability more prominent (Eltahir and Yeh 1999; Wu et al. 2002). An artificial division at the U.S.–Canadian border is caused by the lack of radar and gauge precipitation data in Canada, which hinders effective temporal downscaling and bias correction (Cosgrove et al. 2003). Trends in recharge do not correlate well with trends in annual temperature including the warming trends in most of the study domain (Fig. 10).
Long-term trends in annual temperature (°K yr−1), precipitation (mm yr−1), and recharge (mm yr−1) during 1979–2017 and temporal standard deviation (mm) of annual precipitation and recharge. Stipples indicate trends at the 0.05 significance level.
Citation: Journal of Hydrometeorology 22, 2; 10.1175/JHM-D-20-0130.1
The temporal standard deviation of annual recharge which exhibits spatial patterns similar to those of temporal variability of annual precipitation (Fig. 10). As has been observed in other statistics, the magnitude of the temporal standard deviation of recharge varies among the models and depends on the magnitude of simulated mean annual recharge, similar to the behavior in recharge trends.
4. Discussions
a. Model uncertainties and limitations
Certain spatial patterns in the uncertainty of simulated recharge and their causes have emerged from this study. In the eastern United States, where intermodel variability in ET is large (Xia et al. 2012a), simulated annual recharge also varied widely among the models. In the dry western United States, where uncertainties in simulated runoff and snow are high (Xia et al. 2012a,b), simulated annual recharge generally displayed low skills (as indicated by correlation and KGE), with respect to both the monthly and the annual USGS recharge estimates. In cold and wet climates where recharge is strongly modulated by spring snowmelt, Noah-MP, which simulates a deeper snowpack with more complex physics on snow and frozen soil, performed better than most other models in predicting annual recharge (Table 2). Evaluation using the monthly USGS recharge data shows soil moisture storage plays an important role in the timing of seasonal maximum recharge in cold regions. Although there is inconsistency between the two USGS recharge datasets, these results are consistent with previous studies on related processes (e.g., Xia et al. 2012b, 2017).
Evaluation using in situ soil moisture revealed systematic biases in the timing of deep layer seasonal maximum soil moisture with simulated soil moisture leading the SCAN observations This result suggests deficiencies in model physics. One issue is that these LSMs employ vertically homogeneous soils, while soil texture in the real-world changes with depth, as observed in SCAN soil texture profiles (available at the SCAN website). Studies have shown that adding vertical heterogeneity can lower effective hydraulic conductivity and infiltration rate (Gohardoust et al. 2017; Mohanty and Zhu 2007). The fact that deep soil moisture at the cluster of SCAN sites in the lower Mississippi River basin peaks during different months of the year depending on the location (Fig. 9) also points to the impact of subsurface heterogeneity on seasonal variability of point-scale soil moisture. Note that this discussion on the impact of heterogeneity is more or less specific to agricultural areas with unconsolidated sediments, which is where SCAN sites tend to be located. In regions covered with consolidated rocks, subsurface heterogeneity such as fractures and karsts can increase downward water fluxes (Hartmann et al. 2017).
Another issue is the subsurface runoff (base flow, Fig. 1a) algorithm used by the models. Noah applies a free drainage condition (Table 1) at the bottom of the soil column, which permits unlimited drainage. While this condition may be appropriate in dry climates with thick unsaturated zones, it can cause unrealistically fast drainage in wet climates where, in reality, the water tables might rise into the shallow soil during and after heavy precipitation events, limiting infiltration into the subsurface (Abboud et al. 2018). Other NLDAS-2 models employ empirical relationships to simulate subsurface runoff with parameters often obtained through calibration or tuning with respect to mean annual streamflow or ET. The latter fluxes are strongly influenced by annual precipitation, thus simulated soil moisture is sensitive to changes in precipitation. Moreover, uncertainties such as biases in the reference datasets (e.g., ET) and forcing fields (especially precipitation) used for calibration can have significant impacts on estimated recharge, which is sensitive to precipitation, land cover and soil textures (Moeck et al. 2020). One major cause for the large discrepancy in annual recharge between monthly and annual USGS datasets is the difference in the input ET data.
For models that simulate groundwater, water exchanges between the reservoir and the unsaturated zone significantly affect deep soil moisture. Noah-MP uses a bucket-type linear reservoir to simulate groundwater storage (not head) which cannot properly represent capillary rise, among other groundwater processes (Niu et al. 2007). As the capillary rise is inversely related to water table depth (Yang et al. 2011), underrepresentation of capillary rise would lead to overestimation of downward moisture fluxes, including recharge, in areas with shallow groundwater such as wetlands and floodplains. Capillary rise is often enhanced in the warm season (Yang et al. 2011), which is one reason why unconfined groundwater is observed to have strong seasonal variations (Li et al. 2019b). Because of its underrepresentation of capillary rise, Noah-MP simulates groundwater with weaker seasonality and smaller amplitude variations than observed in wells (Xia et al. 2017). On the other hand, CLSM simulates strong interactions among its water storage compartments, allowing upward movement of water to support ET in the warm season. As a result, its simulated groundwater storage has seasonal changes comparable to observations (Xia et al. 2017; Li et al. 2019a). However, CLSM only has two soil layers, at 2 and 100 cm, which is inadequate to simulate the lagged response of deep soil moisture to a precipitation event.
Seasonal groundwater storage maxima simulated by Noah-MP and CLSM do not always lead observed maxima, even at wells where the water table is deep. This highlights the fact that not all environmental controls on groundwater are represented in the models. For example, recharge can occur through percolation of surface water, vertically through preferential flow paths, or by lateral flows of groundwater (Winter et al. 1998; Wilson and Guan 2004; Markovich et al. 2019), while none of those processes are simulated by the LSMs. The water levels of surface water bodies that are in communication with aquifers also influence groundwater variations (Winter et al. 1998; Michael et al. 2005), but surface waters are not represented in the LSMs used in this study.
b. Future directions on model improvements
Vertical subsurface heterogeneity is not currently represented in the studied LSMs. To enable representation, model physics and parameters would have to be redeveloped for VIC, Mosaic, SAC, and CLSM, which rely on empirical relationships to simulate soil moisture. For Noah and Noah-MP, the soil-moisture-based Richards’ equation is only valid for homogeneous soils because soil moisture is unable to represent energy potential across different media (Jury et al. 1991). Note that the TOPMODEL physics used by some of the LSMs (Table 1) involve a depth-dependent hydraulic conductivity function for calculating subsurface runoff, but this relationship is not part of the soil moisture calculation.
Perhaps the best solution would be to implement a dynamic flow model based on the hydraulic head to simulate subsurface water flows (Celia et al. 1990; Sudicky et al. 2008; Maxwell et al. 2017). Hydraulic head can be used to represent the energy potential in both saturated and unsaturated zones and across different media. Therefore, a head-based flow model can simulate the moisture profile continuously from the surface to the water table and would make it possible to estimate recharge at the water table instead of somewhere above it (Fig. 1b). In fact, a dynamic flow model can simulate scenarios in which the water table rises into the unsaturated zone (Maxwell and Miller 2005), eliminating the artificial boundary between the unsaturated and saturated zones imposed in Noah-MP, for example. In addition, capillary rise can be appropriately simulated at the water table (as opposed to across the intermediate zone, Fig. 1b), using consistent physics governing downward moisture fluxes. Similarly, a dynamic flow model can represent subsurface heterogeneity and simulate recharge through preferential flow paths (Sweetenham et al. 2017). Surface water can be represented in a head-based flow model as a boundary condition and the associated recharge can be simulated based on Darcy’s law instead of an empirically derived simple relationships that do not consider subsurface heterogeneity (Döll et al. 2014; Sutanudjaja et al. 2018).
Simulating groundwater dynamics is challenged by the lack of spatially continuous information on subsurface properties such as hydraulic conductivity and three-dimensional aquifer maps. Coarsely classified global lithology maps and terrain features have been used to derive the aquifer properties and boundary conditions needed to simulate multidimensional groundwater flows (de Graaf et al. 2015; Maxwell et al. 2015; de Graaf et al. 2020). These derived hydrogeological parameters and conditions contribute to uncertainty in simulated head (Reinecke et al. 2019) and may be inadequate for simulating groundwater dynamics at continental scales (de Graaf et al. 2017). Still, a dynamic flow component would enable tuning or calibration of the models without the need to redevelop model physics. For example, a soil layer with low permeability could be configured near the land surface to delay the response of deep soil moisture to precipitation in a dry climate. Similarly, a water table too shallow to be represented in current LSMs could be configured in floodplains and elsewhere as appropriate. Modeling subsurface heterogeneity would improve recharge estimates in dry climates where infiltrated water often reaches aquifers quickly through macropores and in mountain front aquifers where fracture flows are important (Beven and Germann 2013; Wilson and Guan 2004). Other improvements include coupling with three-dimensional dynamic flow models (Kollet and Maxwell 2008; Hein et al. 2019; Tran et al. 2020) or implementing quasi-dynamic models (Miguez-Macho and Fan 2012) to simulate recharge from groundwater lateral flows.
In addition to subsurface flow simulation, the partitioning of precipitation into surface runoff and infiltration is vital for improving recharge estimates. Most evaluation and calibration have focused on total runoff which does not guarantee skills in recharge. In addition, the performance of calibrated models may depend on model structure and physics (Moeck et al. 2016, 2018). A topography-based runoff algorithm that does not represent precipitation partitioning in flat terrains may not perform the same as in mountainous regions. Although subject to uncertainties, the surface runoff estimates used in the USGS recharge datasets should provide useful information for diagnosing this issue.
Due to multitudes and complexity, parameters tuning is an effective approach to improve ET estimates, especially when the source of uncertainty is identified through evaluation or model intercomparison. As indicated earlier, warm season biases were identified in Noah simulated ET and other processes during NLDAS-1 and were significantly reduced through using spatially and seasonally varied leafy area indices, seasonally varied root density profiles and a different function for calculating canopy resistance (Xia et al. 2012a). One of the challenges for such modeling tuning is that it may lead to adverse impacts in areas where no ET observations are available for validating the impacts on simulated processes. Satellite derived ET products provide spatial coverage and the fusion with other ET products may help alleviate some of their limitations such as lack of water balance constraints (Long et al. 2014; Reitz et al. 2017b).
It is clear, from Eq. (1), that improving recharge also needs constraints on groundwater storage changes. The Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (FO) missions detect groundwater storage changes through continuous mapping of Earth’s gravity field (Tapley et al. 2004). By design, groundwater signals are embedded with other components of terrestrial water storage (TWS), soil moisture, snow, and surface water. In the last decades, data assimilation techniques have been used to disaggregate GRACE derived TWS into its component by relying on hydrological models that simulate individual components of TWS (Zaitchik et al. 2008; Su et al. 2010; Houborg et al. 2012; Kumar et al. 2016). A recent global-scale study has shown that GRACE data assimilation significantly reduced the estimation error in CLSM simulated groundwater storage and improved its correlation with in situ groundwater at both regional and point scales (Li et al. 2019b). These positive impacts on groundwater storage estimates do not necessarily suggest positive impacts on recharge which may exhibit different errors relative to that of simulated groundwater storage. For instance, if a model overestimates both groundwater storage changes and base flow, as GRACE data assimilation decreases groundwater storage changes, it would also decrease base flow due to the dependency of base flow on groundwater storage; on the other hand, if model underestimates base flow while overestimating groundwater storage changes, reducing groundwater storage changes by GRACE data assimilation would lead to even smaller base flow. The root cause for this issue is the lack of constraints on simulated fluxes (ET and runoff) during the data assimilation process (Li et al. 2012). For this reason, the impact of GRACE/GRACE-FO data assimilation on recharge may depend on model physics and should be assessed individually for each data assimilation experiment.
In addition to data assimilation, groundwater storage signals can be extracted from GRACE/GRACE-FO TWS by using ancillary data on soil moisture, snow, and surface water (Rodell et al. 2009). These GRACE-derived groundwater storage estimates can be used to constrain the calibration of regional groundwater models which often results in nonunique sets of hydraulic conductivity values when using water level measurements alone (Sun et al. 2012). In particular, Lo et al. (2010) showed that calibration of a land surface model using both GRACE derived groundwater storage and base flow led to more improvements in simulated water table variations than using base flow alone. The latter approach should help improve recharge estimates as well based on Eq. (1). GRACE/GRACE-FO-derived TWS observations should also help with identifying uncertainties in simulated groundwater withdrawals and associated recharge in heavily irrigated regions (Döll et al. 2014).
c. Implications for climate studies
We showed that significant trends in simulated recharge do not correlate with significant trends in precipitation and temperature consistently across the United States (Fig. 10). This is in contrast with other studies that suggested rising temperatures in future climate would, in general, lead to increased ET and hence decreased recharge and groundwater storage (e.g., Crosbie et al. 2013; Condon et al. 2020). Those studies used predicted changes in precipitation and temperature alone without considering changes in other atmospheric fields such as increases in relative humidity and decreases in incoming shortwave radiation due to increased cloud cover, both of which would favor decreasing ET (Sheffield et al. 2012; Li et al. 2019a). Further, rising nighttime temperatures, which have a small impact on ET, are the main component of increasing global mean temperature (Karl et al. 1991). As such atmospheric changes are reflected in some of reanalysis-based forcing datasets (e.g., Sheffield et al. 2012; Li et al. 2019a), it is not surprising that trends in recharge are not easily predictable based on trends in precipitation and temperature. However, future studies are needed to evaluate long term changes in NLDAS-2 forcing fields against other datasets.
In interpreting the results presented here, it is important to note that the LSMs only simulate diffuse recharge, whereas focused recharge is likely to be important in many regions and likely responds differently to changes in climate. In addition, studies have shown that groundwater in highly permeable aquifers or in aquifers close to mountain fronts and rivers responds more strongly to large-scale climate signals than in other areas (Anderson and Emanuel 2008; Hanson et al. 2006). This suggests that representing subsurface heterogeneity may also be important for studying impacts of climate change on recharge and groundwater.
5. Conclusions
This study evaluated diffuse recharge from a suite of land surface models (LSMs) forced by a common set of atmospheric forcing fields in the conterminous United States during the period of 1979–2017. Considering the limited field recharge estimates (Scanlon et al. 2006; Mohan et al. 2018; Moeck et al. 2020), this suite of recharge estimates, if reasonably accurate, would be valuable for groundwater sustainability management and climate change studies.
Our analysis showed that the simulated recharge time series from the different models exhibited similar spatial patterns in mean and interannual variability of annual recharge, suggesting the LSMs may be capable of simulating impacts of changing in precipitation and ET on diffuse recharge. However, large discrepancies existed among simulated recharge magnitudes and the timing of the seasonal maximum, due to differences in simulated ET, surface runoff, snow and snowmelt. Evaluated against annual USGS recharge estimates, the LSMs were shown to have relatively low skill in simulating annual recharge in dry climates, where the models also have difficulty simulating runoff (Xia et al. 2012b). The ensemble mean recharge compares better with the USGS estimated annual recharge, which is another common conclusion in model intercomparison studies. However, the ensemble mean did not compare as well against the monthly USGS recharge dataset, which yields different annual mean recharge than the annual USGS dataset. We conclude that there is uncertainty in these reference recharge datasets and that further evaluation would be warranted when new recharge datasets become available.
Finally, it is worth emphasizing that the NLDAS-2 forcing fields are also imperfect, especially at high elevations where station data used for bias-correcting precipitation are scarce (Cosgrove et al. 2003). In particular, the hourly precipitation data are temporally disaggregated from daily gauge data using radar imagery that is often not available in mountainous regions. Further, their temporal variability may be affected by changes in the number and locations of gauges and changes in the radar products (Ferguson and Mocko 2017). Studies have shown that temporal inconsistency in precipitation data has serious consequences when attempting to simulate long-term trends in land surface states and fluxes (Li et al. 2019a). These uncertainties, along with deficiencies in the model parameterizations, should be considered when using simulated recharge for hydrological research and applications.
Acknowledgments
This work was supported by NASA Western Water Application Office. The NLDAS-2 forcing data used in this study 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). Computing resources were provided by the NASA High-End Computing (HEC) Program through the NASA Center for Climate Simulation (NCCS) at Goddard Space Flight Center.
Data availability statement
NLDAS-2 model output and forcing data are available at NASA GES DISC, https://disc.gsfc.nasa.gov/datasets?keywords=NLDAS, and WLDAS output are available at https://portal.nccs.nasa.gov/datashare/WLDAS/.
REFERENCES
Abboud, J. M., M. C. Ryan, and G. D. Osborn, 2018: Groundwater flooding in a river-connected alluvial aquifer. J. Flood Risk Manage., 11, e12334, https://doi.org/10.1111/jfr3.12334.
Allen, D. M., D. C. Mackie, and M. Wei, 2004: Groundwater and climate change: A sensitivity analysis for the Grand Forks aquifer, southern British Columbia, Canada. Hydrogeol. J., 12, 270–290, https://doi.org/10.1007/s10040-003-0261-9.
Alley, W. M., T. E. Reilly, and O. L. Franke, 1999: Sustainability of ground-water resources. USGS Circular 1186, 79 pp., https://pubs.usgs.gov/circ/circ1186/pdf/circ1186.pdf.
Alley, W. M., R. W. Healy, J. W. LaBaugh, and T. E. Reilly, 2002: Flow and storage in groundwater systems. Science, 296, 1985–1990, https://doi.org/10.1126/science.1067123.
Anderson, W. P., Jr., and R. E. Emanuel, 2008: Effect of interannual and interdecadal climate oscillations on groundwater in North Carolina. Geophys. Res. Lett., 35, L23402, https://doi.org/10.1029/2008GL036054.
Arsenault, K. R., and Coauthors, 2018: The Land surface Data Toolkit (LDT v7.2) – A data fusion environment for land data assimilation systems. Geosci. Model Dev., 11, 3605–3621, https://doi.org/10.5194/gmd-11-3605-2018.
Bair, E. H., K. Rittger, R. E. Davis, T. H. Painter, and J. Dozier, 2016: Validating reconstruction of snow water equivalent in California’s Sierra Nevada using measurements from the NASA Airborne Snow Observatory. Water Resour. Res., 52, 8437–8460, https://doi.org/10.1002/2016WR018704.
Befus, K. M., K. D. Kroeger, C. G. Smith, and P. W. Swarzenski, 2017: The magnitude and origin of groundwater discharge to eastern U.S. and Gulf of Mexico coastal waters. Geophys. Res. Lett., 44, 10 396–10 406, https://doi.org/10.1002/2017GL075238.
Beven, K., and P. Germann, 2013: Macropores and water flow in soils revisited. Water Resour. Res., 49, 3071–3092, https://doi.org/10.1002/wrcr.20156.
Bloomfield, J. P., B. P. Marchant, S. H. Bricker, and R. B. Morgan, 2015: Regional analysis of groundwater droughts using hydro-graph classification. Hydrol. Earth Syst. Sci., 19, 4327–4344, https://doi.org/10.5194/hess-19-4327-2015.
Cao, G., B. R. Scanlon, D. Han, and C. Zheng, 2016: Impacts of thickening unsaturated zone on groundwater recharge in the North China Plain. J. Hydrol., 537, 260–270, https://doi.org/10.1016/j.jhydrol.2016.03.049.
Celia, M., E. T. Bouloutas, and R. L. Zarba, 1990: A General mass-conservative numerical solution for the unsaturated flow equation. Water Resour. Res., 26, 1483–1496, https://doi.org/10.1029/WR026i007p01483.
Chen, F., and Coauthors, 1996: Modeling of land-surface evaporation by four schemes and comparison with FIFE observations. J. Geophys. Res., 101, 7251–7268, https://doi.org/10.1029/95JD02165.
Condon, L. E., A. L. Atchley, and R. M. Maxwell, 2020: Evapotranspiration depletes groundwater under warming over the contiguous United States. Nat. Commun., 11, 873, https://doi.org/10.1038/s41467-020-14688-0.
Cosby, B., G. M. Hornberger, R. B. Clapp, and T. R. Ginn, 1984: A statistical exploration of the relationships of soil moisture characteristics to the physical properties of soils. Water Res. Res., 20, 682–690, https://doi.org/10.1029/WR020i006p00682.
Cosgrove, B. A., and Coauthors, 2003: Real-time and retrospective forcing in the North American Land Data Assimilation System (NLDAS) project. J. Geophys. Res., 108, 8842, https://doi.org/10.1029/2002JD003118.
Crosbie, R. S., I. Jolly, F. Leaney, and C. Petheram, 2010: Can the dataset of field based recharge estimates in Australia be used to predict recharge in data-poor areas? Hydrol. Earth Syst. Sci., 14, 2023–2038, https://doi.org/10.5194/hess-14-2023-2010.
Crosbie, R. S., B. R. Scanlon, F. S. Mpelasoka, R. C. Reedy, J. B. Gates, and L. Zhang, 2013: Potential climate change effects on groundwater recharge in the High Plains Aquifer, USA. Water Resour. Res., 49, 3936–3951, https://doi.org/10.1002/wrcr.20292.
Daly, C., M. Halbleib, J. I. Smith, W. P. Gibson, M. K. Doggett, G. H. Taylor, J. Curtis, and P. A. Pasteris, 2008: Physiographically sensitive mapping of climatological temperature and precipitation across the conterminous United States. Int. J. Climatol., 28, 2031–2064, https://doi.org/10.1002/joc.1688.
de Graaf, I. D., E. H. Sutanudjaja, L. P. H. Van Beek, and M. F. P. Bierkens, 2015: A high-resolution global-scale groundwater model. Hydrol. Earth Syst. Sci., 19, 823–837, https://doi.org/10.5194/hess-19-823-2015.
de Graaf, I. D., R. L. van Beek, T. Gleeson, N. Moosdorf, O. Schmitz, E. H. Sutanudjaja, and M. F. Bierkens, 2017: A global-scale two-layer transient groundwater model: Development and application to groundwater depletion. Adv. Water Resour., 102, 53–67, https://doi.org/10.1016/j.advwatres.2017.01.011.
de Graaf, I. D., L. Condon, and R. Maxwell, 2020: Hyper-resolution continental-scale 3-D aquifer parameterization for groundwater modeling. Water Resour. Res., 56, e2019WR026004, https://doi.org/10.1029/2019WR026004.
Döll, P., H. M. Schmied, C. Schuh, F. T. Portmann, and A. Eicker, 2014: Global-scale assessment of groundwater depletion and related groundwater abstractions: Combining hydrological modeling with information from well observations and GRACE satellites. Water Resour. Res., 50, 5698–5720, https://doi.org/10.1002/2014WR015595.
Edmunds, W. M., and C. B. Gaye, 1994: Estimating the spatial variability of groundwater recharge in the Sahel using chloride. J. Hydrol., 156, 47–59, https://doi.org/10.1016/0022-1694(94)90070-1.
Eltahir, E. A. B., and P. J.-F. Yeh, 1999: On the asymmetric response of aquifer water level to floods and droughts in Illinois. Water Resour. Res., 35, 1199–1217, https://doi.org/10.1029/1998WR900071.
Erlingis, J. M., and Coauthors, 2021: A high-resolution land data assimilation system optimized for the western United States. J. Amer. Water Resour. Assoc., in press.
Famiglietti, J. S., and Coauthors, 2011: Satellites measure recent rates of groundwater depletion in California’s Central Valley. Geophys. Res. Lett., 38, L03403, https://doi.org/10.1029/2010GL046442.
Feng, W., M. Zhong, J.-M. Lemoine, R. Biancale, H.-T. Hsu, and J. Xia, 2013: Evaluation of groundwater depletion in North China using the Gravity Recovery and Climate Experiment (GRACE) data and ground-based measurements. Water Resour. Res., 49, 2110–2118, https://doi.org/10.1002/wrcr.20192.
Ferguson, C., and D. Mocko, 2017: Diagnosing an artificial trend in NLDAS-2 afternoon precipitation. J. Hydrometeor., 18, 1051–1070, https://doi.org/10.1175/JHM-D-16-0251.1.
Flint, A. L., L. E. Flint, E. M. Kwicklis, J. M. Fabryka-Martin, and G. S. Bodvarsson, 2002: Estimating recharge at Yucca Mountain, Nevada, USA: Comparison of methods. Hydrogeol. J., 10, 180–204, https://doi.org/10.1007/s10040-001-0169-1.
Gleeson, T., Y. Wada, M. F. Bierkens, and L. P. Van Beek, 2012: Water balance of global aquifers revealed by groundwater footprint. Nature, 488, 197–200, https://doi.org/10.1038/nature11295.
Gohardoust, M. R., M. Sadeghi, M. Z. Ahmadi, S. B. Jones, and M. Tuller, 2017: Hydraulic conductivity of stratified unsaturated soils: Effects of random variability and layering. J. Hydrol., 546, 81–89, https://doi.org/10.1016/j.jhydrol.2016.12.055.
Green, T. R., M. Taniguchi, H. Kooi, J. J. Gurdak, D. M. Allen, K. M. Hiscock, H. Treidel, and A. Aureli, 2011: Beneath the surface of global change, Impacts of climate change on groundwater. J. Hydrol., 405, 532–560, https://doi.org/10.1016/j.jhydrol.2011.05.002.
Gupta, H. V., H. Kling, K. K. Yilmaz, and G. F. Martinez, 2009: Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. J. Hydrol., 377, 80–91, https://doi.org/10.1016/j.jhydrol.2009.08.003.