Comparing potential recharge estimates from three Land Surface Models across the Western US

Abstract

Groundwater is a major source of water in the western US. However, there are limited recharge estimates available in this region due to the complexity of recharge processes and the challenge of direct observations. Land surface Models (LSMs) could be a valuable tool for estimating current recharge and projecting changes due to future climate change. In this study, simulations of three LSMs (Noah, Mosaic and VIC) obtained from the North American Land Data Assimilation System (NLDAS-2) are used to estimate potential recharge in the western US. Modeled recharge was compared with published recharge estimates for several aquifers in the region. Annual recharge to precipitation ratios across the study basins varied from 0.01–15% for Mosaic, 3.2–42% for Noah, and 6.7–31.8% for VIC simulations. Mosaic consistently underestimates recharge across all basins. Noah captures recharge reasonably well in wetter basins, but overestimates it in drier basins. VIC slightly overestimates recharge in drier basins and slightly underestimates it for wetter basins. While the average annual recharge values vary among the models, the models were consistent in identifying high and low recharge areas in the region. Models agree in seasonality of recharge occurring dominantly during the spring across the region. Overall, our results highlight that LSMs have the potential to capture the spatial and temporal patterns as well as seasonality of recharge at large scales. Therefore, LSMs (specifically VIC and Noah) can be used as a tool for estimating future recharge rates in data limited regions.

1. Introduction

Groundwater is a life-sustaining natural resource that supplies water to billions of people on earth (Gleeson et al. 2012). It plays a central part in irrigated agriculture and sustaining ecosystems (Giordano 2009; Siebert et al 2010), and enables human adaptation to climate variability and change (Taylor et al. 2012). Globally it accounts for 1/3rd of all fresh-water withdrawals, for domestic (36%), agricultural (42%) and industrial purposes (27%) (Doll et al. 2012; Taylor et al. 2012). In the United States (US), ground water is the source of drinking water for 50% of the population and as much as 90% of the population in rural areas, especially in the West (Anderson and Woosley 2005). Reduced reliability of surface water supplies in the western US with projected increases in evaporative demand and uncertain changes in annual precipitation (Rasmussen et al. 2011, 2014) may increase groundwater use (Scanlon 2005). Many areas of the region are already experiencing groundwater depletion caused by sustained groundwater pumping (Faunt 2009; Konikow 2013; Castle et al. 2014).

Groundwater recharge is a flux of water into the saturated zone. Spatial variability of recharge rates is controlled by precipitation and other climate variables (Hoekstra and Meknnen 2012; Hoekstra et al. 2012; Gleeson et al. 2012), vegetation, soils, and geology (Stonestrom et al. 2007). Despite the importance of groundwater in this region, limited recharge estimates are available due to the complexity of recharge processes and the lack of feasible measurement methods (Scanlon et al. 2006). Thus, improving current recharge estimates and understanding spatial variability of recharge processes are essential for sustainably managing this precious resource (Scanlon et al. 2006; Famiglietti and Rodell 2013) to meet human and ecosystem demands in the future (Scanlon et al. 2006).

Recharge estimation methods include water balance accounting, remote sensing, observational methods and environmental tracer analysis, and modeling (Scanlon et al 2006; Healy, 2010). In the western US, groundwater recharge generally occurs at depth, where direct observational methods cannot be applied. Several Land Surface Models (LSMs) (e.g SAC-SMA (Burnash et al. 1973; Burnash 1995; SSiB, Xue et al. 1991); Mosaic (Koster and Suarez 1996); NSSIP (Koster et al. 2000); Variable Infiltration Capacity (VIC; Liang et al. 1994); Noah (Ek et al. 2003, Nui et al. 2011); and CLM (Bonan et al. 2002, 2011)) have been developed over the last few decades to better represent land surface and atmospheric processes as well as improve estimates of various water, energy and carbon fluxes at the land surface. These models could be a valuable tool for estimating current and future recharge estimates due to projected climate change. However, to date, besides currently published recharge estimates (Li et al. 2015) in the eastern US using the VIC model, recharge estimates from these models have not been comprehensively assessed.

LSMs vary in representation of the exchange of energy, mass, momentum and CO₂ exchange between the land surface and the overlying atmosphere (Koster and Suarez 1996; Liang et al. 1994; Bonan et al. 2011; Nui et al. 2011). It is therefore important to understand how differences in model structure affect the simulation of recharge and whether certain LSMs perform better under particular physiographic and climatic settings. In this paper, we compare recharge estimates from three LSMs over the western US with a specific emphasis on 10 aquifer systems where recharge estimates from other approaches are available.

The major questions addressed in this study are: 1. Are recharge estimates in the western US from various LSMs significantly different? 2. Do LSMs provide reasonable estimates of recharge in the western US? 3. Do the amount, seasonality, trend and spatial pattern of recharge vary based on the choice of LSMs? For addressing these questions, simulations of three LSMs (Mosaic, Noah, and VIC) obtained from the North American Land Data Assimilation System-phase 2 (NLDAS-2) were used for assessing recharge estimates across the western US. We used MODIS ET (Mu et al. 2011) and baseflow index (BFI) based recharge (Wolock 2003a) for the whole western US for comparison and evaluation purpose. Simulated recharge from the LSMs was compared with published recharge estimates from 10 aquifers in the region (Northern Plains, Central High Plains, Southern High Plains, San Pedro, Death Valley, Salt Lake Valley, Central Valley, Columbia Plateau, Spokane Valley, and Williston, Fig 1) synthesized by Meixner et al. (2016). These aquifers represent a broad sample of variability in climatological, geological, and hydrological characteristics along with anthropogenic pressures like groundwater pumping on the aquifers. Trends, amounts, and patterns of recharge from the three models were compared statistically to determine their consistency. Statistical analyses (Kolmogorov-Smirnov test, Kendall Tau trend analysis test, spatial pattern correlation test) were conducted using R (version R 3.1.3).

Fig. 1.

Study Region shown with elevation and study basins

2. Methods

2.1. North American Land Data Assimilation System Phase 2 (NLDAS-2)

NLDAS-2 (Mitchell et al. 2004; Xia et al. 2012) integrates observation-based and model reanalysis data to drive LSMs offline. It executes 4? LSMs at 1/8th-degree grid spacing at an hourly temporal scale over central North America, enabled by the Land Information System (LIS; Kumar et al. 2006; Peters-Lidard et al. 2007). LIS is a scalable land data assimilation system that integrates a suite of advanced LSMs, high resolution satellite and observational data, data assimilation and parameter optimization techniques, and high-performance computing tools (Kumar et al. 2006; Peters-Lidard et al. 2007). Outputs from three LSMs (Mosaic, Noah, and VIC) over a 30-year historical period (1981–2010) were used to answer the study questions posed above. The first two years of the simulation (1979–1980) were used as a model spin up period and excluded from the analysis. The data used in this study were generated within NASA’s Earth Science Division and are archived and distributed by the Goddard Earth Sciences (GES) Data and Information Services Center (DISC; http://disc.gsfc.nasa.gov/hydrology/index.shtml).

Past NLDAS-2 multi-model evaluations are performed for evaluating surface water and energy fluxes (Wei et al 2013), such as soil moisture (Mo et al. 2011; Mo et al. 2012; Xia et al. 2014; Xia et al. 2015a; Xia et al. 2015b), evapotranspiration (Long et al. 2014), soil temperature (Xia et al. 2013), and streamflow (Mo et al. 2012; Xia et al. 2012b). Although there have been efforts to compare the continental scale water and energy-fluxes among LSMs for NLDAS-2, (Xia et al. 2012a) model performance was only evaluated with surface water (Xia et al. 2012b). No particular efforts have been made using these models to simulate and characterize the amount and seasonality of recharge in the western US, mostly due to the limited availability of recharge data. However, recently Li et al. (2015) demonstrated that NLDAS/VIC provides reasonable estimates of groundwater recharge in the central and northeastern US.

2.2 Descriptions of LSMs

2.2.1 Formulation of Water and Energy Budget in LSMs

Although the 3 LSMs vary in complexity for the treatment of exchange of energy, mass, momentum and CO₂ between land surface and overlying atmosphere, they follow similar fundamental conceptualization of the energy and water budget. The water balance is calculated based on the continuity equation:

$$ds/dt=P-ET-R-G$$

where, ds/dt is the change in strorage (mm), P is precipitation (mm), R is surface runoff (mm), G is ground water runoff (mm) and ET is evapotranspiration (mm) which is calculated as,

$$ET=CE+BE+T+S$$

where, CE is canopy evaporation (mm), BE is bare soil evaporation (mm), T is transpiration (mm) and S is sublimation (mm).

The models assume gravity-driven, free-drainage from the bottom layer as subsurface runoff/recharge, and surface runoff is the excess water after infiltration.

The surface energy balance is calculated based on the equation:

$$R_n=LE+SH+G+SF$$

where, R_n is the net radiation flux (W/m²), LE (λET) is the latent heat flux (W/m²), SH is the sensible heat flux (W/m²), G is the ground heat flux (W/m²), and SF is the snow phase-change heat flux (W/m²).

2.2.2. Mosaic

The 1D Mosaic LSM (Koster and Suarez 1992; Koster and Suarez 1996, Table 1) which calculates both energy and water balance, accounts for subgrid heterogeneity of land surface characteristics by dividing each grid cell into several subregions, called “tiles,”. Each tile contains a single vegetation or bare soil type (Koster and Suaraz, 1996). Energy and water balance calculations are performed separately over each tile and weighted by fractional coverage to calculate the total fluxes for each grid cell. The vertical structure of the model includes a single canopy layer and three soil layers: a thin surface layer (0–10cm), a middle layer (10–40cm) that encompasses the remainder of the root zone, and a lower “recharge” layer (40–200cm) at the bottom. Mosaic calculates total evapotranspiration as the sum of bare soil evaporation, transpiration, and canopy evaporation. Runoff occur both as overland flow during precipitation events and as delayed baseflow. Mosaic treats baseflow/recharge as a linear function of water content, bedrock slope, and hydraulic conductivity of the bottom layer.

Table 1.

Physical characteristics of study basins

Aquifers	Area (sq. km)	P (mm/yr)	Recharge (mm/yr)	Aquifer material
High Plains(HP)	451,000	535	48	unconsolidated, poorly sorted clay, silt, sand and gravel underlain by bedrock
NHP	250,000	548	73.7	same as HP
CHP	125,000	545	33.7	same as HP
SHP	76,000	472	27.9	same as HP
Central Valley	52,000	650	315.4	sand and gravel
Death Valley	45,300	185	2.8	carbonate and volcanic rock, and alluvium
Colombia	114,000	442	116.7	basalt, sand and gravel
San Pedro	7,560	371	6.5	sand and gravel
Spokane	2,100	689	300.0	sand and gravel
Williston	102,400	382	4.7	sand and gravel

2.2.3. Noah

Noah (Ek et al. 2003, Table 1) is a 1-D column model that simulates soil moisture, soil temperature, snow depth, snow water equivalent, canopy water content, and water and energy flux terms of the surface water and energy balance (Mitchell 2004). For this study, Noah was configured to have a 2-m-deep soil layer divided into the 4 sub-layers:0–10 cm, 10–40 cm, 40–100 cm, and 100 cm–200 cm. The deepest layer acts as a reservoir with gravity drainage at the bottom. The volumetric soil moisture is determined using the diffusive form of Richard’s equation. The total evaporation, in the absence of snow, is the sum of direct evaporation from the topmost soil layer, evaporation of precipitation intercepted by plant canopy, and transpiration from the vegetation canopy. The Noah LSM assumes spatially continuous soil moisture values within tiles pixels, parameterizes surface runoff with a simple infiltration-excess scheme, and treats base- flow/recharge as a linear function of hydraulic conductivity (K) of the bottom soil layer (Schaake et al. 1996).

2.2.4. VIC

The VIC (Liang et al. 1994, Table 1) model incorporated within NLDAS-2, and characterizes the subsurface as three soil layers with spatially variable thickness. The surface is described by different vegetation types plus bare soil. The land cover types are specified by their leaf area index (LAI), canopy resistance and relative fraction of roots in each of the soil layers. Evapotranspiration from each vegetation type is characterized by potential evapotranspiration together with canopy resistance and aerodynamic resistance to water transfers. Associated with each land cover type are a single canopy layer, and multiple soil layers (up to 2 m depth). Subsurface hydrology parameterizations of the VIC LSM is more complex (Liang et al. 1994) because it uses a spatial probability distribution to represent subgrid heterogeneity in soil moisture. It also treats baseflow/recharge as a nonlinear recession curve which is a function of bottom layer soil moisture (it is linear below a threshold and then non-linear above that threshold). The top 2 soil layers are designed to represent the dynamic behavior of the soil column that responds to rainfall events and evapotranspiration, and the lower layers control inter-storm soil moisture behavior. The lower layer only responds to rainfall when the upper layer is fully saturated and thus can separate subsurface flow from quick response storm flow. Roots can extend down to the bottom layer, depending on the vegetation and soil type. The soil characteristics (such as soil texture, hydraulic conductivity, etc.) are held constant for each grid cell. In the model, all the states and output variables are calculated for each land cover tile at each time step and weighted by fractional land cover to calculate the total fluxes for each grid cell.

2.3 Descriptions of study area/basins

2.3.1. Western US

The Western US (Fig 1) is the largest region of the country, covering more than half of the land area of the contiguous US. It is also the most geographically diverse region in the country encompassing the Pacific Coast, the temperate rainforests of the Northwest, the highest mountain ranges (including the Rocky Mountains, the Sierra Nevada, and Cascade Range), the Great Plains, and all of the desert areas located in the US (the Mojave, Sonoran, Great Basin, and Chihuahua deserts). Elevation varies between −86 m to 4402 m above sea level (Fig 1).

The Western US consists primarily of five land-use/land-cover classes: grassland/shrubland (59%), forest (28.1%), agriculture (6.3%), developed (1.5%), and barren (1.9%) (Sleeter at al. 2012). Grassland/shrubland and barren lands are most common in the arid-southwest and interior desert regions, whereas forest dominates in the Pacific Northwest and Rocky Mountains. Agriculture and developed areas are found to some degree in nearly all regions but are concentrated mainly in a relatively few high-density areas (USGS 2012).

As a generalization, the climate of the Western US can be described as overall semiarid; however, parts of the region get extremely high amounts of rain and/or snow, and other parts are true desert and get little rain per year. Annual rainfall (Fig 2) ranges between 58 mm to 5051 mm based on NLDAS 2 data and is greater in the eastern portions, gradually decreasing until reaching the Pacific Coast where it again increases.

Fig. 2.

Average annual precipitation (1981–2010) across the western US

2.3.2. Study Basins

High Plains Aquifer

The High Plains aquifer (HPA, Fig 1) extends into eight States: Colorado, Kansas, Nebraska, New Mexico, Oklahoma, South Dakota, Texas and Wyoming. The aquifer is comprised of unconsolidated, poorly sorted clay, silt, sand and gravel and is underlain by bedrock units. HPA is divided into Northern (NHP), Central (CHP) and Southern (SHP) regions. Average precipitation (P) in HPA is 522 mm/yr and average recharge is 48 mm/yr (Meixner et al 2015). In general, average annual P and recharge increase from south to north (Table 2) and occurs predominantly during summer.

Table 2.

Basic differences in LSMs used in this study

	Mosaic (1D)	Noah (1D)	VIC (1D)
Run time step	15 min	15 min	1 hour
Soil Layer	3	4	3
Soli layer depths	10, 30, 160 cm (constant)	10, 30, 60, 100 cm (constant)	10 cm, variable, variable (variable)
Tiling:Vegetation	Y	N	Y
Tiling: Elevation	N	N	Y
Snow Layers	1	1	2
Soil temperature profile	N	Y	Y
Drainage	Y (linear)	Y(linear)	Y(non-linear)
Soil water: vertical diffusion	Y	Y	N
Rooting depth	40 cm (constant)	100 cm (constant) expect forest (down to 200 cm)	Variable (down to 200 cm)
Rooting density	constant	constant	exponential
Canopy capacity	0–1.6 mm	0.5 mm	0.1–1 mm
Convective P input	Y	N	N
PET	Input	Calculates itself	Input
Diurnal Albedo	Y	N	N

^*Y:Yes; N:No

San Pedro Aquifer

The San Pedro Basin (Fig 1) in southern Arizona is representative of the hydrogeology of a southern Basin and Range aquifer system (Goode and Maddock 2000). It is an alluvial aquifer that is comprised of basin-bounding crystalline and sedimentary rock mountains and unconsolidated sediments of clay, silt, sand, and gravel within the valley (Pool and Dickinson 2006). The basin receives an annual average precipitation and recharge of 400 mm and 6.5 mm respectively (Table 2). The majority of annual rainfall (~54%) in the San Pedro occurs during the summer monsoon season, with the remainder occurring in the winter months as rain and snow from low-intensity storms.

Death Valley Aquifer

The Death Valley Aquifer System (Fig 1) is located in the arid southern Great Basin of Nevada and California. Major aquifers consist of fractured volcanic rock and alluvium. The average annual P and recharge for the aquifer is 185 mm and 2.8 mm respectively (Table 2). Precipitation particularly as snowfallin mountain systems is predominantly in the winter months.

Salt Lake Valley (SLV) Aquifer

SLV aquifer (Fig 1), a representative of northern Basin and Range aquifer system consists of shallow unconfined aquifers underlain by confined to semi-confined sand and gravel aquifers, (Lambert 1995; Cederberg et al. 2009). The average annual P and recharge for the aquifer are 488 mm and 203 mm respectively (Table 2) with most of the P falling as snow during winter and spring.

Central Valley Aquifer

The Central Valley aquifer system of California (Fig 1) is an unconsolidated sand and gravel aquifer that underlies the Sacramento and San Joaquin Valleys of central California. The average annual P and recharge for the aquifer are 650 mm and 315 mm respectively (Table 2). About 85% of the precipitation falls from November to April in Central Valley.

Columbia Plateau Aquifer

The Columbia Plateau aquifer system (Fig 1) in Washington, Oregon, and Idaho (Kahle et al. 2011) consists of productive basalt aquifers characterized by highly permeable interflow zones separated by less permeable basalt-flow interiors. Extensive sedimentary aquifers consisting of valley-fill deposits lie atop the basalts along major drainages. With an average annual P of 440 mm mostly occurring during winter months, recharge is estimated to be 162 mm (Table 2).

Spokane Valley-Rathdrum Prairie Glacial Aquifer

The Spokane Valley-Rathdrum Prairie aquifer (Fig 1) is a glacial aquifer in northwestern Idaho and northeastern Washington (Hutson et al. 2004). The aquifer is composed of coarse-grained sediments with fine-grained layers interspersed (Hsieh et al. 2007; Kahle and Bartolino 2007). Annual average P is 689 mm/yr and mostly is concentrated during winter, and average recharge is 300 mm/yr (Table 2).

Williston Basin Glacial Aquifer System

The Williston Basin (Fig 1) is present within southern Canada, northeastern Montana, and western North Dakota (Soller et al. 2012). The aquifer is composed of till, clay, silt, sand, and gravels (Fullerton et al. 2004). The average P and recharge estimates are 382 mm/yr and 4.7 mm/yr respectively (Table 2). Both P and recharge are summer dominated.

2.4. Evaluation Datasets

2.4.1 MODIS Evapotranspiration data

The MOD16 (Mu et al 2011) global evapotranspiration (ET) datasets (Fig 3) are regular 1-km² land surface ET datasets for the global vegetated land areas at 8-day, monthly and annual intervals. The dataset covers the time period 2000 – 2010. The ET algorithm is based on the Penman-Monteith equation (Monteith 1965). Terrestrial ET includes evaporation from wet and moist soil, from rain water intercepted by the canopy before it reaches the ground, and the transpiration through stomata on plant leaves and stems. Evaporation of water intercepted by the canopy is a very important water flux for ecosystems with a high LAI. It should be noted that the MODIS ET estimates used for evaluation have lots of uncertainties associated with input data (e.g. LAI, PAR), inaccuracy of measured ET data from eddy covariance flux towers, scaling from flux towers to landscape, and algorithms (associated with processes and parameters). However, it is the best available dataset for the region due to its spatial and temporal coverage.

Fig. 3.

Annual average ET estimates (2000–2010) from LSMs, MODIS, model ensemble mean and model spread

2.4.2 Baseflow Index (BFI) recharge

A spatially distributed recharge map (Wolock 2003a, Fig 5) was created by multiplying a grid of base-flow index (BFI) values (Wolock 2003b) by a grid of streamflow values (Gebert et al. 1987) derived from a 1951–1980 mean annual runoff contour map generated for the whole USA. The assumptions inherent in this recharge estimation approach are that in long term: (1) recharge is equal to discharge, and (2) the BFI reasonably represents the proportion of natural ground-water discharge to streamflow. The BFI grid (1 km resolution) was interpolated from the BFI values of 8,249 U.S. Geological Survey stream gages (Wolock 2003c) using the inverse distance weighting interpolation method. The BFI values are computed using an automated hydrograph separation computer program called the BFI program (Wahl and Wahl 1988; Wahl and Wahl 1995). However it should be noted that the BFI-based recharge itself is a very rough estimate and should not be treated as an observation due to high uncertainty related to this dataset. The recharge dataset likely reflects general patterns across broad geographic regions, but recharge values at specific sites are unlikely to be accurate.

Fig. 5.

Scatter plots of recharge between models and models vs BFI-based recharge

2.4.3 Basin wide literature estimates of recharge

The literature based recharge estimates for the study basins synthesized by Meixner et al. (2016) are used for evaluating model estimates. These recharge estimates come from various sources and different approaches (observational, environmental tracer analysis, and modeling) are used for making these estimates. Although the methodology is not consistent over the study basins, these estimates are the best available.

3. Results and Discussions

3.1. Comparing ET: Among models and with MODIS ET

The models tend to agree on the spatial pattern of ET (Table 3, Fig 3) with each other and with the MODIS ET following the pattern of P (Fig 2), though ET rates vary. MODIS ET was generally lower than LSMs ET. Mosaic consistently generated higher ET compared to Noah and VIC. Model ensemble mean slightly improved the spatial pattern of ET when compared with MODIS ET (Fig 3).

Table 3.

Pattern correlation (Pearson’s r) matrix on ET estimates among LSMs, model ensemble mean (ENS) and MODIS-ET

	Mosaic	VIC	Noah	ENS	MODIS
Mosaic	1.00	0.89*	0.86*	0.97*	0.87*
VIC	0.89*	1.00	0.74*	0.86*	0.77*
Noah	0.86*	0.74*	1.00	0.92*	0.75*
ENS	0.97*	0.86*	0.92*	1.00	0.93*
MODIS	0.87*	0.77*	0.75*	0.93*	1.00

^*Statistically significant (p<0.05)

ET more or less followed the pattern of P across the region (Fig 2, Fig 3). Annual average precipitation ranges between 58 mm to 5051 mm based on NLDAS 2 data (Fig 2). A gradual decrease in P from east towards west before a significant increase in P at the west coast was observed. Among the basins examined, Death Valley (P̄= 185 mm) and Spokane Valley (P̄ = 689 mm) are the driest and the wettest basins respectively (Table 1). Average annual ET was estimated between 58 mm and 1260 mm for Mosaic, between 36 mm and 1123 mm for Noah, between 21 mm and 986 mm for VIC and between 49 mm and 1026 mm based on model ensemble mean. The models and MODIS ET (Fig 3, Table 3) showed the lowest ET in southern regions and higher ET in western coast and lower eastern regions (Fig 3). Higher ET on the western coast is related to higher water availability from higher P. Higher ET in south eastern corner was due to the combined effect of T and P (high T and moderate P).

MODIS estimated ET (ranged between 35 mm and 1175 mm) was generally lower than LSMs ET over the western US (Fig 3). Annual ET is the highest for the Mosaic LSM model, and lowest for MODIS. LSM’s ET was higher than the MODIS ET for most of the study basins. Mean annual ET values across the study basins were between 176 mm and 597 mm (87% and 99% of P) based on Mosaic, 120 mm and 454 mm (59% and 91% of P) based on Noah, 153 mm and 485 mm (54% and 89% of P) based on VIC and 170 mm and 490 mm (71 and 92% of P) based on MODIS (Table 4). The Spokane valley has the lowest ET/P ratio according to all models (Table 5). The San Pedro basin has the highest ET/P ratio based on Mosaic and Noah, and Central High Plains (CHP) based on VIC estimates.

Table 4.

Water balance comparison between models for study basins and western US

		Mosaic			Noah			VIC
	P (mm)	ET (mm)	SR (mm)	Recharge (mm)	ET (mm)	SR (mm)	Recharge (mm)	ET (mm)	SR (mm)	Recharge (mm)
Death Valley	185	176.3	7.0	1.6	119.9	12.1	52.8	151.8	10.4	22.5
Colombia	442	403.5	25.9	14.3	296.4	22.9	124.7	243.7	58.1	140.6
Williston	382	375.4	5.2	0.5	332.7	33.1	12.4	310.8	36.3	35.2
San Pedro	371	367.7	2.8	0.2	339.8	14.0	16.8	318.3	27.5	24.9
SHP	472	456.2	7.8	7.1	409.3	28.6	33.0	395.8	32.8	43.3
NHP	548	524.5	9.7	12.8	433.4	39.6	73.9	463.9	41.5	42.4
CHP	545	525.2	9.1	11.0	454.3	34.8	55.6	427.7	26.8	90.9
Central Valley	650	441.6	111.9	97.6	313.1	72.4	266.7	351.5	97.3	201.7
Spokane Valley	689	597.1	71.8	20.7	345.8	53.8	289.6	402.1	118.1	168.6
SLV	488	452.0	24.2	10.3	373.9	33.6	79.1	365.7	41.2	81.4
Western US	561	472.4	43.2	45.5	360.2	61.6	139.1	367.3	66.2	127.8

Table 5.

Water balance comparison (in %) between models for study basins and western US

		Mosaic			Noah			VIC			Literature
	P (mm)	ET (%)	SR (%)	Recharge (%)	ET (%)	SR (%)	Recharge (%)	ET (%)	SR (%)	Recharge (%)	Recharge (%)
Death Valley	185	95.4	3.8	0.9	64.9	6.5	28.6	82.2	5.6	12.2	1.5
Colombia	442	91.2	5.9	3.2	67.0	5.2	28.2	55.1	13.1	31.8	26.4
Williston	382	98.2	1.4	0.1	87.0	8.7	3.2	81.3	9.5	9.2	1.2
San Pedro	371	99.2	0.7	0.01	91.7	3.8	4.5	85.9	7.4	6.7	1.7
SHP	472	96.7	1.7	1.5	86.7	6.1	7.0	83.9	6.9	9.2	5.9
NHP	548	95.7	1.8	2.3	79.1	7.2	13.5	84.7	7.6	7.7	13.4
CHP	545	96.3	1.7	2.0	83.3	6.4	10.2	78.4	4.9	16.7	6.2
Central Valley	650	67.9	17.2	15.0	48.1	11.1	41.0	54.0	15.0	31.0	48.5
Spokane Valley	689	86.7	10.4	3.0	50.2	7.8	42.0	58.4	17.1	24.5	43.5
SLV	488	92.6	5.0	2.1	76.6	6.9	16.2	74.9	8.4	16.7	41.6
Western US	561	84.1	7.7	8.1	64.2	11.0	24.8	65.4	11.8	22.8	N/A

Mosaic consistently generated higher ET compared to Noah and VIC for most of the Western US (Fig 4). Overall, using the MODIS estimates as the standard, Mosaic overestimated ET by 36% (Fig 4). Noah and VIC follow a similar pattern with overestimation for low ET areas and underestimation for higher ET areas (Fig 3), but overall the bias was minimum (4% for Noah and 9% for VIC). Model ensemble mean overestimated MODIS ET by 15%. It should be noted that MODIS based ET is based on an empirical retrieval algorithm that requires calibration using ?data from a network of in situ measurements which themselves require calibration. MODIS area averaged ET rates have not yet been proven to be more accurate than LSMs, hence the biases shown here should not be interpreted as errors.

Fig. 4.

Average annual recharge estimates (1981–2010) from LSMs, BFI-based recharge, model ensemble mean and model spread

The specific breakdown of ET differs among models with Mosaic more dominated by passive processes and the other two more vegetative processes. Over the western US, Mosaic generated most of the ET through bare soil evaporation (47%) followed by transpiration (33%), canopy evaporation (18%) and sublimation (2%). Noah produced most of the ET through transpiration (41%) and bare soil evaporation (39%), followed by canopy evaporation (16%) and sublimation (4%). VIC on the other hand generated a majority of its ET as transpiration (82%) with other contributions from canopy (13%) and sublimation (4%). The very high contribution of ET through transpiration and limited contribution from bare soil by VIC is related to the tiling process in VIC which classifies a majority of land areas to some vegetation group. Moreover, the root zone depth extends throughout the 2m soil layer in VIC, while root zone depth is up to 1 m in the case of Mosaic and Noah (excluding forest land cover in Noah).

Relatively larger magnitudes of ET by Mosaic compared to other models could be ascribed to greater upward diffusion of water from deeper soil layers to the shallow root zone (Mitchell et al. 2004; Long et al. 2013). This process has a significant influence on the recharge estimates because the recharge in the LSMs is the function of water content in the bottom soil layer. These differences could be related to energy balance or water balance constraints (Mitchell et al. 2004; Peters-Lidard et al. 2011; Rodell et al. 2011; Xia et al. 2012b; Cai et al. 2014).

3.2. Comparing Recharge across the Western US: Among models and with BFI-based recharge

A similar spatial pattern (Table 6, Fig 4) of recharge was observed based on LSMs and also with BFI, although recharge rates vary among models as in the case of ET. While Mosaic consistently generated lower recharge compared to Noah, VIC and BFI, VIC overestimated recharge at low recharge zones and underestimated recharge in medium to higher recharge zones compared to Noah. Noah and VIC models overestimated recharge compared to BFI.

Table 6.

Pattern correlation (Pearson’s r) matrix on recharge estimates among LSMs, model ensemble mean (ENS) and BFI-recharge

	Mosaic	VIC	Noah	ENS	BFI
Mosaic	1.00	0.85*	0.91*	0.95*	0.74*
VIC	0.85*	1.00	0.91*	0.95*	0.77*
Noah	0.91*	0.91*	1.00	0.98*	0.86*
ENS	0.95*	0.95*	0.98*	1.00	0.83*
BFI	0.74*	0.77*	0.86*	0.83*	1.00

^*Statistically significant (p<0.05)

Average annual recharge rates varied between 0 and 4128 mm based on Noah, 0 and 3479 mm based on Mosaic, 0 and 2209 mm based on VIC, 0 and 3272 mm based on model ensemble and between 0 and 2031 mm based on BFI- based estimates at 1/8 degrees grid scale (Fig 4). The average recharge rates for the whole western US was estimated to be 139 mm based on Noah, 46 based on Mosaic, 123 based on VIC, 103 mm based on mean model ensemble and 82 mm based on BFI.

Although recharge rates differ among LSMs and BFI-based estimates, high and low recharge zones are similar among them (Fig 4). It was observed that BFI-based recharge captures higher recharge zones for the west coast, and it predicts lower rates for the Eastern US compared to LSMs (Fig 4). Mosaic showed slightly different patterns from BFI- recharge in other regions except the west coast mostly because Mosaic generated lower recharge compared to Noah, VIC and BFI (Fig 5). While Mosaic consistently generated lower recharge compared to Noah, VIC overestimated recharge at low recharge zones and underestimated recharge in medium to higher recharge zones compared to Noah (Fig 5). There was a stronger relationship between Noah and BFI-based recharge (R²:0.76) and model ensemble (R2= compared to VIC (R²:0.62) and Mosaic (R²:0.56) with BFI-based recharge (Fig 6?). Results from a Kolmogorov-Smirnov (K-S) test suggested that the recharge estimates from the three models as well as the ensemble average are significantly different from each other and from the BFI-based recharge estimates (Table 7). Spatial pattern and magnitude of Model mean were more similar to VIC and Noah model and did not necessarily better than those models when compared with the BFI based recharge (Fig 4). Model estimates differ less in the higher recharge zones in the east and west coast and more in the inner dry regions (Fig 4).

Fig. 6.

Comparison of model and ensemble model mean (ENS_MOD) recharge estimates with literature estimates (EST) in 10 study basins

Table 7.

K-S test for comparing distribution of recharge

K-S Test: H0: Samples are drawn from the same distribution
	Dmax
	Mosaic	VIC	Noah	ENS	BFI
Mosaic	0	0.66	0.65	0.68	0.61
VIC	0.66	0	0.046	0.12	0.37
Noah	0.65	0.046	0	0.11	0.38
ENS	0.68	0.12	0.11	0	0.34
BFI	0.61	0.37	0.38	0.34	0
	p-value
Mosaic	1	< 0.0001	< 0.0001	< 0.0001	< 0.0001
VIC	< 0.0001	1	< 0.0001	< 0.0001	< 0.0001
Noah	< 0.0001	< 0.0001	1	< 0.0001	< 0.0001
ENS	< 0.0001	< 0.0001	< 0.0001	1	< 0.0001
BFI	< 0.0001	< 0.0001	< 0.0001	< 0.0001	1

It should be noted that the BFI-based recharge itself is a very rough estimate and should not be treated as an observation due to high uncertainty related to this dataset. The dataset is likely to underestimate natural recharge in arid regions where ET is significant. Also, ground-water discharge to streams does not occur in “losing” streams which are more common in arid regions (Wolock 2003c). As a result, the BFI-based recharge consistently underestimated recharge compared to literature estimates in the study basins.

3.3. Evaluating LSMs’s recharge estimates at a basin scale

Mosaic consistently underestimated recharge across all the basins where estimates are available. Noah captured recharge reasonably well in wetter basins in xx, but overestimated it in drier basins (xx and xx). VIC overestimated recharge in the drier basins and underestimated it for wetter aquifers. Over the study basins, recharge estimates varied between 0.2 mm/yr to 97.6 mm/yr based on the Mosaic model, between 12.4 mm/yr to 289.6 mm/yr based on the Noah model, and between 22.5 mm/yr to 201.7 mm/yr based on the VIC model (Table 4 & Fig 6). The literature based recharge estimates ranged between 2.8 mm/yr and 315.5 mm/yr (Table 1, Fig 6). Model ensemble mean ranged between 14.1 mm/yr and 189.1 mm/yr (Fig 6) and was not ? necessarily better than Noah and VIC when evaluated with the literature estimates Ensemble mean is affected by highly skewed lower Mosaic estimates (Fig 6). Although models have predicted different recharge rates for the study basins, the patterns of predicted recharge were similar. Models agree in identifying drier and wetter aquifers (i.e. low vs. high recharge aquifers, Fig 6) although the driest and wettest aquifers identified by the models varied slightly among each other. The driest and wettest aquifers were the Death Valley aquifer and the Central Valley aquifer respectively based on the literature estimates (Fig 6). Similar results were obtained for the VIC model. However, Mosaic and Noah produced different results.

Based on literature estimates for individual aquifers in the region, about 1% to 49% of the precipitation becomes recharge (Table 5), lowest for the Williston basin and highest for the Central Valley. The Williston basin is in a semi-arid region, and thus the fraction of precipitation, which occurs primarily during summer that becomes recharge is relatively small. Recharge in the Central Valley comes from irrigation return flows, diffuse recharge directly from precipitation and from mountain system recharge in the form of leakage from streams originating in the Sierra Nevada Mountains. These basins where a higher proportion of P becomes recharge are more permeable and have lower ET rates.

Based on Mosaic, about 0.01% to 15% of precipitation becomes recharge in the study basins which is much lower than literature estimates (Table 5). Based on Noah, about 3.2% to 42% of P becomes recharge in the study basins which is within similar range compared to literature estimates (Table 5). Based on VIC, about 6.7% to 31.8% of the precipitation becomes recharge in the study basins which is slightly higher for the drier basin and slightly lower for the wetter basins when compared to literature estimates. Models agreed with the literatures estimates that basins like Spokane valley, Central Valley and Columbia produce higher percent of P as recharge (Table 5).

Overall, Mosaic consistently underestimated recharge significantly across the basins (Fig 6). VIC slightly overestimated recharge in the drier basins (Death Valley, Williston basin, San Pedro basin, SHP and CHP) but slightly underestimated in wetter basins (NHP, Colombia, SLK, Spokane Valley and Central Valley, Fig 6). Noah, on the other hand overestimated recharge in the drier basins but capture recharge reasonably well in the wetter basins except for SLV where it underestimated recharge (Fig 6). Thus, based on the analysis of these 10 basins, although none of the models were found to be capturing the recharge magnitude across the whole western US, it can be said that the Noah model showed a great promise in capturing the recharge in wetter regions. Mosaic seems to work better in drier basins which could just be an artifact that it underestimates recharge throughout the region. VIC seems to balance between Noah and Mosaic and seems to work for both dry and wetter regions if a single model is to be chosen across the western US. However overall, all three models (especially VIC and Noah) showed a lot of promise that with some advancements/modifications in hydrologic process representation and with some calibration at local scale/aquifer these models can be a useful tool for estimating current recharge and also for forecasting the effect of projected climate change on recharge.

3.4. Seasonality of Recharge

The models were fairly consistent with respect to the seasonality of recharge, which was largest during the spring. Over the study basins, models tend to agree on the seasonality of recharge occurring dominantly during spring months (MAM) except in the SHP basin (Fig 7). This spring time dominance of recharge is mostly pronounced in Mosaic. Since VIC has a more damped response to recharge, it produced similar recharge throughout the year for many basins. Seasonality of recharge did not necessarily follow the seasonality of precipitation in the aquifers (Fig 7).

Fig. 7.

Seasonlity of recharge based on LSMs and P based on NLDAS-2 data

Seasonally higher recharge in spring for the basins could be due to additional sources of snowmelt from winter (DJF) P which tends to melt at the beginning of spring when temperature is sufficient to melt but not high enough to lose a lot of water from evaporation (Dunne and Leopold 1978; Clark and Fritz 1997, Ajami et al. 2012; Jasechko et al. 2014) in addition to rain occurring in spring time. Several field monitoring studies in Sweden (Rodhe 1981), Idaho (Flerchinger et al. 1992), and the United States mid-west (Delin et al. 2007; Dripps 2012) have also found that the spring snowmelt constitutes the bulk of annual groundwater recharge at the middle latitudes examined here.

3.2.2. Potential reasons for differences in recharge estimates among models

Differences in recharge estimates among models can be attributed to differences in (1) ET calculations/estimates, (2) model structure particularly the thickness of bottom soil layer and (3) parameterizations. Recharge estimates from Mosaic were significantly smaller than those of Noah and VIC (Figs. 4, 5, 6). The lower estimates of recharge by Mosaic were directly related to Mosaic’s very high estimated ET. The model converted most of the precipitation to evaporation leaving much less water available to run off or infiltrate and percolate down as recharge. All of these LSMs characterize ET using soil moisture stress factor that impacts evaporation from the top soil layer and vegetation transpiration. As noted previously, the Noah model in NLDAS-2 has four soil layers with spatially invariant thicknesses of 10, 30, 60, and 100 cm. The first three layers form the root zone in non-forested regions, with the fourth layer included in forest regions. The Mosaic model has three layers with thicknesses of 10, 30, and 160 cm with the first two layers corresponding to the root zone. Mosaic has a greater ability to transfer water from the deep layer to the surface/root zone through vertical diffusion, and therefore shows higher ET rates under normal conditions (Long et al. 2013). This process dries up the bottom soil layer and leaves minimal water to become recharge through drainage. Although vertical diffusion does occur in Noah, the magnitude is much smaller compared to Mosaic. No vertical diffusion between 3^rd and 2^nd layer occurs in VIC model which also accounts for the sub-grid heterogeneity of vegetation and soil moisture. In VIC, rooting depth extends to the bottom layer unlike Noah and Mosaic. Although the general conceptualization and basic structure of the models are similar, they vary in certain processes and formulations. These differences in the parameterizations can give rise to large variability in the outputs depending upon the variables of interest. The multi-model analysis carried out under the Global Soil Wetness Project-2 (GSWP-2) (Dirmeyer et al. 2006) illustrated that LSM variables, especially those associated with snow processes (i.e., snow water equivalent) and soil water (i.e., soil moisture in the lower layers), have a large spread. The same is true for groundwater recharge (Xia et al. 2012a).

The thickness of the bottom layer, which is the source of recharge, is 160 cm in Mosaic model, 100 cm in Noah model, and of variable depth in VIC model. This variability leads to different amounts of free drainage. It is likely that differences in soil wetness, related to evapotranspiration and surface runoff rates, have a greater impact on the modeled recharge estimates than the free drainage formulations themselves (ref?). The relative bias analysis of soil moisture in the US showed that the models have small relative biases for the Eastern US where soils are normally wet but large relative biases in the western region where soils are drier (Xia et al. 2014). The disparity in mean annual evaporation and runoff ratio among the LSMs was also most obvious over the western mountainous regions (Xia et al. 2012a).

3.6. Limitations of LSMs

Like most LSMs, those used in this study, were developed using many simplifications necessary to represent complex physical processes across large spatial scales with limited computational power and with imperfect inputs. These LSMs have soil columns with depths of 2 m, divided into multiple (3 or 4) layers, while neglecting deeper soil moisture and groundwater. Vertical flows of soil water are estimated using the Richards equation while the horizontal flow is ignored. Groundwater recharge is parameterized by a gravitational percolation term, which is a linear/nonlinear function of bottom soil layer drainage affected by soil type, soil moisture content, and slope. It derives from a simple infiltration/saturation excess scheme used for both surface runoff and drainage. None of the models take into account the horizontal flow of groundwater. The partitioning of saturation excess into surface runoff and drainage and how they vary in space are also quite different from one LSM to another (Lohmann et al. 1998, 2004; Boone et al. 2004). Nevertheless, LSMs provide spatially and temporally continuous estimates of hydrological variables that would be impossible to obtain using observations alone, and often the results are surprisingly good considering their limitations (Dirmeyer et al. 2006; Syed et al. 2008; Jimenez et al. 2011; Wang et al. 2011; Li et al. 2015).

4. Summary and Conclusions

Three LSMs: Mosaic, Noah and VIC were used to estimate recharge and assess its spatial pattern and temporal trend in the western US. While Mosaic estimates were consistently low compared to the BFI based recharge, Noah recharge estimates were generally higher. VIC has mixed results with higher estimates at lower recharge zones and lower estimates at high recharge zones when compared with the BFI based recharge. Models were consistent in identifying high and low recharge zones although rates vary. When evaluated with published estimates of recharge in 10 aquifers across the western US, Mosaic was consistent in underestimating recharge significantly across all the basins. VIC slightly overestimated recharge in the dry aquifers and slightly underestimated it in the wetter aquifers. Noah captured recharge reasonably well for wetter basins (SHP, NHP, Colombia, Spokane and Central Valley), but overestimated it in the other basins. The models accurately identified low and high recharge aquifers, although their rankings based on recharge magnitude differed. The models were fairly consistent with respect to the recharge seasonality, which was largest during the spring. VIC’s recharge seasonality was dampened compared to Noah and Mosaic. This consistency among models was greater in the south than in the north, with its more snow dominated regions.

Overall, LSMs have the potential to capture the spatial and temporal patterns, as well as seasonality of recharge across the western US. Mosaic in particular requires calibration to capture the magnitude of recharge. Noah is more useful in capturing recharge in wetter regions with default parameters and VIC could be useful for both drier and wetter conditions but might require some calibration to improve estimations. In general, all three models (especially VIC and Noah) showed promise that with advancements/modifications in hydrologic process representation and with additional calibration at local/aquifer scale, these models can be a very useful tool for estimating current recharge and also for forecasting the effect of projected climate change on recharge.

Even though the source of meteorological forcing data produced as part of NLDAS-2 for all these LSMs was the same, differences in recharge estimates among models emerged due to differences in ET calculations/estimates, model structure particularly the thickness of the bottom layer, and parameterizations. Calibration of these LSMs could improve their ability to estimate recharge. However, it should be noted that carefully calibrating LSMs at a regional scale and at a grid level can be computationally and labor-intensive and observational data for calibration are also limited. Improving model inputs and adding process complexity especially associated with groundwater mechanisms in future could help reduce uncertainty in recharge estimates. Recharge estimates were highly controlled by precipitation and there was not much of an imprint of topography on the recharge estimates, even for major mountain chains, given that such patterns are evident in ET and likely in precipitation.