Augmenting Legacy Gaging Data with Emerging Datasets for Sustainable Water Management: Water Balance Analysis in the Upper Green River Basin, WY (1991–2023)

Abstract

Water balance calculations at the watershed scale are fundamental to water resource planning and the sustainable management of limited water supplies. These calculations rely on stream and canal gaging networks operated by local, state and federal entities, whose availability has varied over time due to cost, staffing constraints, and limitations on suitable gaging locations. The Green River Basin (GRB) above Fontenelle Dam in Wyoming illustrates this trend, where the number of operational stream gaging sites has varied over time and the majority of locations have less than 15 years of streamflow records. Recent advancements in the ability to perform streamflow reconstruction and estimate agricultural water use offer a new avenue for estimating the water balance for watersheds with discontinuous gage observations. But the use of these datasets and approaches has not been tested. Therefore, this paper proposes and tests a novel framework that combines discontinuous streamflow observations with new datasets (OpenET, ET-Demands, and GEOGLOWS) to calculate monthly water balances in the GRB from water year 1991 to 2023. Focusing on two main test basins, the Green River and the New Fork River, the integration of modern datasets enables the successful calculation of the water balance in the GRB with good agreement with downstream gaging records, achieving a Nash–Sutcliffe efficiency (NSE) of 0.88 for the New Fork River and 0.80 for the Green River. By improving the ability to quantify water balance components in data-limited basins, this framework supports more transparent water accounting and informed decision-making for sustainable water management, including irrigation planning, drought response, and long-term resource allocation in semi-arid river systems.

1. Practical Applications

A water balance calculation represents an effort to quantify all changes in water storage over a specific spatial and temporal extent. In the Western United States, these calculations are critical to sustainable water management due to the limited availability of water resources and the legal ramifications of water use and delivery. Incomplete and discontinuous observations of water resources often limit our ability to estimate the water balance over an extended period. Recent developments in remote sensing and continental- to global-scale hydrology modeling offer the ability to develop continuous time series of input data for extended water balance analyses. In this study, we determined that we could make use of discontinuous observed datasets, continental-scale remote sensing and global-scale hydrologic modeling datasets to perform water balance studies. We used two tributaries to the Colorado River in Wyoming to study this novel methodology.

2. Introduction

Water is a finite resource that is required by both anthropogenic activities and environmental systems, and they compete for their share of the resource [1]. Thus, water scarcity is a critical impediment to sustainable development [2]. Water balance calculations are fundamental to understanding scarcity and use of water within a natural water system (i.e., watershed). Water use and availability are particularly important in the arid American West. Recent droughts have strained water supplies [3], highlighting a need for a better understanding of water use, which can lead to improved and sustainable water management. For streams and rivers that traverse state boundaries, the need to understand water availability and water use is often legally significant, such as in the Colorado River Compact [4] and subsequent laws, treaties, acts, decrees, statutes, decisions, regulations, criteria, guidelines, and contracts that make up what is commonly referred to as the “Law of the River”. There is increasing pressure to understand and account for water use throughout the system.

Streamflow measurements are foundational to water balance studies. The United States Geological Survey (USGS) operates and maintains over 10,000 stream gages [5]. These gages operate over different timeframes, and they often only have a few years of overlapping stream data, if any. This makes it difficult to conduct a water balance based solely on streamflow observations. To emphasize this point, the reader is encouraged to visit the USGS “Gages Through the Ages” website (https://www.usgs.gov/tools/gages-through-ages-data-visualization-story (accessed on 13 May 2025)) to visualize a map of active United States stream gages through time.

Computational hydrologic models are a popular means of simulating streamflow where and when gaged observations of streamflow are not available. For specific watersheds, models such as the Soil and Water Assessment Tool (SWAT) can be developed, calibrated, and validated to simulate a continuous time series of streamflow and other hydrologic system components using observations of past meteorology [6]. Over the past 10–15 years, the ability to estimate hydrologic systems accurately and at hyper-resolution has been improving at continental to global scales [7]. These hyper-resolution hydrologic models can use reanalysis of past meteorology to estimate streamflow for streams throughout entire continental-scale hydrologic systems and now dependably offer operational products. A prominent example of the progress made in continental-scale hydrology models that provide streamflow output is the operational National Water Model (NWM), which provides time series of estimated streamflow for streams in much of the United States [8]. At a global scale, Version 2 of the Group on Earth Observations Global Water Sustainability (GEOGLOWS) European Centre for Medium Range Weather Forecasts (ECMWF) Streamflow Model provides readily available reanalysis streamflow data for most streams and rivers for the world from 1940 to the present. For brevity, the GEOGLOWS ECMWF Streamflow Model Version 2 and associated data will hereafter be referred to as GEOGLOWS. Despite extensive calibration efforts, one main issue with these streamflow simulation models is that they cannot totally replicate the accuracy of stream gage streamflow measurements. To improve the accuracy of GEOGLOWS data at gage locations, Sanchez Lozano et al. [9,10] developed and tested a flow duration curve (FDC) method, extended from Farmer et al. [11], that uses monthly FDCs from gage locations and bias-corrects synthetic streamflow estimates from GEOGLOWS. Ultimately, this allows the use of stream gage data with full or partial records to bias-correct the GEOGLOWS streamflow from 1940 to the current. This bias-corrected streamflow can be used to reconstruct the missing observations at the gage location and expand the temporal breadth of water balance studies.

Beyond streamflow, canal diversions for irrigation play a crucial role in many water balance assessments in agricultural watersheds in the Western United States. The 17 westernmost states in the contiguous United States (CONUS) accounted for 81% of total irrigation withdrawals and 74% of total irrigated land area in the United States [12]. Although states often publicly report canal diversions, the available records are often limited and typically only account for water diverted, not the amount of water that returns to the river from runoff, seepage, or tailwater returns. These return flows are also not properly accounted for in current hydrologic modeling platforms, such as the NWM [13] or GEOGLOWS. Recently, Follum et al. [14] found that canal diversions could be estimated based on consumptive use by the crops ($C_u$) and assumed system efficiencies related to the crop type, irrigation method, and conveyance method.

$C_u$ represents the amount of water depleted from the system during crop production, with crop evapotranspiration (${ET}_a$, cm) often being the primary component of $C_u$ in the Western United States. $C_u$ has typically been calculated based on point-scale measurements that have then been applied to larger domains (i.e., the entire watershed). Recent advancements have led to the development and operational use of remotely sensed and satellite-based technologies in conjunction with energy-balance approaches [15,16,17,18] to estimate ${ET}_a$ at the field scale (~30 m spatial resolution). OpenET [19,20] is a platform that provides gridded 30 m resolution estimates of ${ET}_a$ from both individual models and ensemble averages. The Google Earth Engine implementation of the Mapping Evapotranspiration at high Resolution with Internalized Calibration model (eeMETRIC), based on Allen et al. [16,17], and the Operational Simplified Surface Energy Balance model (SSEBop), based on Senay et al. [21], are two of the ensemble members included within OpenET. Volk et al. [20] found that SSEBop had a higher accuracy in calculating evapotranspiration in wetland and riparian sites, while the U.S. Bureau of Reclamation (USBR) and the Upper Colorado River Commission chose eeMETRIC as a unified approach for measuring agricultural water use in the Upper Colorado River Basin [22]. Accounting for effective precipitation and soil moisture, the USBR collaborated with the Desert Research Institute to calculate monthly vector-based field-scale $C_u$ data throughout the Colorado River Basin using OpenET eeMETRIC data within the ET-Demands model [23,24].

Despite these recent advancements in the ability to perform streamflow reconstruction and estimate agricultural water use, we are not aware of previous studies that combined these datasets to produce a water balance estimate in data-limited basins of the Western United States. Therefore, we identified a need for an integrated multi-source water balance calculation technique for river basins with discontinuous gage systems. We hypothesized that a water balance estimate could be developed for determining periods of gains and losses in streamflow that are caused by non-agricultural components of the hydrologic system. Therefore, the objectives of this study were to answer the following research questions:

• How well do bias-corrected large-spatial-extent streamflow reconstructions and estimated $C_u$ perform in Western United States watersheds?
• Can we combine streamflow reconstructions, estimated $C_u$, and estimates of riparian evapotranspiration to accurately estimate a monthly water balance?
• Does the estimated water balance provide insight into periods of time when the stream is gaining or losing water as it travels through the basin’s river network?

3. Materials and Methods

Figure 1 presents the framework of the analysis undertaken in this research using a system flowchart. To answer our research questions, we enacted this framework on the GRB above Fontenelle Reservoir in Wyoming. First, using GEOGLOWS streamflow estimates, we reconstruct bias-corrected streamflow data at all stream gage sites. Second, we calculate diversion amounts using ET-Demands data. Third, we calculate riparian evapotranspiration by using calculations from the SSEBop method within OpenET. We then conduct a water balance for the New Fork River and Green River (the New Fork River is a major tributary to the Green River), comparing the water balance values to gage records. Lastly, we discuss data requirements and provide insight for application of the water balance approach in other basins throughout the Western United States. The proceeding sections offer additional details about each component of Figure 1.

3.1. Test Basin

The GRB above Fontenelle Reservoir has an area of 9881.7 km² and serves as the test case for this study (Figure 2). The New Fork River is a major tributary to the Green River, with both rivers being fed by high-elevation snowpack from the Wind River Mountain Range (peak elevation of 4207 m) and the Wyoming Mountain Range (peak elevation of 3470 m). Seasonal runoff from glacial melt also contributes to streamflow generation during warm periods of the year [25]. Both the New Fork River and the Green River start in the high-elevation mountains and merge prior to flowing into Fontenelle Reservoir (approximate elevation of 1980 m). The majority of the GRB is an intermontane desert with annual precipitation ranging from 152.4 cm in the higher-elevation mountains to 17.8 cm at Fontenelle Reservoir [26]. Although most precipitation occurs as snow, the basin does receive some rainfall, with most rainfall occurring in short, intense storms [27]. The GRB is an important source of water for agriculture, mining, and municipal uses. The GRB is also Wyoming’s largest contribution to the Colorado River, and therefore subject to certain portions of the “Law of the River”. As settlers began moving to the GRB in the 1800s, irrigation systems were found to be necessary for crop production. In the 1900s, the development of dams and reservoirs accompanied the expansion of irrigation systems in the GRB. This expansion included the completion of Fontenelle Dam in 1964. Today, most of the irrigation within the GRB is in the riparian corridors along rivers and creeks and is supplied via unlined canal systems [27]. The GRB was chosen as a test case because it has a history of discontinuous stream gaging and the movement of water in the basin is heavily influenced by agricultural usage. Furthermore, understanding water balance in the GRB is important as states that encompass the Colorado River Basin seek to undertake sustainable water management practices.

3.2. Water Balance

A water balance is a comparison of inflows to outflows for a given domain. For this study, two domains are investigated, as shown in Figure 2. The first is the New Fork River between the upstream USGS stream gage 09193000 (New Fork River below New Fork Lake) and the downstream USGS stream gage 09205000 (New Fork River near Big Piney, WY), approximately 1950 m upstream of the confluence with the Green River. The second is the Green River, with an upstream boundary using USGS stream gage 09188500 (Green River at Warren Bridge, near Daniel, WY) and a downstream boundary at USGS stream gage 09209400 (Green River near La Barge, WY) that is just upstream of Fontenelle Reservoir. The Green River domain encompasses the entire basin (including the New Fork River), and therefore also uses the USGS stream gage 09193000 as an upper boundary for the New Fork River. The monthly average streamflow at these stream gages designates the inflow and outflow boundaries for each water balance we calculated, with the outflows being designated as $Q_{outflow}$ (m³ s⁻¹) and the inflows being designated as tributaries ($Q_{tributary}$, m³ s⁻¹). In addition to USGS stream gages 09193000 and 09188500, there are numerous other tributaries ($Q_{tributary}$) that contribute to the basin (Figure 2), all with varying degrees of streamflow measurement. In 1940, there were over 20 USGS stream gages in the GRB. Today, there are 13 stream gages operated by the USGS and State of Wyoming, with the majority being seasonally operated and having less than 15 years of record. As will be discussed, a total of 21 stream gages (current and historical) are used in this study.

Transit losses occur as water moves along the river, including transpiration from riparian zones ($Q_{riparian}$, m³ s⁻¹), evaporation from the water surface ($Q_{water surface evap}$, m³ s⁻¹), and changes in bank storage (${\Delta Q}_{bank storage}$, m³ s⁻¹). ${\Delta Q}_{bank storage}$ is negative if water is filling the porous zones of the banks (typically during high flow events in the spring runoff), or positive if the water is draining back to the river (typically on the falling side of the hydrograph after the spring runoff). During the peak of the irrigation season (typically April through July), one of the largest components of the water balance is diversions for irrigation canals ($Q_{diversion}$, m³ s⁻¹). A portion of $Q_{diversion}$ returns to the river as both surface water (tail water returns and overland flow) and groundwater, caused by saturation of the canals and fields during the irrigation season [29,30]. The magnitude and timing of these return flows are difficult to determine and are therefore grouped into a single term, $Q_{gain/loss}$ (m³ s⁻¹), which will be evaluated separately. The residual term $Q_{gain/loss}$ represents the combined effect of several hydrologic processes that cannot be separated with the data and time scale used in this study. Previous work in Wyoming and similar basins has shown that gains and losses in streamflow are influenced by a range of interacting processes—such as evapotranspiration, bank and channel storage, groundwater exchange, and irrigation return flows—and that it is often not possible to isolate these components individually because of data limitations [31]. For this reason, water balance studies typically rely on combined or “net” loss terms that represent the overall system response rather than individual physical processes (see, for example, [32]).

$Q_{gain/loss}$ is positive if the stream is gaining water (such as from natural springs known to be in the GRB), or negative if water is being removed from the river (such as water seeping into the groundwater system). Thus, the water balance at the outlet of the study domains ($Q_{est}$, m³ s⁻¹)—USGS stream gage 09205000 for the New Fork River and USGS stream gage 09209400 for the Green River—can be calculated as shown in Equation (1):

$$Q_{est}=Q_{tributary}-Q_{diversion}-Q_{riparian}-Q_{water surface evap}+Q_{gain/loss}+{\Delta Q}_{bank storage}$$

(1)

The aim of this study is to calculate $Q_{est}$ for the New Fork River and Green River using the datasets described in the proceeding sections and then compare to $Q_{outflow}$. Because $Q_{gain/loss}$ is unknown, it will initially be left out of the water balance and then further evaluated by replacing $Q_{outflow}$ with $Q_{est}$ in Equation (1) and solving for $Q_{gain/loss}$.

3.3. Streamflow Reconstructions ($Q_{tributary}$)

GEOGLOWS provides daily retrospective streamflow estimates at a global scale from 1940 to near present. These data represent naturalized historical flow [33,34] and do not include the Stream Analysis for Bias Estimation and Reduction (SABER) method for bias correction that will be deployed in future versions of the dataset [35,36]. However, at streams with gage records that overlap spatially and temporally with the GEOGLOWS hydrography (as shown in Figure 1), we deployed the gage-based bias correction techniques developed by Sanchez Lozano et al. [9,10] and Farmer et al. [11] that are available in the GEOGLOWS Python package version 1.7.0 [37,38]. To deploy the GEOGLOWS Python package, we used the observed time series of discharge from a stream gage, which overlapped with a fraction of the discharge time series available in the GEOGLOWS streamflow data. There were two sources of observed discharge that were combined in several instances: the U.S. Geological Survey’s National Water Information System (NWIS) and the state of Wyoming’s State Engineer’s Office (SEO). The observed and simulated time series were used to construct observed and simulated FDCs for each month of the year. The GEOGLOWS Python package converted the simulated discharge into a bias-corrected discharge by taking the simulated streamflow at an ordinate, determining that simulated streamflow’s monthly exceedance probability on the simulated FDC, and then finding the streamflow on the observed FDC that matches that monthly exceedance probability on the simulated FDC. This streamflow that was extracted from the observed FDC represented the bias-corrected streamflow. Figure 3 illustrates this bias correction process. This technique enabled us to use a limited time series of observed streamflow data to bias-correct simulated streamflow covering a much larger timeframe, effectively expanding the temporal coverage of streamflow data for this investigation. This method assumed stationarity in the monthly FDCs as a product of the limited observations of streamflow used in this study. A combination of observed data at each stream gage and bias-corrected GEOGLOWS streamflow to fill in the time gaps within the observed data provided a continuous estimate of $Q_{tributary}$ in Equation (1) for strategic stream locations (Figure 4). Using these stream gage reconstructions, the water balance also gained greater spatial coverage of the watershed, as is illustrated in Figure 4 by the expanded domain. Active stream gages cover approximately 760.5 km² (23.4%) of the New Fork River Basin (NFB; 3247.6 km²), and 2680.5 km² (27.1%) of the GRB (9881.7 km²). By including the stream gage reconstructions, approximately 1529.2 km² (47.1%) of the NFB and 5062.4 km² (51.2%) of the GRB are covered. We did not include ungaged tributary contributions in the water balance because the remaining streams, as illustrated in Figure 3, exhibit limited streamflow due to their predominantly desert drainage areas.

3.4. Canal Diversion ($Q_{diversion}$) Based on Consumptive Use ($C_u$)

The State of Wyoming measures and publicly reports several canal diversions ($Q_{diversion}$) within the GRB (https://seoflow.wyo.gov/ (accessed on 13 May 2025)). The state does not measure or record return flow data. The majority of continuous, publicly available $Q_{diversion}$ data begins in May of 2012, which limits the timeframe for a water balance. Additionally, the $Q_{diversion}$ data does not capture all water diverted for irrigation within the GRB. Therefore, we use monthly $C_u$ data as a proxy to estimate the total irrigation water diverted within both domains. Of the water diverted from rivers for irrigation ($Q_{diversion}$), a portion of water is consumed by the crops ($C_u$), some is accounted for as incidental losses (i.e., evaporation, deep percolation, and/or conveyance evapotranspiration), and some water is returned back to the river (component of $Q_{gain/loss}$). Irrigation systems inherently have inefficiencies, meaning $Q_{diversion}>C_u$. The withdraw ratio ($W_r$) is expressed as the ratio of $Q_{diversion}$ to $C_u$ for the crops irrigated. $W_r$ varies greatly depending on the conveyance system, irrigation methods, and hydrologic conditions (precipitation, snowpack, soil moisture, etc.). $Q_{diversion}$ can therefore be estimated as a function of $C_u$ and $W_r$, as shown in Equation (2):

$$Q_{diversion}=C_u{W}_r$$

(2)

A comprehensive study of $W_r$ within the GRB does not exist; however, Hasfurther [39] calculated that 72% of diverted water for irrigation returns to the river system within 1 year at a 72.5 km² test area near the New Fork River. This indicates that even if the majority of the remaining 28% of diverted water was consumed by crops ($C_u$), that the $W_r$ value is close to 4 during an average precipitation year. $W_r$ will be further evaluated using observed canal diversion and $C_u$ data in Section 4.

$C_u$ can be calculated using the ET-Demands dataset [23], which has both a spatial and a tabular component. The spatial component, as shown in Figure 5, maps field locations and provides the area of the field ($A_f$, m²). The tabular component provides a monthly time series of ${ET}_a$ and effective precipitation ($EP$, cm) data for each irrigated field between 1991 and 2023. The ET-Demands dataset was provided by the USBR and Desert Research Institute. For each field, monthly $C_u$ values are calculated by subtracting $EP$ from ${ET}_a$, then multiplying by $A_f$. As described in the ET-Demands user documentation [40], excess effective precipitation (${ET}_a- EP$) from the previous month is carried over into the next month for each field, which provides a method to account for antecedent soil moisture. Any fields that are hydrologically above the $Q_{tributary}$ locations were not used in the calculation of $C_u$. The diversions to irrigate these fields, as well as any return flows from these fields, would have been accounted for within the $Q_{tributary}$ values.

3.5. Transit Losses

Transit losses occur through several mechanisms, including ${\Delta Q}_{bank storage}$, $Q_{water surface evap}$, $Q_{riparian}$, and $Q_{gain/loss}$. Understanding these components and their relative loss contributions is crucial for an effective water balance. Estimated conveyance losses in Wyoming range between 0.21% and 1.03% per km (0.34% and 1.66% per mile) [30]. The New Fork River demonstrates higher losses at 0.53% per km (0.85% per mile), while the Green River shows total loss rates of 0.2% to 0.4% per km (0.4% to 0.6% per mile) for short-term releases [41]. It is important to note that these transit loss estimates were based on several measurements throughout the system during short-term releases of reservoir water, and do not account for the overall gains and losses within the system during longer timeframes.

$Q_{water surface evap}$ typically contributes about 6% of total transit losses [31], with turbulent flow increasing evaporation by up to 15% compared to still water [42]. Precise quantification of $Q_{water surface evap}$ remains challenging due to the complex interaction between vegetation, groundwater, and surface water systems [31]. The Western Regional Climate Center’s pan evaporation data in combination with the length and average width of the rivers is used in this study to calculate monthly $Q_{water surface evap}$ values for the Green and New Fork Rivers.

$Q_{riparian}$ becomes increasingly significant over longer periods, particularly in semi-arid regions [43]. On the Green River, $Q_{riparian}$ rates peak during summer months when vegetation actively draws from groundwater sources, with rates declining significantly during winter dormancy [44,45]. This study uses OpenET’s SSEBop model to estimate $Q_{riparian}$ losses [46], because Volk et al. [20] found that SSEBop had the highest relative accuracy for wetland/riparian sites and avoided the high-bias tendency in the spring observed in several other OpenET models. The SSEBop model uses satellite-derived surface temperature data to determine water stress and calculate monthly ${ET}_a$ rates [21]. This approach offers operational advantages through its simplicity and scalability, enabling consistent ${ET}_a$ estimates across large geographic areas without requiring extensive ground-based measurements. This SSEBop-derived riparian ET should be interpreted as an operational estimate rather than a site-validated measurement. No local eddy-covariance, lysimeter, or other ground-based riparian ET observations were available in the study area to independently validate OpenET estimates. Therefore, SSEBop-derived riparian ET was used as an operational estimate supported by published model intercomparison results rather than as a locally calibrated measurement.

Riparian corridors were delineated using the Colorado Riparian Polygons (CO-RIP) dataset, which combines topographic and hydrologic data through the Valley Bottom Extraction Tool [47]. This standardized approach provides a consistent framework for riparian mapping across diverse regions. Monthly $Q_{riparian}$ values were then calculated for each domain using SSEBop ${ET}_a$ values within the CO-RIP corridors.

$Q_{bank storage}$ can account for up to 70% of initial losses in alluvial stream sections [32]. This effect is particularly strong in areas with permeable deposits and strong stream–aquifer connectivity [48]. However, the effects of ${\Delta Q}_{bank storage}$ stabilize after approximately 10 days, with stored water returning to the stream [41]. ${\Delta Q}_{bank storage}$ is often characterized as a timing shift in water rather than a permanent loss [41]. Due to $C_u$, $Q_{water surface evap}$, and $Q_{riparian}$ data all being on a monthly timestep, all water balance calculations will also be at a monthly timestep. Therefore, the effects of ${\Delta Q}_{bank storage}$, which often occur on a sub-monthly timescale, are considered negligible within the water balance. Removing ${\Delta Q}_{bank storage}$ and combining Equations (1) and (2) results in Equation (3):

$$Q_{est}=Q_{tributary}-C_u{W}_r-Q_{riparian}-Q_{water surface evap}+Q_{gain/loss}$$

(3)

where $Q_{gain/loss}$ is the only unknown term on the right side of the equation.

4. Results and Discussion

4.1. Streamflow Reconstructions

Twenty-one USGS stream gages were reconstructed using GEOGLOWS flow rates and bias-correcting methods. For each stream gage that was reconstructed,

Table 1. Summary of bias correction results for monthly streamflows at USGS stream gages. USGS gage 09195000 was used in the subsequent water balance but only had approximately one year of data and could not be used in this testing phase.

USGS Stream Gage ID	GEOGLOWS Stream Order	GEOGLOWS ID	Year Gage Active	Total Days of Observed Data	NSE Before Bias Correction	NSE After Bias Correction	PBIAS Before Bias Correction	PBIAS After Bias Correction	RMSE Before Bias Correction	RMSE After Bias Correction
09188500	4	710265920	1931–2024	33,696	0.87	0.91	2.36	−4.65	5.52	4.65
09189000	4	710257986	1938–1954	5397	−3.60	0.48	−176.17	30.42	5.44	1.83
09190000	3	710214342	1931–1985	9495	0.65	0.90	−49.29	11.71	2.27	1.23
09191500	4	710256014	1938–1954	5837	0.36	0.66	−44.54	34.61	3.03	2.21
09193000 *	2	710132986	1938–2013 2015–2024	14,689	0.39	0.82	48.37	−1.40	2.22	1.21
09194500	3	710214338	1938–1941	1188	−6.44	−0.16	−172.33	47.83	0.54	0.21
09195500	2	710178629	1938–1941	1157	−23.23	−0.10	−6.14	−0.44	0.20	0.04
09196000	4	710269897	1938–1944	2253	−4.40	0.86	−103.25	19.50	4.50	0.73
09197000 **	3	710242119	1910–2016	9278	0.51	0.86	47.22	22.43	4.68	2.52
09198500	3	710228232	1938–1971	12,052	0.76	0.94	26.87	6.39	2.68	1.32
09199500	3	710242121	1938–1971	12,018	0.76	0.93	10.60	8.65	1.13	0.61
09202000 ***	4	710257993	1938–1973 2012–2023	14,370	0.34	0.84	54.91	10.63	10.74	5.29
09203000	3	710303635	1938–1992	19,723	0.43	0.87	44.06	−8.03	3.81	1.83
09204000	3	710250063	1938–1971	11,244	0.47	0.92	37.33	6.34	2.15	0.81
09205000	6	710369111	1954–2025	25,690	0.87	0.90	10.49	1.48	8.95	7.66
09205500	3	710287763	1915–1972	15,158	0.86	0.93	16.40	3.29	0.88	0.61
09206000 ****	3	710226262	1939–1954 2012–2024	7197	0.44	0.77	25.43	−4.51	0.91	0.59
09207500	3	710277851	1938–1942	1522	−0.03	0.84	−37.79	2.85	0.71	0.28
09207700	4	710371105	1965–1973	1919	−0.59	0.38	−120.27	48.33	0.41	0.25
09208500	3	710539754	1940–1949	3286	0.61	0.80	28.10	16.60	1.73	1.22
09209400	6	710561578	1963–2024	22,373	0.77	0.89	−22.11	−4.53	23.20	15.81
			Average	10,931	−1.39	0.73	−18.08	11.78	4.08	2.42

* NWIS data was augmented with State of Wyoming data collected from 2015 to 2024 at a co-located gage. ** NWIS data was augmented with SEO data collected from 2015 to 2024 at a co-located gage. USGS data was summed with SEO/USGS data collected at canal gages Freemont/09196940 and High Ditch/09196960 at coincidental times. *** NWIS data was augmented with SEO data collected from 2012 to 2023 at a co-located gage. **** NWIS data was augmented with SEO data collected from 2012 to 2024 at a co-located gage.

shows the USGS gage identification number, years the gage was active, and the nearest GEOGLOWS stream segment. For each, a gage split-sample calibration and validation process occurred, where 70% of the $Q_{obs}$ data were withheld for training the bias correction method and the remaining 30% of data were used for testing. We computed the $NSE$ [49], percent bias (PBIAS), and root mean square error (RMSE) between the monthly mean observed flow ($Q_{obs}$, m³ s⁻¹) and both the original monthly mean GEOGLOWS streamflow (m³ s⁻¹) and the monthly mean bias-corrected streamflow ($Q_{bc}$, m³ s⁻¹). $NSE$ was calculated using the HydroStats Python package version 0.78 [50,51]. Negative PBIAS values indicate over-estimation and positive PBIAS values indicate under-estimation. All gages contributing to $Q_{tributary}$ were tested except USGS stream gage 09195000 because it did not have enough observed data to perform the 70/30 training and testing split. We also included three additional USGS gages in the basins to test our bias correction process, including the comparison stream gages 09205000 and 09209400, along with an additional gage on the New Fork River (USGS stream gage 09196000). As shown in Table 1, the original $NSE$ values ranged from −23.23 to 0.87 with an average of −1.39, PBIAS ranged from −176.16 to 54.91 with an average of −18.08, and RMSE ranged from 0.20 to 23.20 with an average of 4.08. However, after bias correction, $NSE$ for $Q_{bc}$ ranged from −0.16 to 0.94 with an average of 0.73, PBIAS ranged from −8.03 to 48.33 with an average of 11.78, and RMSE ranged from 0.04 to 7.66 with an average of 2.42. In all but one PBIAS instance at gage 09188500, the individual metrics improved with bias correction. Composite metrics improved substantially for the gages analyzed, with average RMSE decreasing by 41% in our gage sample.

Acceptable $NSE$ has varied by study, but Moriasi et al. [52] established $NSE>0.50$ as a satisfactory watershed model performance at a monthly time step after a thorough review of the literature. After bias correction, this threshold was reached by 17 of the 21 stream gages in the testing sample. Figure 6 demonstrates a general bias of GEOGLOWS to over-predict streamflow during wet periods, but this is generally resolved by use of the bias correction method. Figure 6 shows a comparison between $Q_{obs}$ and $Q_{bc}$ at the four sites with $NSE<0.50$. Here, we see that despite a relatively low $NSE$, the $Q_{bc}$ provides a much better estimate of the magnitude of $Q_{obs}$ and generally follows the same behavior as $Q_{obs}$. In fact, RMSE was reduced by 38–79% at these four gage locations, despite a relatively low NSE. However, there are noted instances of greater improvements that lead to $NSE>0.50$, such as those found at USGS gage 09196000 (Figure 7), which led to more skillful $Q_{bc}$ during the testing period. A strong predictor of gage performance after bias correction was performance of the original GEOGLOWS dataset. For instance, Figure 8 illustrates an elasticity model for predicting RMSE after bias correction using RMSE before bias correction. Figure 8 indicates through visualization and a high R² (coefficient of determination) that log-transformed RMSE before bias correction is correlated with RMSE after bias correction. Figure 8 also demonstrates that bias-corrected GEOGLOWS streamflow RMSE was elastic with respect to RMSE in the original GEOGLOWS data and that a 1% increase in RMSE before bias correction can yield a 1.094% increase in RMSE after bias correction [53].

Use of the bias correction method allowed for a continuous monthly flow record for each stream gage location between 1991 and 2023, giving us an approximate $Q_{tributary}$. Bias correction generally improved the performance of the GEOGLOW streamflow estimates. But what was of importance was how errors from our bias-corrected streamflow accumulated to distort and create uncertainty in the $Q_{tributary}$ at our comparison gages. Therefore, we summed up the average error at all gage locations for each month above each comparison gage. This sum included both under- (positive error) and over-estimation (negative error) of the original GEOGLOWS and bias-corrected GEOGLOWS streamflow to give a composite error metric that assesses seasonal performance of the $Q_{tributary}$ estimates, excluding any error that results from excluding desert tributaries. Figure 9 illustrates this assessment of aggregate monthly errors in $Q_{tributary}$ above 09205000 on the New Fork River and 09209400 on the Green River before and after bias correction. This assessment was not a split sample but contains errors from the entire gage record of observations. What we see is that aggregate monthly error is significantly reduced when the bias correction technique is applied to the original GEOGLOWS data. The aggregate streamflow error magnitude peaked in May at 3.72 m³ s⁻¹ for the New Fork River at 09205000 and in June at −4.08 m³ s⁻¹ for the Green River at 09209400 once bias correction was applied. For both comparison gages, May and June have the highest magnitude of composite error, with all other months having a composite error magnitude of <1 m³ s⁻¹. As we see in the Section 3.2 below, these magnitudes of error were a relatively small fraction of the overall water balance in May and June for our comparison gages and gave us confidence in the utility of $Q_{bc}$ to compute an accurate water balance. However, this error could have contributed $Q_{gain/loss}$ identified in the Section 3.2 below, particularly in May and June. This aggregation of error was an estimate of uncertainty that was passed from $Q_{tributary}$ to the water balance calculations.

4.2. Consumptive Use ($C_u$) and Canal Diversion ($Q_{diversion}$)

We use $C_u{W}_r$ as a proxy for $Q_{diversion}$ (Equation (2)) because $Q_{diversion}$ data is limited. $C_u$ was calculated using ET-Demands data for both the New Fork River domain and the Green River domain. To verify the methodology, we calculated $C_u$ values for the entire Green River Basin (which encompasses our domains) that are within Wyoming and compared them to the state-level values calculated by USBR [54] for the same region (Figure 10). The USBR [54] values are between 9% and 10% higher than the values we calculated. The USBR [54] values account for incidentally irrigated areas, defined by the State of Wyoming as areas surrounding fields that are incidentally irrigated by seepage or surface runoff, and are assumed to be approximately 10% of $C_u$ [55]. It should be noted that the incidental irrigation values are currently being revaluated by the Upper Colorado River Commission and the Upper Division States in coordination with USBR. When the incidental irrigated area (10% of $C_u$) is accounted for, the values calculated using the methodology shown in this paper and the USBR [54] are nearly identical.

Figure 11 shows a comparison of $C_u$ and $Q_{diversion}$ data at five canals within the GRB. The canals were chosen because they have well-defined irrigation domains. Although diversions for irrigation in the GRB occur from April to September, the majority of irrigation diversions for crops often occur in April through July. Therefore, in this study, we focus on the diversion flows from April to July and refer to this period as the study irrigation season. The overall $W_r$ values during the study irrigation season ranged between 3.7 and 8.2, with a median of 5.1, which is close to the value of 4 previously discussed based on the Hasfurther [39] study. $W_r$ values will vary based on the crop type, irrigation method (sprinkler vs. flood), canal losses, etc., and therefore, for this study, the $W_r$ value is set to the median value of 5.1 for the entire irrigation season. Figure 8 also shows that $Q_{diversion}$ values can vary greatly from month to month. Based on available canal records within the GRB during the study irrigation season, 5.2% of the irrigation water is diverted in April, 23.8% is diverted in May, 42.5% is diverted in June, and 28.5% is diverted in July. To reflect the monthly distribution of irrigation water, the volume of water required for each study irrigation season (April–July) is first calculated using Equation (2). This volume of water is then distributed to monthly $Q_{diversion}$ values using the monthly percentages (5.2% in April, 23.8% in May, 42.5% in June, and 28.5% in July). This approach retains the $W_r$ values throughout the study irrigation season while also allowing for monthly variability in flow rates.

4.3. Water Balance Evaluation

Figure 12 shows the water balance calculation (Equation (3)) from 1991 to 2023 for the New Fork River (USGS gage 09205000) and the Green River (USGS gage 09209400). Figure 13 depicts relative contributions of $Q_{tributary}$, $C_u{W}_r$, and $Q_{riparian}+ Q_{water surface evap}$ to the water balance calculation. Figure 13 shows box-and-whisker plots that demonstrate that the dominant hydrologic processes we accounted for are $Q_{tributary}$ and $C_u{W}_r$, with $Q_{riparian}+ Q_{water surface evap}$ being a much more minor component. Because it is unknown, $Q_{gain/loss}$ is initially set to 0 in the water balance and will be explored further in the next section. Figure 12 shows generally good agreement between the water balance estimates and gage outflows. In the New Fork watershed above USGS gage 09205000, the $NSE$ is 0.88. In the Green River watershed above USGS gage 09209400, the $NSE$ is 0.80. These $NSE$ values indicated that a large proportion of observed $Q_{obs}$ variation is explained by the water balance technique even without accounting for $Q_{gain/loss}$. However, the calculated water balance underpredicts $Q_{obs}$ during portions of the April–July irrigation season, indicating that one or more positive inflow components are not fully represented or that one or more loss terms are overestimated. A $Q_{est}$ less than $Q_{obs}$ indicates that a source of water is missing, such as return flows from irrigation diversions, ungaged tributaries, local inflows, etc. A $Q_{est}$ greater than $Q_{obs}$ indicates that there is an unaccounted-for loss in the river, such as recharge to the groundwater system, unaccounted for diversions, etc. $Q_{gain/loss}$ represents the unaccounted losses and gains in the water balance, which are further evaluated in the proceeding section.

Negative values of $Q_{gain/loss}$ do not indicate negative streamflow. Rather, they indicate months when the measured outflow is less than the water balance estimate after accounting for represented inflows and losses, implying a net residual loss from the modeled reach. These values may reflect physical processes such as groundwater recharge, temporary bank storage, ungaged depletions, or riparian ET, but may also arise from timing mismatches among monthly water balance components. For example, irrigation diversions, canal seepage, field-scale deep percolation, and shallow groundwater return flows may occur over different monthly timescales, causing losses in one month and delayed gains in subsequent months. Because these processes are not explicitly represented in the current accounting framework, $Q_{gain/loss}$ should be interpreted as a net residual closure term rather than as a separately validated physical flux.

4.4. Evaluation of $Q_{gain/loss}$

$Q_{gain/loss}$ was calculated by substituting $Q_{obs}$ (USGS gage 09205000 for the Green River and USGS gage 09209400 for the New Fork River) into Equation (3) in place of $Q_{est}$, and solving for the unknown $Q_{gain/loss}$. Figure 14 depicts box-and-whisker plots of monthly $Q_{gain/loss}$. Positive values in Figure 14 indicate that the stream gained unaccounted water during that month, likely from return flows from irrigation diversions, groundwater discharge, or local inflows from ungaged tributaries. Negative values suggest the stream is losing water during that month, likely attributable to groundwater recharge or ungaged depletion. Uncertainty in riparian ET estimates may transfer into the residual gain/loss term, particularly during spring green-up and the early irrigation season, when vegetation dynamics, shallow groundwater access, and inundation may affect surface temperature-based ET estimates.

Seasonal patterns in $Q_{gain/loss}$ are consistent with the dominant hydrologic processes in the basin. During spring and early summer, snowmelt from low- and mid-elevation areas, activation of ungaged tributaries, rising groundwater levels, and bank storage effects may contribute to positive residual gains. During the irrigation season, $Q_{gain/loss}$ reflects a more complex combination of canal diversions, consumptive use, canal seepage, field-scale recharge, delayed return flows, riparian evapotranspiration, and groundwater exchange. Because these processes may occur on different timescales, losses in one month may appear as gains in subsequent months. During late fall and winter, when irrigation diversions and riparian vegetation water use are reduced, $Q_{gain/loss}$ is more likely to reflect groundwater discharge, baseflow contributions, and any remaining limitations in the reconstructed tributary inflows. Thus, the seasonal behavior of $Q_{gain/loss}$ should be interpreted as a diagnostic of net reach-scale gain/loss behavior rather than as attribution to a single physical process.

Both the New Fork River and Green River systems are known at times to be gaining systems [56], specifically over the March–April period during the onset of melt in mountain snowpack [57], which is evident in Figure 14. Although individual components of $Q_{gain/loss}$ can be conceptually identified, their quantification requires detailed field measurements or physically based modeling frameworks. Prior research has demonstrated that even with targeted investigations, significant uncertainty remains in separating these components, particularly in gaining stream systems and irrigated basins [31,32,41,58].

Acknowledging the limitations of our calculations, we normalized the $Q_{gain/loss}$ values for each water year to the yearly average $Q_{obs}$, as pictured in Figure 15. Figure 15 is intended to provide an approximation of the bounds of yearly streamflow gains (positive values) and losses (negative values) for each watershed in our comparison, relative to the average observed discharge. The New Fork River (USGS gage 09205000) has normalized $Q_{gain/loss}$ values that range from 0.06 to 0.99, with a median of 0.36. The Green River (USGS 09209400) has normalized $Q_{gain/loss}$ values that range from −0.01 to 0.95, with a median of 0.40. The water years with higher normalized $Q_{gain/loss}$ values correspond with lower-yield water years, such as the early 2000s (see Figure 12), where $Q_{gain/loss}$ was a more significant component of the overall water balance. The results indicate there is potential for these watersheds to fluctuate yearly but that the central tendency is for the watersheds to be gaining streams. Negative $Q_{gain/loss}$ values do not represent negative streamflow; rather, they indicate months in which the residual term acts as a net loss from the modeled river reach.

Operationally, seasonal patterns in $Q_{gain/loss}$ suggest that water use demands along the mainstem Green and New Fork Rivers above Fontenelle Reservoir are generally met, as these reaches are largely unregulated and diversions from the main channels are sufficient to satisfy demand. Water exports to Fontenelle Reservoir tend to increase during the early growing season. However, interannual variability and uncertainty in the modeling approach indicate that the growing season is also the most uncertain period in this analysis. Improvements in irrigation efficiency and crop selection may influence these exports, although the framework presented here does not explicitly account for how such changes affect the overall water balance.

4.5. Transit Losses and Gains

Within the water balance (Equation (3)), $Q_{riparian}$, $Q_{water surface evap}$, and $Q_{gain/loss}$ are considered transit losses or gains along the rivers. The median monthly values of each transit component are shown in

Table 2. Median monthly flow rates for transit losses and gains, $Q_{irrigation}, Q_{tributary}$, and $C_u$(converted to m³ s⁻¹). Shading indicates the largest component of the water balance when compared to $Q_{obs}$. Both $C_u$ and $Q_{diversion}$ are only shown for the study irrigation season (April–July). The shading of certain cells indicates the largest component of the water balance.

		Median Monthly Flow Rates (m3 s−1)
	Month	Qriparian		Qwater surface evap		Qgain/loss		Qdiversion		Qtributary		Qobs	Cu,irrig
New Fork River (USGS 09205000)	1	−0.2	(−4.0%)	0.0	(0.0%)	3.6	(61.2%)	-	-	2.6	(44.7%)	5.9	-
	2	−0.2	(−3.9%)	0.0	(0.0%)	3.4	(58.2%)	-	-	2.6	(45.0%)	5.9	-
	3	−0.3	(−3.3%)	−0.2	(−2.7%)	5.4	(70.1%)	-	-	2.9	(37.0%)	7.7	-
	4	−0.6	(−5.4%)	−0.3	(−2.4%)	9.5	(89.8%)	−4.2	(−39.9%)	5.6	(52.9%)	10.5	0.3
	5	−1.3	(−4.0%)	−0.3	(−0.9%)	5.5	(17.5%)	−18.0	(−57.3%)	48.9	(155.5%)	31.4	0.9
	6	−2.6	(−3.5%)	−0.3	(−0.3%)	6.8	(8.9%)	−34.1	(−44.6%)	106.8	(139.4%)	76.6	4.2
	7	−3.2	(−11.3%)	−0.2	(−0.6%)	20.6	(72.8%)	−22.2	(−78.5%)	37.1	(131.1%)	28.3	6.2
	8	−2.6	(−21.6%)	−0.1	(−1.0%)	4.8	(40.3%)	-	-	11.8	(98.7%)	12.0	-
	9	−1.7	(−18.6%)	0.0	(0.0%)	3.6	(39.1%)	-	-	7.6	(81.2%)	9.3	-
	10	−0.8	(−8.5%)	0.0	(0.0%)	6.0	(62.2%)	-	-	4.7	(49.0%)	9.6	-
	11	−0.3	(−3.6%)	0.0	(0.0%)	5.6	(62.5%)	-	-	3.7	(40.9%)	9.0	-
	12	−0.2	(−2.7%)	0.0	(0.0%)	4.2	(61.1%)	-	-	2.7	(39.6%)	6.9	-
	Water Year	−1.2	(−6.8%)	−0.1	(−0.6%)	6.6	(38.5%)	−6.6	(−38.3%)	19.2	(111.6%)	17.2	-
Green River (USGS 09209400)	1	−0.5	(−3.8%)	0.0	(0.0%)	3.4	(28.5%)	-	-	9.0	(75.2%)	12.0	-
	2	−0.5	(−3.6%)	0.0	(0.0%)	4.2	(32.1%)	-	-	9.2	(71.3%)	12.9	-
	3	−0.6	(−2.9%)	−0.7	(−3.7%)	10.0	(50.5%)	-	-	11.4	(57.3%)	19.8	-
	4	−1.1	(−3.2%)	−0.9	(−2.5%)	19.1	(54.5%)	−10.5	(−30.0%)	28.1	(79.8%)	35.1	0.9
	5	−2.5	(−3.5%)	−1.0	(−1.4%)	16.3	(23.2%)	−46.1	(−65.7%)	108.1	(154.1%)	70.2	3.9
	6	−5.5	(−3.9%)	−0.9	(−0.6%)	48.6	(34.4%)	−85.6	(−60.7%)	189.0	(134.0%)	141.0	14.3
	7	−6.6	(−10.2%)	−0.6	(−0.9%)	50.9	(78.4%)	−55.6	(−85.5%)	77.1	(118.7%)	65.0	19.7
	8	−5.4	(−21.4%)	−0.4	(−1.6%)	5.3	(20.9%)	-	-	26.9	(106.9%)	25.2	-
	9	−3.5	(−20.5%)	0.0	(0.0%)	3.0	(17.1%)	-	-	18.6	(107.9%)	17.2	-
	10	−1.6	(−8.6%)	0.0	(0.0%)	6.2	(32.2%)	-	-	15.2	(79.2%)	19.2	-
	11	−0.7	(−3.6%)	0.0	(0.0%)	7.8	(40.7%)	-	-	12.5	(65.5%)	19.0	-
	12	−0.4	(−2.8%)	0.0	(0.0%)	4.4	(33.3%)	-	-	9.8	(73.9%)	13.3	-
	Water Year	−2.4	(−6.4%)	−0.4	(−1.0%)	14.2	(37.6%)	−16.5	(−43.6%)	41.7	(110.2%)	37.8	-

. Values in parentheses are the percentage of each water balance component compared to the median monthly observed flow rate ($Q_{obs}$) at the outlet of the New Fork River (USGS 09205000) and the Green River above Fontenelle Reservoir (USGS 09209400). $Q_{riparian}$ peaks in July but becomes a larger component of the water balance (as a percentage of flow) in August and September. $Q_{gain/loss}$ values fluctuate throughout the year (as also shown in Figure 14). Table 2 highlights that $Q_{gain/loss}$ is a major component of the water balance, specifically in March and April, when the low-elevation snowpack is melting because $Q_{gain/loss}$ captures ungaged tributaries, which are typically located in the lower elevations of the basin (see Figure 3). Between October and April, the water balance in the New Fork River is predominantly $Q_{gain/loss}$, likely from ungaged tributaries and irrigation water returning to the stream [39]. Inflows from the gaged tributaries ($Q_{tributary}$) are the largest components of the water balance between May and September for the New Fork and for every month in the Green River.

Also shown is the median monthly observed flow ($Q_{obs}$) at the outlet of the New Fork River (USGS 09205000) and the Green River above Fontenelle Reservoir (USGS 09209400). Values in parentheses are each component’s percentage relative to $Q_{obs}$. Shading indicates the largest component of the water balance when compared to $Q_{obs}$. Both $C_u$ and $Q_{diversion}$ are only shown for the study irrigation season (April–July).

4.6. Future Work

This study demonstrates a framework that integrates limited in situ observations with modern remote sensing and global hydrologic datasets to produce reasonable water balance estimates. Several promising directions remain for future investigation:

Expansion to Additional Basins: The methods developed here are directly transferable to other sections of the Green River and broader Upper Colorado River Basin, especially basins that have historical streamflow data but limited current streamflow data. Aside from available streamflow observation data, OpenET and GEOGLOWS data are available throughout the United States. However, ET-Demands data have only been developed for select basins [23], and thus the ET-Demands data will need to be calculated for other basins, or alternative approaches to calculating $C_u$ and $Q_{diversion}$ will be necessary for application of this methodology outside of the Upper Colorado River Basin.

Representation of Explicit Groundwater–Surface Water Exchange: The current water balance approach absorbs groundwater recharge and discharge processes into a residual gain/loss term ($Q_{gain/loss}$). Incorporating explicit models of groundwater–surface water interactions would provide better fidelity, helping distinguish between anthropogenic depletion and natural variability in baseflow contributions. This might be achieved by coupling with regional groundwater simulations or by integrating well monitoring information. Although many tributaries to the New Fork River and Green River were accounted for within this study, several ungaged tributaries were not included. Methods to calculate these ungaged tributaries could impact the results.

Return Flow Lag Modeling: Return flows from irrigation systems do not always return to the river within the same month they are withdrawn. Future refinements should account for these time-lagged return flows, particularly those moving through shallow groundwater pathways. Developing lag-response functions or integrating empirical delay distributions would improve monthly water balance closure and better reflect the temporal dynamics of recharge and return flow.

Integration of SABER for Automated Bias Correction: The manual bias correction process employed here could be enhanced through implementation of the SABER post-processing framework [35,36], which will be integrated into future GEOGLOWS versions. SABER provides a consistent, scalable method for correcting model biases in both gaged and ungaged river basins. This would allow the incorporation of tributaries that have no historical gaging.

Vegetation Dynamics and Ecological Water Use: Associated with $Q_{gain/loss}$ but not discussed is increased non-crop plant growth during the spring and early summer. Greening and maximum greening [59] occur when transpiration rates of natural plants are higher due to increased growth rates and maximized leaf areas, which would effectively decrease $Q_{gain/loss}$. Driven by CO₂ fertilization, there have been greening trends in global semi-arid regions and drylands, like the GRB, that are generally leading to increased water use by plants and, in some cases, decreased streamflow [60,61].

Daily Scale Enhancements: Because ET-Demands data is only available at a monthly time step, the framework described in this analysis was restricted to a monthly time step. New approaches to CONUS-level $C_u$ estimates have been developed that operate on a daily time step for hydrologic unit code 12 (HUC12) spatial scales [62]. Future utilization of these datasets within the framework presented here could provide enhanced ability to investigate short-term variability, caused by components of the hydrologic system such as return flows, and the effects of extreme events on the water budget.

5. Conclusions

This study presents a novel framework for developing monthly water balances in data-limited basins by integrating historical observations and modern hydrologic modeling and datasets, including bias-corrected GEOGLOWS streamflow estimates, OpenET-based riparian evapotranspiration, and ET-Demands-derived consumptive use. Focusing on the Green River Basin (GRB) above Fontenelle Reservoir in Wyoming, we demonstrate that even in the absence of complete and continuous in situ observations, skillful hydrologic accounting is achievable over multi-decade timescales. Highlights include the following:

• By applying flow duration curve-based bias correction to GEOGLOWS simulations, we improved the GEOGLOWS reconstructed streamflow time series. The composite error of these improved data at tributary locations above our comparison stream gages appeared to have been reduced substantially. These enhanced inflow estimates improved the spatial and temporal breadth of our river-scale water balances for both the Green and New Fork Rivers.
• Leveraging the ET-Demands dataset, we calculated monthly consumptive use ($C_u$) data, which was then compared favorably with data from USBR [54]. The $C_u$ data combined with an irrigation withdraw ratio developed from gaged canals was used to calculate diversion estimates ($Q_{diversion}$ for the entire basins, underscoring the potential of remote sensing products to replace or augment sparse ground-based records.
• The residual gain/loss term ($Q_{gain/loss}$) provides insight into the underlying hydrologic processes not explicitly modeled, such as groundwater interactions, local runoff, and lagged return flows. Differences between our estimated monthly water balance and streamflow at our two downstream gage locations was greatest during the April–July time frame. This roughly aligned with an active hydrologic period when snowmelt, anthropogenic influences, and the growth of vegetation are greatest.
• Seasonal patterns suggest substantial subsurface contributions to streamflow during dry months and highlight challenges associated with irrigation withdrawals and return flow dynamics.
• This study highlights a framework by which to link data, models, and methods to develop the major components of a monthly water balance in data-limited basins. Such basins are prevalent throughout the Western United States. Future efforts can better evaluate the different components of the water balance, specifically attributing the magnitude of $Q_{gain/loss}$ to irrigation surface water return, irrigation subsurface return, and local inflows.