Main

The Colorado River is a major water source and economic engine for the southwestern United States. Nearly 90% of its water originates from snow-dominated watersheds in Colorado, Utah and Wyoming1 to support 40 million people and US$1.4 trillion in economic activity2. The Colorado River is currently affected by a multi-decadal drought3,4 with legal and political tensions rising to develop voluntary cuts in water use as a result of diminishing supply5. However, declining streamflow and reservoir storage cannot be explained solely by lower precipitation in headwater basins. In 2021, the Upper Colorado River basin received 80% normal snowpack but delivered only 30% of average streamflow to its receiving reservoir6. Warming in mountain basins may be part of the explanation7,8. Warming reduces snowpack accumulation by altering energy exchanges between the snow surface and atmosphere to increase vapour loss from the snowpack (that is, sublimation) in lower-humidity regions9. Warming also shifts more precipitation towards rain and enhances mid-season melt events to accelerate snow–albedo feedbacks10 and reduce the snow-covered area. Changes in snow dynamics potentially alter hydrograph timing and reduce streamflow11,12,13. However, these responses are not ubiquitous. Uncertainty is introduced because multiple, and often competing, mechanisms affect how important seasonal snow dynamics are to streamflow generation within a given basin14. For example, the timing and intensity of water inputs (rain and snowmelt) need to consider their synchronicity with energy availability and subsurface storage properties.

In this Article, we conceptualized streamflow source as containing three water components. Surface runoff (SR) is water flowing over the land surface outside of stream channels. It occurs as either infiltration or saturation excess. Interflow is the subsurface lateral flow through the soil zone. The soil zone represents shallow sediment affected by macropores from both root and animal activity, as well as near-surface weathered bedrock. The soil zone is generally less than 1 m in depth, allows both matrix and preferential flow and is fully coupled to the deeper subsurface flow domain. Interflow can support down gradient changes in soil moisture, evapotranspiration, groundwater generation (that is, recharge) and streamflow. Groundwater flow is defined as water movement through the underlying alluvial and bedrock units. Both unsaturated and saturated conditions are simulated. Changes in groundwater storage account for recharge, groundwater evapotranspiration, groundwater discharge into the soil zone and interaction of groundwater (gains and losses) with the stream network. The prevailing assumption has been that mountain systems have limited groundwater storage and streamflow is primarily reliant on land surface processes that control SR and interflow15,16. However, research also indicates that groundwater contributes 10–50% of the annual streamflow in mountain catchments of the Colorado River17,18,19, with growing evidence that groundwater flow to streams can occur at a considerably greater depth than previously believed20, even in fractured crystalline21 and metamorphic rock22.

Although the release of stored groundwater has the potential to buffer streamflow against short-term climate extremes13,23,24, how groundwater storage will respond to long-term warming and how this may affect streamflow generation in mountain basins is unclear. Understanding groundwater flow to mountain streams is made difficult by the need to quantify the interaction of climate and snow dynamics25,26,27,28, soil water storage and loss to vegetation29, and their combined effects on groundwater recharge30. Mountain basins are also comprised of complex lithologic and structural features that dictate highly heterogeneous distributions of permeability, drainable storage, water table elevations and groundwater flow depths. Modelling all of these processes in mountain basins at the appropriate resolution (100–250 m; ref. 31) is numerically expensive, while obtaining sufficient data to constrain integrated hydrologic models is difficult32,33 and exacerbated by challenges in site access and harsh climate34.

To address these challenges, we focused on the East River in Colorado (750 km2; Fig. 1a). The East River is representative of many headwater basins of the Colorado River35. The climate is continental subarctic, and it spans nearly 2 km in topographic relief (2,440–4,300 m). The East River contains an extensive atmosphere-to-bedrock observational network with investment across multiple US federal agencies, national laboratories, university partners and regional stakeholder groups36. The data collected at the site include airborne mapping of snowpack37,38, vegetation39 and bedrock resistivity40, as well as dense arrays that monitor snowpack, groundwater and streamflow41. These data were used to parameterize and validate an integrated hydrologic model42 at a 100 m grid resolution to represent topographic complexity. A daily timestep was used to limit temporal smoothing of energy and water partitioning among the snowpack, vegetation, soil zone and geologic units.

Fig. 1: Simulating groundwater at high spatial resolution and to depth in the headwaters of the Colorado River.
figure 1

a, The East River (black star, 750 km2) in the Colorado River basin (blue shading) in the southwest United States. The Upper Colorado River basin is shown in striped shading. b, The groundwater model represents Tertiary igneous intrusive (Tmi); Permian, Cretaceous and Tertiary sandstones (PPm, Kd and Tw, respectively); Cretaceous mudstone (Kmv); Cretaceous shale (Km) and Quaternary alluvium (Qa). A description of the geologic units is provided in Supplementary Table 2. c, The simulated depth to the water table at the 100 m resolution for the model start date of 1 October 1986. Depths obtained from the spin-up simulation described in Methods. Open water is indicative of groundwater-fed, perennial streamflow and inundated floodplains. Simulated perennial streams align with the United States Geological Survey (USGS) designated network indicated with black lines.

Full size image

Details on the land surface model construction and calibration are presented in previous work43. Herein, the land surface model parameterization was largely maintained for the integrated hydrological model. The integrated hydrologic model includes a three-dimensional groundwater flow model44 to simulate variably saturated groundwater flow as well as streamflow gaining and losing conditions45. The groundwater model accounts for eight stratigraphic units46 and extends 400 m below the land surface (Fig. 1b). Inclusion of the deeper subsurface was based on gas tracer observations and hydrologic modelling of a nested subcatchment within the East River that found 20% of groundwater flow occurs in excess of 300 m below the land surface21. Figure 1c shows the simulated depth to the water table at the start of the historical simulation based on a fully spun-up model accounting for spatially and temporally variable climate forcings. Simulated open water indicates groundwater-fed perennial streams and inundated floodplains. An overview of model construction is provided in Methods. Supplementary Information provides details on data sources (Supplementary Table 1 and Supplementary Fig. 1), updates to the land surface model with dynamic coupling to the groundwater model (Supplementary Figs. 2–15) and groundwater model construction (Supplementary Table 2 and Supplementary Figs. 16–20).

We assumed +4 °C warming to represent end-of-century conditions based on local temperature trends (Supplementary Table 3). We applied a constant +4 °C temperature increase continuously across all water years, and then we applied +4 °C independently to each season (autumn, winter, spring and summer) based on historical climate sequencing for water years 1987–2022 (water year: 1 October to 30 September). The approach is meant as a sensitivity analysis that allows a direct comparison to the historical baseline condition. The approach does not account for incremental warming, seasonal differences in warming, non-uniform spatial response to temperature47, the influence of the Clausius–Clapeyron relationship on precipitation or changes in extreme event sequencing48. We did not simulate possible changes in vegetation water use response to elevated CO2 (ref. 49) or the consequences of forest mortality50,51. Using this experimental design, we explored historical water budgets and groundwater exchanged with streams. We then analysed how sensitive projected streamflow generation is to a warmer climate and how and where changes in groundwater storage may influence changes in streamflow.

Groundwater is a stable and important stream water source

The basin-scale water budget averaged over the historical simulation period is shown in Fig. 2. The spatial distribution of climate across the East River follows previous work52 that used airborne light-detection and ranging (LiDAR) of snowpack37. LiDAR-adjusted precipitation inputs implicitly redistributes snow by wind and avalanche to move snow off ridges and into high-elevation cirque valleys to produce the deepest snowpack near the tree line. Modelled historical conditions indicate that basin average annual precipitation is high (947 ± 176 mm, mean ± s.d.) and arrives primarily as snow. The summer monsoon delivers approximately 20 ± 6% of the annual water input and this is important for sustaining vegetation water demand in the late summer53. Annually, the East River is generally energy limited such that aridity, or the ratio of potential evapotranspiration to precipitation (PET/P), is less than 1.0. Despite being a relatively wet and cold climate, the average annual simulated evapotranspiration accounts for slightly more than half of the water exported out of the basin, with the majority of evapotranspiration derived from the soil zone and smaller contributions obtained from sublimation, canopy evaporation and groundwater use by phreatophytes. Simulated evapotranspiration aligns with observed data54 (Supplementary Fig. 12) and, excluding groundwater evapotranspiration, fits the Budyko framework55, such that the fraction of evapotranspiration to precipitation (ET/P) increases at a decreasing rate with aridity (Fig. 2b). Although evapotranspiration sourced from groundwater is a fairly small portion of the average annual water budget at the basin scale (2 ± 1%), it is important locally along lower elevation floodplain reaches where water table elevations are shallow (Supplementary Fig. 13). In years when basin-scale aridity is high, groundwater evapotranspiration approaches 8% of the total precipitation input.

Fig. 2: Historical basin-scale water budget for water years 1987–2022.
figure 2

a, A schematic representation of the average annual water budget. Water fluxes are provided as the percentage of average annual precipitation with standard deviations. Boxed water budget terms are stream sources. b, The ratio of annual evapotranspiration to precipitation (ET/P) as a function of aridity, or the ratio of PET/P with the classic Budyko formulation55 provided for comparison. Subl., sublimation; canopy, canopy evaporation; soil, soil evapotranspiration; Groundwater ET, groundwater evapotranspiration; Total ET, sum of all these terms. c, The annual streamflow source defined as interflow + surface runoff (SR) and groundwater flow as a function of aridity. Interflow volumes in streams are corrected for seepage recharge. The correlation coefficient is provided (ρ) with statistical significance indicated (P value) for a one-tailed ordinary least squares regression model. Interflow + SR data are linearized using a log–log transformation. Our results indicate that (1) on average, 53% of precipitation is lost to the atmosphere as total evapotranspiration despite being a wet and cold basin, (2) groundwater generation is the combination of areal recharge below the rooting zone and seepage from non-perennial streams in topographic convergent zones, (3) interflow supports down gradient soil moisture, recharge, evapotranspiration and streamflow, (4) interflow to streams is highly variable as a function of interannual climate and (5) groundwater is an important and relatively stable source of streamflow.

Full size image

Groundwater recharge was simulated with two separate mechanisms. Areal recharge is the vertical flux of water below the soil zone. Areal recharge is relatively low in the East River (4 ± 1%) because of its steep topography and low permeability bedrock that biases excess soil water towards lateral interflow. Seepage recharge occurs when stream water flows into the subsurface. This primarily occurs from non-perennial river channels that are disconnected from the saturated groundwater system (also referred to as gradient-driven flow). Seepage was estimated to be the dominant recharge mechanism in the East River (10 ± 1%). It largely occurs in the spring when interflow moves high-elevation snowmelt down-gradient towards topographic convergent zones to form low-order streams (Fig. 3). These streams are considered net losing. In contrast, higher-order streams occur along valley bottoms. They aggregate streamflow from upland portions of the East River and tend towards perennial, or groundwater-gaining, conditions. Both areal recharge and seepage recharge decline when annual conditions are warm and dry (Supplementary Fig. 21).

Fig. 3: Groundwater interactions with streamflow vary across space and time in mountain basins.
figure 3

a, The simulated average annual streamflow in the East River for average climate conditions to illustrate the downstream aggregation of stream water through the river network. b, The spatial distribution of net groundwater gaining and net groundwater losing streams. c, A daily time series of net groundwater into or out of stream channels aggregated across the East River. Groundwater losses in low-order tributaries of the East River exceed groundwater gains in higher-order streams during snowmelt.

Full size image

Although a portion of interflow-derived streamflow seeps into the subsurface to support groundwater recharge along low-order tributaries, interflow still represents the bulk of the stream water source exiting the East River (70 ± 29%). Interflow in streams is highly responsive to interannual variability in aridity and when combined with surface runoff, plots non-linearly as an inverse function of evapotranspiration to precipitation (ET/P) (Fig. 2c), which suggests a tight coupling with vegetation water use. For example, under warm and dry conditions in the East River (PET/P >1), more than 80% of the precipitation input is consumed by evapotranspiration and interflow decreases. Net groundwater discharge to streams is estimated at approximately 12 ± 2% the average annual water budget and volumetrically represents 26 ± 3% of annual streamflow, although extreme wet and dry water years stretch the fraction of groundwater contributing to streamflow between 15% and 43%, respectively, in response to interflow contributions. In contrast to interflow, groundwater volumetric inputs to streams are a stable source of streamflow that are relatively insensitive to interannual climate variability.

Forest water use drives groundwater storage declines

Groundwater storage aggregates climate over a 4 year period in the East River (Methods). Historically, this multi-year aggregation has helped to buffer extremely dry water years and promote groundwater recovery back to the historical mean condition (Fig. 4a). The onset of low-to-no monsoon rain from 2016 to 2020 combined with the shock of an extremely dry water year in 2018 has prompted sustained declines in simulated groundwater storage, with average annual groundwater storage losses simulated at greater than two standard deviations of the historical mean by the close of the historical simulation. This agrees with groundwater observations in the East River that indicate hillslope water table declines are not recovering from the extremely dry conditions of 2018 despite above average snow accumulation in 2019 and 2022 (ref. 56) (Supplementary Fig. 22). Groundwater observations and model results highlight that groundwater recovery time increases with higher drought severity57 and that compounding climate events, such as increased snow-drought58 coupled with decreased monsoon storm frequency59, will further hinder recovery.

Fig. 4: Substantial groundwater storage declines occur with warming and are most prominent in the subalpine forest primarily due to summer temperature increases.
figure 4

a, Simulated East River average annual groundwater (GW) storage anomalies for the historical condition and +4 °C warming scenarios. Anomaly is (μ − x)/σ, where μ is the historical average change in annual storage, x is the annual change in storage and σ is the historical standard deviation of change in annual storage. b, Simulated change in water table elevation at the 100 m grid resolution between the warming scenario and historical condition at the end of the simulation (water year 2022), <0 indicates the water table goes down with warming. c, The average water table change (∆) across elevation based on seasonal warming and warming all year. Ecological zones are identified. Maximum conifer density occurs at the transition between the upper and lower subalpine. All photos taken by the lead author, R. Carroll.

Full size image

Applying spatially uniform +4 °C warming to historical climate sequencing results in a groundwater system that is unable to recover to the historical average condition even during the very wet periods simulated. The largest drops in groundwater storage occur after each exceptionally dry water year and losses exceed the historical low point after the first of these events. At the end of the hypothetical warming simulation, groundwater storage loss exceeds six standard deviations of the historical mean and does not appear to stabilize. Water table declines in response to multi-decadal warming are non-uniform across the East River, with mountain ridges and valley bottoms appearing less affected than hillslopes (Fig. 4b). Averaged water table declines across elevation suggests that groundwater losses are greatest near the transition between the upper and lower subalpine zones (Fig. 4c). This shift in ecological zones is distinguished by the transition between energy and water limitation during historically low-precipitation water years and coincides with the maximum density of conifer cover43. Isolating seasonal warming effects on water table declines indicates winter warming raises water table elevations, whereas summer warming is responsible for the largest water table declines.

To understand groundwater storage loss, we need to consider how the timing of warming affects recharge processes (Fig. 5). Changes in recharge are largely controlled by changes in soil moisture, with low snow years experiencing the largest recharge declines (Supplementary Fig. 23). Winter warming occurs when atmospheric demand is low and the vegetation is not transpiring. Winter warming produces early snowmelt that increases soil moisture and interflow to streams and subsequent seepage recharge into the subsurface. Increased winter recharge comes at the expense of spring recharge, but the net effect of winter warming is an increase in annual groundwater generation and increased water table elevations in the East River. Spring warming and, to a lesser degree autumn warming, also promote earlier snowmelt and increase early season recharge compared with historical conditions, and do so at the loss of later-season recharge. Unlike winter warming, autumn and spring warming typically result in a net decrease in annual recharge. Exceptions occur during wetter and cooler water years (for example, 1995 and 2011). Summer warming occurs when potential evapotranspiration is high, with the largest increases in evapotranspiration occurring in the subalpine zone where conifer density is highest (Supplementary Fig. 24). Summer warming results in relatively large reductions in soil moisture throughout the entire water year but, unlike warming in the other seasons, interannual variability in recharge related to summer warming is not sensitive to reductions in soil moisture. Instead, changes in evapotranspiration are a better predictor of recharge decline.

Fig. 5: The timing of warming affects simulated changes in recharge.
figure 5

Simulated changes in recharge associated with warming in the autumn, winter or spring are best described by changes in soil moisture. Summer warming is responsible for the largest annual reductions in recharge, with interannual variability best described by increased evapotranspiration. A change (Δ) <0 indicates a reduction compared with the historical simulation. Warming scenarios: +4 °C applied all year, limited to the autumn (September to November), winter (December to February), spring (March to May) or summer (June to August). a,b, The changes from the historical simulation for water years 2016–2021 for different warming scenarios: daily soil moisture (a) and recharge (b). c, Annually aggregated changes in recharge for different warming scenarios for water years 1987–2022. d, The annual change in recharge related to annual change in soil moisture for different warming scenarios. e, The changes in annual recharge as a function of changes in annual evapotranspiration due to summer warming. The first year of the summer warming simulation produces a smaller than expected reduction in recharge due to stored soil water at the start of the simulation that moderates response. This dissipates after the first year. The shaded area represents the 99th confidence interval. ρ is the correlation coefficient and the P value indicates statistical significance for a one-tailed ordinary least squares regression model.

Full size image

A multiple linear regression analysis was applied to simulated average water table declines in ten nested East River catchments. The regression results indicate that high-relief catchments covered by conifer forests are expected to experience the largest declines in groundwater storage in response to long-term warming (Methods). High-relief catchments occur in the uplands, while lower-relief catchments are more representative of valley reaches where groundwater tends to pool and are less sensitive to change (Fig. 4b). Regression results imply that conifer forest water use is an important control on simulated water table declines. Forests have higher rates of evapotranspiration compared to non-woody vegetation60 and potentially out-compete for access to deeper-water pools61, whereas dense conifer canopies are responsible for large interception losses to evaporation that are expected to increase in a warmer climate. As discussed above, increased evapotranspiration reduces simulated recharge and results in lower water tables within the subalpine62.

Applying the regression model across the mountainous regions of the Upper Colorado River basin helps distinguish those catchments potentially most at risk of groundwater decline in a warmer future (Fig. 6). Specifically, mountainous areas in the Colorado Headwater and the Gunnison basins are estimated to be at elevated risk. These two basins contribute nearly 40% of the annual Upper Colorado streamflow entering the receiving reservoir, Lake Powell, and are critical to managing water resources across the entire Colorado River basin. Our regression results can be used to identify individual catchments for a more detailed study of how forests could be conjunctively managed to improve groundwater sustainability63. Detailed analysis would require expanding our approach to include a more robust assessment of forest recruitment and growth strategies, stress effects on water use64 and possible forest succession as a result of different climate disturbances51,65, as well as addressing the tradeoffs of forest thinning on carbon storage and biodiversity.

Fig. 6: Groundwater loss due to long-term warming is non-uniformly distributed across the Upper Colorado River basin.
figure 6

Results indicate the Colorado Headwaters (2) and Gunnison (3) hydrologic unit code level-4 basins are most at risk of high or extreme water table declines. The location of the East River in the Gunnison Basin is outlined in black. The location of the Upper Colorado River basin with respect to the western United States is provided in Fig. 1a.

Full size image

Groundwater storage declines amplify streamflow loss

Groundwater flow into streams is not strongly correlated to interannual aridity (Fig. 2c) but does respond as a direct exponential function of groundwater storage (ρ = 0.85 and P = 0.00) (Supplementary Fig. 25). As discussed in the previous section, seepage recharge increases with warming in the winter and early spring, and this will occur at the expense of streamflow exported out of the basin. To isolate the effects of groundwater exchanges on streamflow generation, we applied +4 °C warming to individual water years with the groundwater condition initialized to the East River’s average historical state and then reran the model with groundwater initialized at a depleted state. The model results indicate that accounting for stream water and groundwater exchanges related to groundwater storage decline is important (Fig. 7). Without groundwater depletion, warming produces only modest declines in average annual streamflow (6–13%). The effects on annual streamflow are lessened because sharp decreases in summer streamflow are largely offset by increased winter and early spring streamflow. Accounting for depleted groundwater storage with warming nearly doubles average annual streamflow loss (15–34%). Streamflow declines largely occur because of increased seepage recharge in the winter, with the fraction of streamflow decline increasing at higher aridity.

Fig. 7: Long-term warming and associated water table declines will affect seasonal streamflow with significantly reduced annual average and minimum flows.
figure 7

a, The effects of increased temperature (∆T) and reduced groundwater storage (∆S) on a low-precipitation water year hydrograph with the total effect of both (Total) and comparisons made to historical conditions provided. Letters along the x axis denote months (that is, O for October, N for November and so on). b,c, The change from the historical condition for annual streamflow (b) and the 7 day minimum streamflow (c) across an aridity gradient (PET/P). The shading represents the 95% confidence interval with central lines representing a one-tailed ordinary least squares regression model and aridity log-transformed. The East River becomes non-perennial in the late summer (fraction of flow reduction approaches 1) when precipitation (P) is ≤700 mm.

Full size image

Assuming only temperature increases and no groundwater storage depletion results in a 7 day minimum streamflow that is dramatically reduced from the historical climate by 21–81%, with the largest decreases occurring in water years when aridity is high. Preferential partitioning of soil water to support forest water use is the primary driver of these severe streamflow declines, although simulated seasonal riparian groundwater use is also important. Accounting for diminished groundwater storage on summer flows is estimated to reduce the 7 day minimum streamflow by an additional 12–27%. Streamflow declines related to groundwater storage loss are a consequence of decreased groundwater inflows to streams through lowered water table elevations. Although groundwater volumetric inflow reductions are relatively small, they push the East River towards a dry riverbed during below-average precipitation (<700 mm yr−1), which is not predicted if the groundwater system is ignored. We conclude that simulating groundwater storage and allowing complex groundwater exchanges with mountain stream water is important (Fig. 8). Disregarding these interactions are expected to underestimate streamflow response to long-term warming.

Fig. 8: The seasonal effects of increasing temperature on streamflow declines include complex exchanges between surface and groundwater that need to be explicitly modelled to avoid underestimating streamflow losses given a warmer future climate.
figure 8

Winter: warming will decrease snow coverage and promote the earlier onset of infiltration and runoff, although the net effect of annual warming on recharge will be less than under historical conditions. An increased winter stream stage coupled with lower water table elevations will drive increased seepage loss (that is, gradient-driven loss) in non-perennial streams to significantly lower annual streamflow. Summer: higher temperatures will promote increased evapotranspiration in the forested areas and riparian zones to limit the lateral movement of subsurface flow into stream channels. A decrease in groundwater entering streams will occur through falling water tables. Summer streamflow will be lower with a possible transition towards non-perennial conditions when evapotranspiration losses exceed groundwater inflows to streams. Qs = interflow + surface runoff to streams; Qg = groundwater flow to streams.

Full size image

Summary

Mountain snowpacks are a critical source for downstream water use worldwide66. Dwindling stream water supplies are aggravating increased lowland demand such as the Colorado River basin, promoting political tensions and a need to better understand the mechanisms of mountain discharge for better adaptive management. Hydrologic modelling studies have identified increased evapotranspiration as a primary mechanism of streamflow reduction in a warmer climate, largely through decreased albedo with snow loss12 and longer growing seasons, but with spatial variability in evapotranspiration response largely dictated by elevation67. Although groundwater contributions are increasingly acknowledged as important to mountain streamflow, limited analysis has been done at the resolution necessary for estimating recharge processes in complex terrain and linking these contributions to land surface processes associated with snow and evapotranspiration dynamics. We use a high-resolution, integrated hydrological model that extends 400 m into the subsurface to determine whether groundwater will buffer or amplify streamflow loss in a warmer world. Our results for a representative headwater basin in the Colorado River indicate groundwater storage aggregates climate over a 4 year period to minimize climate extremes, that groundwater contributions have historically been an important and stable contribution to stream water (26 ± 3%) and are directly responsive to groundwater storage. Groundwater recharge is estimated to occur through the percolation of infiltrated water below the soil zone (areal recharge) and via streamflow loss into the groundwater system (seepage recharge). Seepage recharge occurs primarily in the spring when low-order, non-perennial streams form from snowmelt transported via interflow into topographic convergent zones.

We applied constant warming to multiple decades of historical climate. Estimated warming is consistent with end-of-century conditions based on local temperature trends. Under these conditions, groundwater storage falls below the historical minimum condition after the first extremely dry water year and is unable to recover to historical conditions even after simulating multiple wet periods. Warming increases recharge in the winter and early spring but does so at the expense of recharge later in the year to promote a net loss in groundwater generation. Elevated evapotranspiration from summer warming results in small reductions in daily recharge, particularly in the subalpine forests where conifer densities are highest, but it is the most influential process on recharge reduction and associated groundwater storage loss when aggregated annually.

Warming promotes increased seepage recharge from streams into the groundwater system during the winter and early spring to partially buffer recharge loss, but in doing so, nearly doubles annual streamflow loss compared with warming simulations if groundwater storage declines are ignored. Diminished groundwater storage also reduces groundwater inflows to perennial streams and forces the minimum summer flow towards intermittent conditions during dry water years. Water table declines are non-uniformly distributed across the East River, with upland areas occupied by conifer forests simulated at particular risk. Our research reinforces the need to better understand seasonal forest water use strategies as they adapt, acclimate and possibly migrate in response to increased temperatures to quantify recharge dynamics. We also stress that including detailed spatial and temporal groundwater exchange with surface water is critical to forecasting mountain streamflow changes in a warmer future. Extending our results across the region helps identify those mountainous catchments in the Upper Colorado River basin that may benefit from the joint management of forest and groundwater resources to minimize streamflow reductions given a warmer future climate.

Methods

Model development

Hydrologic water budgets rely on the United States Geological Survey (USGS) groundwater and surface water flow model (GSFLOW)42. GSFLOW dynamically couples the USGS precipitation–runoff modelling system68 (PRMS) and the Newton formulation of the USGS three-dimensional modular groundwater flow model44 (MODFLOW). PRMS parameterization of the East River is described in previous work69 but provided here for completeness. The data collection sites for model development are given in Supplementary Fig. 1 with links to data provided in Supplementary Table 1. USGS tools for GSFLOW input70 helped parameterize elevation71 and vegetation characteristics39,72 of dominant cover type, summer and winter cover density, canopy interception characteristics for snow and rain, and transmission coefficients for shortwave solar radiation using vegetation classification maps72 (Supplementary Fig. 2). Climate data for water years 1987–2022 were distributed using the daily lapse rate of minimum and maximum temperature values from two snow telemetry stations (SNOTEL), Schofield site 737 and Butte site 380, and adjusted for aspect. Temperatures were bias corrected for sensor replacement73 in the early 2000s. Shortwave solar radiation uses a degree–day method applicable for the Rocky Mountains74 and calibrated to weather stations operated by the Rocky Mountain Biological Laboratory (RMBL) (Supplementary Fig. 3). Potential evapotranspiration was calculated using a modified version of the Jensen–Haise formation that is dependent on temperature and solar radiation75. Simulated evapotranspiration was compared with a local eddy covariance flux station54 (Supplementary Fig. 12). Observed daily precipitation at the Schofield SNOTEL station was spatially distributed similar to previous work52. Precipitation was defined as either rain or snow based on a temperature threshold calibrated to match snow water equivalent (SWE) at the Butte SNOTEL station. The distribution of snowfall relied on airborne LiDAR maps of snow depth37 converted to SWE based on ground surveys76 and snow density modelling77. The snow module parameterization used flight data for a wet water year near peak accumulation (7 April 2019; Supplementary Fig. 7) and was adjusted for simulated losses associated with canopy interception, sublimation and early melt before the flight. Simulated SWE was validated using flight data collected during an average water year (4 April 2016; Supplementary Fig. 4), two flights during a dry water year (30 March 2018 and 24 May 2018; Supplementary Figs. 5 and 6), snow pits dug during 2016–2020 (Supplementary Fig. 8) and snow depth at RMBL weather stations (Supplementary Fig. 9). Rainfall was distributed using the 30 year monthly parameter–elevation regressions on independent slopes model78.

Maximum soil water storage was estimated as the product of rooting depth72 and available water content as determined from soil type79 and adjusted to match PRMS-derived streamflow across the basin. Additional PRMS parameter adjustment was done to match the farthest downstream gage (east at Almont, USGS 09112500) when fully coupled to MODFLOW. The Almont stream discharge Nash Sutcliffe efficiency (NSE) of 0.81 and relative root mean squared error (rrmse) of 5% (log discharge NSE of 0.81 and rrmse of 9%). Hydrographs for 15 stream gages are provided in Supplementary Figs. 14 and 15. Smaller catchments at higher elevation were simulated with rrmse ranging from 6% to 15% but tend towards low NSE values (<0.50). Error in estimating low-order daily streamflow is partly due to difficulty developing rating curves at these sites where shifting channel morphology and sensor malfunction or loss tended to occur with runoff and required rebuilding rating curves annually. In addition, small and flashy catchments are difficult to model at the scale of the East River (hydrographic unit code (HUC)10). USGS stations are simulated with rrmse less than 6% and NSE values tend towards values in excess of 0.8. Across all discharge stations, annual streamflow is simulated with an 8% error.

A three-dimensional geologic model was developed for incorporation into the hydrological model to simulate subsurface processes over space and time. The geologic model was developed using Seequent’s three-dimensional Leapfrog geological modelling software, which is based on implicit modelling using a radial basis function to generate surface discontinuities (for example, horizons and faults) within the domain of interest80,81,82. The initial structural model was based on a series of USGS maps for the area at different scales (1:24k and 1:250k) with higher resolution maps containing complex structures with strike and dip, and fault throw information (Supplementary Fig. 16). Maps of different geologic units were georeferenced and incorporated into Leapfrog and overlain on the digital elevation model. Structural discs were assigned to the boundaries where strike and dip information was observed on the geologic maps and fit to the published cross-sections. The horizon boundaries were digitized with polylines that captured the geologic structure. The structural model was then generated that honoured the input data using the radial basis function.

Top and bottom information for each of the geologic units was extracted from Leapfrog as text files and processed for use in the hydrologic model. Specifically, eight geologic units are included in the groundwater model and divided into six layers using a grid-overlay option (Supplementary Fig. 17). The grid-overlay option was previously used for a subcatchment analysis in the East River52 and chosen on the basis of its ease of defining discontinuous geologic units, draping surface alluvium and designating a shallow weathered zone. Geologic units were parameterized using a zonal approach. Layer thicknesses were defined as 10, 20, 40, 80, 120 and 130 m for a total thickness of 400 m. Layers were limited to keep the number of active groundwater cells below 500,000 for numerical tractability. Surface alluvial spatial extent was obtained from 1:24K geologic maps. For regions in the model with information limited to the 1:150K geologic map and no mapped surface alluvium, these surface units were inferred to occur along stream reaches. The decay of hydraulic conductivity (K) with depth (Y) is defined using an exponential function \(K={K}_{0}{e}^{(-{A}_{0}Y)}\), with surface hydraulic conductivity (K0) and a decay coefficient (A0) calibrated to match observed groundwater water levels (Supplementary Fig. 18 and Supplementary Table 2). K0 values align with previous catchment-scale groundwater models in the basin21 and A0 = 0.05 is an acceptable value83 that effectively maintains groundwater levels and perennial streams at high elevations while preventing basin flooding (Supplementary Figs. 19 and 20). The simulated perennial and non-perennial conditions match USGS classifications (Figs. 1b and 3b and Supplementary Fig. 26). Water levels were found insensitive to geologic subzonal heterogeneity based on the statistical distribution of resistivity derived from an airborne electrical magnetic survey. The transient spin-up was initiated with steady-state water levels and run for 72 years (the historical climate period simulated twice) to stabilize unsaturated and saturated subsurface storage to initiate the historical transient run (Fig. 1c). Transient model simulations are organized in sequential 1 year increments with restart84 files that are executed through a batch file. This approach allows model assessment and output management during execution and provides starting conditions for subsequent experiments assessing initial storage on changes in streamflow.

Observed climate trends and simulated warming

Annual and seasonal Schofield SNOTEL climate trend analysis applied linear least squares to assess historical seasonal change in the East River (Supplementary Table 3). Seasons are defined as autumn: September, October and November; winter: December, January and February; spring: March, April and May; and summer: June, July and August. Mean, minimum and daily temperature before May 2003 applied bias correction73 to account for senor replacement. The results indicate statistical significance in temperature increase across all seasons except spring, and maximum daily temperature is increasing faster than minimum daily temperature. Annually, the average daily temperature is increasing 0.44 °C per decade. Seasonally, autumn is experiencing the largest temperature increases at nearly double the annual rate. There has been no significant change in annual precipitation in the area, although a decrease in autumn precipitation of nearly 10% per decade (P < 0.05) has occurred. Annual precipitation appears to follow a 12 year trend based on \(P=A\sin \left({wx}+C(\frac{\pi }{180})\right)+K\), where A is the amplitude set to the historical annual standard deviation (175.8 mm yr−1), C is the lag in years, K is the SNOTEL historical mean (947.3 mm yr−1) and w is the angular frequency. The lag and angular frequency were adjusted to minimize the sum squared error. Differences in means for each of the three 12 year cycles were tested using the independence of means t-test, and while the last cycle has lower annual and seasonal precipitation means, there are no statistical differences between cycles.

Water budget analysis

Initial subsurface storage was set to the spin-up conditions obtained for 1 October 1987 to allow comparisons between different model runs. Annual water budgets averaged over the water year were done across all active model cells. Regressions are based on a least squares approach with the 99% or 95% confidence intervals provided where indicated. Low monsoon conditions were defined as an annual rain anomaly (z-score) less than −0.5, which equates to annual rain less than 162 mm. Low monsoon rain is to have occurred in water years 1989, 1996, 2002, 2004, 2008, 2011–2012 and 2016–2017. A no-monsoon condition was defined as an annual rain anomaly less than −1.0, or rain less than 133 mm and simulated to have occurred in 1988, 1994 and 2018–2020. Budyko55 estimates were calculated as:

$$Y={\left(1-\exp \left(-X\tanh \left(\frac{1}{X}\right)\right)\right)}^{0.5},$$

where Y = PET/P and X = ET/P.

Annual deviation of groundwater storage from the historical average is assessed as a function of climate lag with a multiple linear regression. Climate variability is represented by annual aridity (PET/P). Aridity information for a given lag is kept in the multiple linear regression if the statistical significance of all independent variables is maintained (P < 0.01) and the coefficient of determination improves with added information. Our results indicate that aridity with lag 0 (same year) has no predictive ability on simulated groundwater storage deviation (r2 = 0.05 and P = 0.19). Instead, groundwater storage deviation is best explained using 4 years of climate information that excludes lag 0 (r2 = 0.85) with groundwater storage of −3.1 – 66.5(X4) − 86.1(X3) − 106.6(X2) − 124.3(X1); where X4 is aridity lag 4 years (P = 0.001), X3 is aridity lag 3 years (P = 7.5 × 10−5), X2 is aridity lag 2 years (P = 2.7 × 10−6) and X1 is aridity lag 1 year (P = 1.0 × 10−7). Daily cell-by-cell output for various water budget components are generated by PRMS in CSV format and by MODFLOW in binary format. One-dimensional elevational plots aggregate simulated output across 50 m elevational increments. Aggregated values for the warming scenarios are then subtracted from the historical, or baseline, scenario (for example, Fig. 4c and Supplementary Fig. 24). Subbasins are defined by contributing area to either a stream gage for streamflow assessment, or by the USGS HUC level-12 catchment85.

Changes in streamflow from the historical condition were divided into those associated only with temperature increase (ΔT), those associated only with groundwater storage declines (ΔS) and those associated with both increased temperature and reduced storage (total). Changes in streamflow were assessed across a range of historical water year climate inputs that span the full range of historical conditions and capture historical variability in the timing of climate inputs (Supplementary Fig. 27). The experimental design ran three simulations for each water year tested: (1) historical: initial storage set to the 2016 historical condition with no change to historical climate input; (2) ΔT: initial storage set to 2016 historical condition and +4 °C applied to historical climate inputs and (3) total: initial conditions set to 2016 after continuous warming and +4 °C applied to historical climate input. Storage loss effects (ΔS) were isolated by subtracting (3) from (2) (Supplementary Fig. 28). Initial conditions defined by 2016 chosen since historically its groundwater storage is near the historical average and under long-term warming it has deviated from average (Fig. 4a).

Upscaling groundwater risk to warming

MLR models of East River catchments

Water table declines in the East River were calculated by subtracting water table elevations of the +4 °C warming from the historical period at the end of the multi-decadal simulation. Cell-by-cell differences were averaged across each of the ten HUC12 catchments nested within the East River. Catchments range in size from 37 to 112 km2 and contain 3,700–11,200 GSFLOW 100 m grid cells, respectively. Water table declines were then regressed against a total of 35 climate and catchment characteristics. Climate characteristics of minimum and maximum temperature, minimum and maximum vapour pressure deficit, precipitation and solar radiation on a sloped surface use parameter–elevation regressions on independent slopes model 30 year annual average (years 1991–2020). Supplementary Table 1 identifies the parameters and data sources used in the multiple linear regression (MLR) model. A Python code was written to cycle through each combination of possible independent variables with the maximum number of independent variables kept less than or equal to four given the small sample size. For each iteration, the variable inflation factor (VIF) was calculated to assess multi-collinearity between independent variables. The calculation regresses each independent variable (for example, X1) with the other independent variables (for example, X2, X3 and X4) and calculates the coefficient of determination with VIF(X1) of 1/(1 − r2) for each combination. MLR models containing a maximum VIF value greater than 5 were assumed to contain excessive multi-collinearity and discarded. For acceptable models, the Akaike Information Criterion (AIC) was calculated using, \(\mathrm{AIC}=2k-\mathrm{ln}\left({\sum }_{i=1}^{i=n}{\left(O-P\right)}^{2}\right)\), where k is the number of predictor variables, O is the average HUC12 water table decline estimated by GSFLOW and P is the regression-predicted water table decline. AIC accounts for covariance in predictor variables and limits the redundancy of information in the regression analysis that can falsely inflate the strength of the regression. A lower value of AIC is considered a better model based on explanatory power and model simplicity. Model performance was evaluated using the leave-one-out cross-validation, in which all data points except one were used to train the model and were then tested on the single data point. The process was repeated for each data point. The mean squared error in training observations was then averaged across all iterations for a given MLR model.

The best MLR model was assumed to contain the highest predictive power (r2) while maintaining statistical significance (P < 0.01) of independent variables, with lower AIC and VIF considered for MLR models of equal predictive abilities. Of the 1.3 million possible combinations, 27,558 MRL models contained a VIF <5 and 41 models had a coefficient of determination >0.90. The best MLR model was defined as (dWT) of −7.13 + 1.48(TWI) + 0.0019(relief) + 3.81(fraction area grass) + 3.14(fraction area conifer) (r2 = 0.96 and P = 0.004). All independent variables are significant at P ≤ 0.002. The dependent variable dWT is the change in water table elevation with dWT >0, indicating the water table elevation becomes lower with warming. The topographic wetness index (TWI) is a measure of potential moisture accumulation (TWI is ln(flow accumulation / (tan(slope) + 0.01)). It was calculated using ArcGIS Hydrology tools with a 30 m DEM for fill, flow direction and accumulation. The slope is in radians. Higher values indicate locations where water is likely to collect based on topographic depressions or a lower slope. The fraction of the area that is conifer forest is the variable of largest predictive ability. Other metrics of evaluation include AIC of −12, maximum VIF of 3.25 and leave-one-out cross-validation mean squared error of 0.03 m.

Upscaling across the Upper Colorado River basin

Mountainous catchments in the Upper Colorado River basin are defined using a ruggedness metric, where ruggedness is calculated as topographic relief (elevation maximum minus minimum) across a focal averaging of roughly 7 km (231 30 m grid cells) applied across a rectangle about the point. Roughness values were then averaged for a given HUC12 catchment using zonal statistics in ArcGIS. A comparison with the Global Mountain Biodiversity Assessment Inventory v2 (ref. 86) indicates close alignment for a roughness of approximately 400 m (Supplementary Fig. 29). The results suggest 1,726 HUC12 catchments (150,786 km2) or 41% of the Upper Colorado River basin area are classified as mountainous. Independent variables for predicting change in groundwater elevation related to warming (TWI, relief, fraction area grass, fraction area conifer) were calculated for all mountain catchments (Supplementary Figs. 30–33). The quartile method was applied to estimated water table declines. Groundwater risk was defined as low (<0.72 m), moderate (0.72–1.58 m), high (1.58–2.24 m) and extreme (2.24–4.67 m). Limitations to the MLR include ignoring anthropogenic effects on groundwater storage from irrigation, groundwater pumping and reservoir storage. Future work will need to evaluate the assumed linear relationships between climate parameters and change in groundwater.