1 Introduction
The Upper Colorado River Basin (UCRB) (Figure 1a ) constitutes approximately 90% of the total flow of the Colorado River, one of the most over-allocated water resource regions in the world (Christensen et al., 2004 ; Fleck & Udall, 2021 ; Lukas & Payton, 2020 ; McCabe & Wolock, 2007 ; Xiao et al., 2018 ). Since the onset of the ongoing Millennium drought (“drought”) (2000–present), record streamflow deficits have occurred relative to both the observational record (1906–1999) and the baseline period used in this analysis (1964–1999) (Christensen et al., 2004 ; Fleck & Udall, 2021 ; Hoerling et al., 2019 ; Lukas & Payton, 2020 ; McCabe & Wolock, 2007 ; Schmidt et al., 2023 ; Williams et al., 2022 ; Xiao et al., 2018 ). The mountain winter snowpack is the most critical component of UCRB streamflow (Christensen et al., 2004 ; Heldmyer et al., 2023 ; Li et al., 2017 ; Palmer, 1988 ), but during the drought period, streamflow has frequently been less than anticipated given the amount of winter precipitation (Milly & Dunne, 2020 ; Udall & Overpeck, 2017 ; Woodhouse et al., 2016 ) (Figure 1b ). Annual precipitation decreases over the same period also do not fully account for the observed streamflow decline, suggesting there is a certain amount of “missing water” from the UCRB (Figure 1c ) (McCabe et al., 2017 ; Milly & Dunne, 2020 ; Woodhouse et al., 2016 ).
(a) Map of UCRB with the selected headwater basin locations and UCRB outlet at Lee’s Ferry (red star). (b) April 1 SWE snow course measurements (1964–2022) in Black Gore Creek basin and area-normalized annual streamflow (USGS gage ID 09066000) in mm for baseline (blue) and Millennium drought (red) years. Using the best fit line, resultant streamflow estimates from 500 mm SWE are displayed in blue (baseline) and red (Millennium drought) boxes. (c)10-year rolling average of normalized anomalies for UCRB naturalized streamflow as estimated at Lee’s Ferry (dark blue line) and UCRB precipitation from PRISM (light blue line).
To address this conundrum, previous modeling studies explored the sensitivity of streamflow to precipitation and temperature. Streamflow response to decreased precipitation is magnified during drier years (Woodhouse & Pederson, 2018 ). On the other hand, temperature increases in the UCRB have been linked to streamflow declines through (a) increases in modeled wintertime sublimation and annual evapotranspiration (ET) (Christensen & Lettenmaier, 2006 ; Ficklin et al., 2013 ; McCabe et al., 2017 ; Udall & Overpeck, 2017 ; Xiao et al., 2018 ), and (b) earlier snow disappearance, which reduces surface albedo and increases potential evapotranspiration (PET) (Milly & Dunne, 2020 ). While it is agreed that increased temperature exacerbates UCRB streamflow declines, estimates of this temperature sensitivity vary between studies (Hoerling et al., 2019 ; McCabe & Wolock, 2007 ; Milly & Dunne, 2020 ; Vano et al., 2012 ; Xiao et al., 2018 ). The variation is likely due to different model estimations of precipitation and temperature and inherent model differences in how precipitation is allocated to streamflow, which diverge considerably between models (Vano et al., 2012 ). Thus, the choice of model can significantly influence how streamflow changes are explained (Bennett et al., 2019 ).
To avoid challenges with model choices, we focus on observed precipitation and streamflow changes, and on estimates of PET as an approximation of non-water stressed ET, within unregulated UCRB headwater basins over 1964–2022. We specifically examine changes by season because seasonality strongly influences the energy available to transpire or evaporate water and the water availability in soil (Carroll et al., 2020 , 2024 ; Foster et al., 2016 ; Hidalgo et al., 2005 ; Knowles et al., 2015 ; Lapides et al., 2022 ; Meira Neto et al., 2020 ; Milly & Dunne, 2020 ; Woodhouse et al., 2016 ; S. Zhao et al., 2021 ), and seasonality has recently been linked to unexpected streamflow losses in California (Lapides et al., 2022 ). In semi-arid basins like the Colorado, PET magnitudes closely match actual ET (AET) observations during non-water stressed conditions (Fisher et al., 2005 ; Maes et al., 2019 ), which occur in the spring during and following snowmelt. In winter, AET sometimes matches PET, but with limited energy available, PET remains low. In summer and fall, even with sufficient energy, drier conditions constrain AET.
Spring also marks a rapid transition in snow-dominated headwater regions. As snow melts, surface albedo decreases over newly exposed ground, PET rises quickly, and the annual phenological cycle begins. About 30% of total UCRB precipitation occurs during this dynamic season (Zhao et al., 2023 ). Observations indicate a decreasing trend in spring precipitation in the UCRB over the last century, and future projections forecast these decreases to continue (Kalra & Ahmad, 2011 ; McAfee & Russell, 2008 ; Yin, 2005 ). Additionally, drier springs correlate with warmer temperatures throughout the western US, even after removing the effects of climate change-driven temperature increases (Koster et al., 2009 ; Trenberth & Shea, 2005 ).
Drier and warmer springs shift the timing of plant growth earlier, while decreases in precipitation (and its associated cloud cover) increase mean temperature and surface radiative input, both crucial contributors to PET (Betts et al., 2014 ; Cayan et al., 2001 ; Fischer et al., 2009 ; Hidalgo et al., 2005 ; Schwartz & Reiter, 2000 ; Sumargo & Cayan, 2018 ; W. Zhao & Khalil, 1993 ). Thus, we expect years with lower spring precipitation to correspond with higher spring PET.
Additionally, basin elevation has been identified as a primary driver for how a basin responds to temperature, precipitation, and land cover change (Biederman et al., 2022 ; Carroll et al., 2019 ; Foster et al., 2016 ; Mayer & Naman, 2011 ; Tennant et al., 2015 ) (see Supporting Information S1 ). Therefore, we hypothesize that reduced spring precipitation and higher spring PET do not impact all basins equally during drier springs. Snow in lower elevation headwater basins melts out earlier (Fassnacht et al., 2003 ; Lundquist et al., 2004 ). Earlier snowmelt impacts spring PET over these lower, but not higher, elevation basins because surface albedo decreases earlier, increasing the radiative input into the surface, and plant-driven ET starts earlier with the exposure of plants to sunlight and available water from snowmelt (Milly & Dunne, 2020 ; Nehemy et al., 2021 ).
This study uniquely explores how the relationship between spring precipitation and PET affects streamflow throughout the UCRB headwaters. The combined effects from this relationship have not been previously linked to observed streamflow deficits in the UCRB. To confront this research gap, we primarily use observation-based data to (a) establish the significance of spring precipitation and streamflow decreases since 2000, (b) determine the relationship between spring precipitation decrease and PET, and (c) evaluate the role this relationship has in explaining streamflow declines in headwater basins at different elevations.
2 Data and Methods
2.1 Study Basins
Twenty-six UCRB headwater basins (Figure 1a ) were selected from the USGS Hydro-Climatic Data Network (HCDN), which identifies natural streams minimally influenced by humans (Slack, Lumb, & Landwehr, 1994 ). Prior work has shown that 80% of UCRB streamflow originates from only 15% of its area (Christensen & Lettenmaier, 2006 ). Similarly, the selected basins account for 25% of the average UCRB streamflow but cover only 4% of its land area. These basins span the geographic extent of the UCRB (Figure 2a ) and have comparable elevation range (2,290–4,200 m) and forest cover (28%) to the entire headwater region’s elevation range (2,270–4,300 m) and forest cover (25%).
(a) UCRB boundary with higher, middle, and lower elevation basin groupings. Arrows point to basins used in Figure 4 . (b) Spring precipitation percent change (shading) and annual streamflow volume difference (symbols) between baseline and drought. Symbols represent significant (red triangles) and insignificant (white circles) streamflow volume (Q ) decreases from study basins. Symbol size represents the Q percentage change.
2.2 Seasonal Precipitation Change
Gridded UCRB precipitation data were collected from the Parameter-elevation Regressions on Independent Slopes Model, PRISM (PRISM Climate Group, 2014 ), for the 1964–2022 period. The monthly precipitation product was spatially averaged over each selected basin. Seasonal and annual aggregates of precipitation were created for each basin for each water year in the data record (see Text S2 in Supporting Information S1 ). We evaluated annual and seasonal precipitation changes in each basin by comparing mean precipitation totals between the baseline (1964–1999) and drought (2000–2022) periods.
To verify results from PRISM, we compared monthly PRISM estimates against two other commonly used gridded precipitation data sets (NClimgrid and ERA5-Land reanalysis, see Text S2 in Supporting Information S1 ). We observed consistent trends and strong agreement between products, especially during winter, spring, and fall seasons.
2.3 Basin Streamflow Changes Between Periods With Elevation
Area-normalized water year (October–September) streamflow was calculated from daily mean observations (U.S. Geological Survey, 2016 ). For the entire UCRB, we utilized the Bureau of Reclamation naturalized streamflow record at Lee’s Ferry, AZ. These data have been used in several model evaluation and water balance studies in the UCRB (Hoerling et al., 2019 ; Vano et al., 2012 ; Woodhouse et al., 2016 ; Xiao et al., 2018 ).
We separated headwater basins into higher (>3,250 m), middle (2,950–3,250 m), and lower (<2,950 m) elevation groupings based on their area-normalized mean elevation (Figure 2a , Text S2 in Supporting Information S1 ). Like our evaluation of precipitation change, we compared streamflow volume changes for the UCRB and headwater gages by comparing mean annual streamflow totals between periods.
2.4 Spring Precipitation and PET Relationship
To estimate monthly PET, we used an offline “energy-only” PET (PET-EO) method (Maes et al., 2019 ; Milly & Dunne, 2016 , 2017 ) derived from ERA5-Land reanalysis net radiation estimates (Muñoz-Sabater et al., 2021 ). Maes et al. (2019 ) found the PET-EO method better matched in-situ measurements of non-water stressed ET than the commonly used Penman-Monteith and Priestley-Taylor PET methods. While magnitudes of PET vary, the relationship between spring precipitation and PET is consistent across all methods examined (Text S2 in Supporting Information S1 ).
Net radiation during spring over non-snow-covered terrain is primarily controlled by changes in cloud cover, which impacts incoming solar radiation (Hidalgo et al., 2005 ); however, snow albedo also strongly influences the surface energy balance by controlling outgoing shortwave radiation when snow is present (Milly & Dunne, 2020 ). While the ERA5-Land radiation variables that drive the PET-EO method do account for modeled surface albedo changes over the year (Hersbach et al., 2020 ), the ERA5-Land albedo in the UCRB headwater regions was generally insensitive to snow cover and did not match satellite albedo observations during snow-covered periods (Figure S11 in Supporting Information S1 ). Modeled cloud cover from ERA5 correlated well with observations from meteorological stations distributed throughout the UCRB (United States. National Oceanic and Atmospheric Administration, 1998 ) from 1997 to 2022 (Text S2 in Supporting Information S1 ).
Based on these analyses, the unadjusted ERA5-Land radiation components represent cloud-related processes well, but not the albedo of snow-covered terrain. Thus, we used the net longwave and shortwave radiation values to estimate the impact that changes to cloud cover have on PET (PET-EO-cloud). We examined PET-EO-cloud estimates against modeled ERA5 cloud cover and PRISM precipitation to illustrate the relationship between spring PET, cloud cover, and precipitation over the entire 1964–2022 period.
To better account for the impact of snow cover on PET changes, our second method (PET-EO-snow) adjusted the surface albedo over snow using MODIS observations of fractional snow-covered area, fSCA, on April 15 (the midpoint of March-April-May) from 2001 to 2017 (A. P. Tran et al., 2019 ). Here, we assume snow covered grid-cells (fSCA = 1) have a surface albedo of 0.8, while snow-free grid-cells (fSCA = 0) have an albedo equal to the ERA5-Land modeled snow-free albedo value which varies by land cover, but is generally about 0.2. For fSCA values between 0 and 1, albedo was calculated by combining the snow-covered and snow-free albedo in proportion to fSCA. For further details, see Text S2 in Supporting Information S1 . Using these fSCA and PET estimates, we grouped average spring values by 100-m elevation bins over the UCRB to explore how spring snow cover and PET changed with elevation during dry versus wet springs between 2001 and 2017. We classified springs as dry (wet) if their precipitation was more than one standard deviation below (above) the seasonal average.
3 Results
3.1 How Much Have Spring Precipitation and Streamflow Decreased Since 2000?
Annual precipitation decreases were significant (p < 0.05) across the UCRB, with a 9% reduction between the baseline and drought periods. In the headwater basins, 17 of 26 basins had significant annual precipitation decreases, of which 13 were lower and middle elevation basins. Seasonally, spring precipitation decreases were largest, averaging 14% across the study basins (Figure 2b ). Most headwater basins, except one, showed spring precipitation decreases up to 29% between periods, with significant decreases in 16 of the 26 basins. Lower and middle elevation headwater basins were particularly affected, as 12 of 17 basins showed significant decreases. Basin-specific results for annual and seasonal precipitation changes are detailed in Table S2 of Supporting Information S1 .
At both the headwater basin and UCRB scales, spring precipitation reductions accounted for 40%–65% of the annual precipitation decreases between periods (Figure S15 in Supporting Information S1 ). Fall and winter showed limited change, although some basins had significant summer precipitation decreases. However, unlike spring, these summertime decreases did not correlate with observed streamflow, and only accounted for 15% of the average annual precipitation decreases (Text S2 and Figure S9 in Supporting Information S1 ).
Annual streamflow reductions occurred at all headwater basin gages and in the UCRB naturalized streamflow record. Significant decreases occurred in four of the nine higher elevation basins, ranging from −15% to −19%. Thirteen of the 17 middle and lower elevation basins showed significant annual streamflow decreases (−8% to −36%). Most basins with significant annual streamflow reductions also had significant spring precipitation decreases (Figure 2b , Table S3 in Supporting Information S1 ), with the largest relative streamflow declines concentrated in the lower elevation basins (Figures 2b and 4b ).
3.2 How Much Can the Spring Precipitation-PET Relationship Explain Streamflow Deficits Since 2000?
Across the headwater basins analyzed, we found drier springs had less cloud cover and greater PET (Figure 3a ). Spring was the only season where we found PET to be consistently sensitive to precipitation, with a 10% precipitation reduction corresponding to a 1%–3% PET increase (Figure S13E in Supporting Information S1 ). Aligning with this sensitivity, we found spring PET increased between 2% and 10% since 2000 across the headwater basins.
(a) Spring potential evapotranspiration versus spring precipitation, color-coded by ERA5 cloud cover. Values represent spatial average across analyzed headwater basins for 1964–2022 derived from the PET-EO-cloud estimates. (b) Mean spring precipitation changes (blue squares) and mean spring precipitation plus PET-EO-cloud changes (orange circles) plotted against mean streamflow deficits between baseline and drought periods for all headwater basins with significant streamflow declines between periods.
In basins with significant streamflow decreases since 2000, spring precipitation decreases alone fall short of explaining observed streamflow deficits (Figure 3b ), particularly in basins with the largest streamflow changes. Regressing spring precipitation changes against streamflow declines accounted for 56% of the variance (as quantified by R 2 ). However, by including the sum of water losses (spring PET-EO-cloud and precipitation changes), the variance explained increased to 67%, and the Nash-Sutcliffe efficiency increased from 0.02 to 0.62 as observed points moved closer to the 1-to-1 line (Figure 3b ).
3.3 Are Spring Precipitation and PET Impacts on Streamflow Elevation-Dependent?
Throughout the UCRB, we found earlier snowmelt timing, lower fSCA, and greater spring PET in lower elevation basins (Figure 4a ) compared to higher basins (Figure 4d ). April 15 fSCA also differed with elevation when comparing wet and dry springs, with the fSCA difference (ΔSCA) expanding to a maximum of 0.25 in the lower headwater elevation band (2,400–3,000 m) before decreasing again at the lowest elevations of the UCRB (Figure 4a ). During years with low spring precipitation, we observed an amplified area-normalized spring PET response across all elevations compared to years with high spring precipitation in both the PET-EO-cloud and PET-EO-snow estimates (Figure 4b ).
(a) Hypsometric plot of UCRB area fraction (black), tree-covered area (green), and fSCA for wet (blue) and dry (red) spring conditions. Difference between wet and dry years is symbolized as ΔSCA (b) Hypsometric plot showing area-normalized spring potential evapotranspiration (PET) using the PET-EO-cloud (solid) and PET-EO-snow (dashed) methods for wet (blue) and dry (red) spring conditions. Difference between wet and dry years is symbolized as ΔH , ΔM , ΔL for higher, middle, and lower elevation basins, respectively. (c) Hypsometric plot of total spring PET, accounting for area, using the PET-EO-snow method for wet (blue) and dry (red) spring conditions. Lower, middle, and higher elevation basin bands highlighted in red, green, and blue, respectively. (d) Basin-normalized average daily streamflow during baseline (gray) and drought (colored) periods, in mm/day, for representative higher, middle, and lower elevation basins. Locations marked in Figure 2a . Text boxes show percentage change in streamflow volume relative to baseline. Mean snow disappearance date from MODIS snow-covered area data (2000–2018) indicates when 50% of each basin’s snow-covered area disappeared (O’Leary et al., 2020 S1 .
In the headwater regions above 2,400 m, differences in spring PET between wet and dry springs also increased with decreasing elevation. The mean differences in PET between wet and dry springs for the higher (ΔH), middle (ΔM), and lower (ΔL) headwater basin elevation bands were 25, 37, and 50 mm, respectively, for the PET-EO-cloud estimates, and 28, 56, and 78 mm, respectively, for the PET-EO-snow estimates (Figure 4b ). Considering that a larger proportion of the UCRB’s area is contained within the lower elevation headwater regions, more basin total spring PET occurred at these lower elevations during dry springs compared to wet springs (Figure 4c ).
4 Discussion
4.1 Consequences of Spring Precipitation and PET Changes on Streamflow
Springtime precipitation decreases since 2000 reduced water input across the selected UCRB headwater regions and the greater basin area (Figure 2b ). Nevertheless, precipitation decreases alone did not fully explain the observed streamflow deficits (blue squares, Figure 3b ). To account for the remaining “missing water,” we showed that drier springs have greater PET due to feedback from reduced cloud cover and earlier snowmelt timing that increased net radiation (Figures 3a and 4b ). Combining spring PET increases with observed precipitation decreases provides a better explanation for UCRB streamflow deficits during the ongoing drought (orange circles, Figure 3b ).
Reduced cloud cover allows greater incident solar radiation to reach the surface, leading to warmer temperatures and greater net radiation. Over snow-free terrain, this cloud-radiative feedback drives PET increases (Figures 3a and 4b ), in agreement with observations from California (Hidalgo et al., 2005 ). Over snow-covered terrain, greater incoming radiation during drier springs accelerates snowmelt, reducing snow cover earlier at all elevations to varying extents (Figure 4a ). This reduction in snow coverage leads to a decrease in outgoing radiation resulting from the lowered surface albedo, consequentially increasing net radiation and PET. This earlier ground exposure and increased energy allows plants to respond over a longer spring period when ET remains energy-limited.
While prior work has investigated long-term trends toward earlier snowmelt timing in the western US (Clow, 2010 ; Fritze et al., 2011 ), the feedback-driven changes to snow-covered area and net radiation presented here are related to inter-annual variability in spring weather patterns (wetter/drier). Given uncertain climate projections of spring precipitation, it is unclear whether the springtime reductions observed since 2000 will persist. Further research is needed to better understand the relationship between long-term climate signals and spring weather patterns in the UCRB. If climate change directly impacts spring precipitation, streamflow losses would be expected to amplify.
4.2 Explaining Spring Precipitation and PET Impacts at Different Elevations
Just as the headwater regions (elevations above 2,300 m) receive the most precipitation in the basin, they also had the greatest net radiation and PET differences between wet and dry springs (Figures 4b and 4c ), while the lowest elevations exhibited a minimal signal of basin-wide changes since 2000 (Figures 4c and Text S2 in Supporting Information S1 ). The highest elevation basins (elevations above 3,300 m) experienced more modest reductions in annual streamflow volume caused by the smaller percentage-wise spring precipitation changes since 2000 and smaller fSCA and PET shifts during drier springs (Figure 4b , Table S2 in Supporting Information S1 ). These basins are less vulnerable to these shifts due to their comparatively deeper snowpacks and later snowmelt timing (Figure 4d ), which limits PET during spring and buffers against the spring precipitation-PET relationship’s influence on streamflow.
Spring precipitation percent declines and PET increases were largest within the lower elevation headwater zones (between 2,300 and 2,900 m) of the UCRB. At these elevations, drier springs also exhibited a larger fSCA reduction and greater spring PET increase (Figures 4a and 4b ). Without extensive or long-lasting snow cover, the combined effects of spring precipitation decrease and PET increase impacted streamflow the most in lower elevation headwater zones (Figure 4d ). This result corroborates findings from Knowles et al. (2015 ), which showed that dry years preserved ET at the expense of streamflow in a lower elevation region of the UCRB headwaters. However, focusing solely on spring changes, we demonstrate that spring PET and precipitation changes closely matched annual streamflow deficits in all basins during the ongoing drought (Figure 3b , Figure S17 in Supporting Information S1 ), further supporting that spring PET is a good indicator of AET. Thus, energy availability during spring is likely the primary control on AET variability in the UCRB headwaters. Ignoring other water balance losses, our results indicate that drier, warmer springs contributed to more total AET from a streamflow perspective. The data set compiled herein, which tries to avoid modeling, may help with assessing which hydrologic models are appropriately sensitive to springtime precipitation and snowmelt influences on ET.
4.3 Importance of Accurate Seasonal Precipitation Forecasts
Drier springs in the UCRB correlate significantly to warmer spring temperatures (Figure S16 in Supporting Information S1 ). Since 2000, we observed a significant shift toward drier spring conditions, with a north-south-oriented precipitation gradient (Figure 2b ) that could signal the predicted climate change-driven northward shift in storm tracks (Harvey et al., 2020 ). Our findings corroborate model results from Heldmyer et al. (2023 ), which showed when warm and dry conditions occur together, a greater loss in UCRB streamflow can be expected. Nonetheless, our research emphasizes the critical role of spring precipitation in driving UCRB streamflow, underscoring the need for further investigation into the underlying causes of these precipitation reductions. Recent research has used oceanic teleconnections, specifically North Pacific sea surface temperature, to predict spring precipitation and streamflow in the UCRB with encouraging results (Zhao et al., 2021 , 2023 ). A better forecast of spring precipitation would benefit water managers in the UCRB region (Lehner et al., 2017 ; Zhao et al., 2023 ).
5 Conclusion
Since the onset of the ongoing Millennium drought in the UCRB (2000–present), annual precipitation and April 1 SWE have provided less streamflow than historically observed. We focused on changes during the spring season to explain the observed inconsistency between precipitation and annual streamflow declines in select UCRB headwater basins during the drought. Our analysis revealed significant spring precipitation decreases in these basins between the baseline (1964–1999) and drought periods, which notably outweighed decreases during other seasons. We found years with less spring precipitation and cloud cover corresponded with warmer temperatures and greater spring PET. Including both spring precipitation reductions and spring PET increases better accounted for annual streamflow deficits in these headwater basins compared to precipitation decreases alone. By combining these water losses, we explained 67% of the variance in annual streamflow deficits within these headwater basins since 2000. The compounding impacts of these changes reached a maximum in the lower elevation headwater basins, where the largest relative reductions in annual streamflow volumes were observed.
With continued projections of increasing temperature during spring, it will be paramount to highlight the importance of springtime conditions throughout the UCRB and to better resolve our estimates of seasonal spring precipitation when forecasting water availability. Further work should focus on understanding regional precipitation changes during spring and the interplay between precipitation and other hydrologic fluxes during this dynamic season.
Acknowledgments
We gratefully acknowledge funding support for this analysis from the National Science Foundation (Award 2139836), Sublimation of Snow, and the Department of Energy Environmental System Science Division (Award DE-SC0024075), Seasonal Cycles Unravel Mysteries of Missing Mountain Water. Additionally, we thank the Sublimation of Snow project members and the Mountain Hydrology Research Group at the University of Washington for their valuable feedback and insight.
