1 Introduction

California water resources are vulnerable to anthropogenic climate change, in part because the state's water storage infrastructure was built under the assumption that climate patterns would remain stable. The majority of California's annual precipitation arrives between October 1 and April 1, and must be stored for use during later dry periods (Dettinger et al., 2011). Historically, snowpack has provided a natural water storage service, retaining winter and spring precipitation and delivering that water slowly to downstream agricultural regions, urban areas, and ecosystems as the snowpack melts. Water is also stored in a network of reservoirs throughout the state, including those along the western slope of the Sierra Nevada (M. D. Dettinger & Anderson, 2015). The southern Sierra Nevada—which we define as the Stanislaus, Tuolomne, Merced, San Joaquin, Kings, Kaweah, Tule, and Kern watersheds for this study—is a unique and critically important subrange due its physical characteristics and economic importance. These factors include relatively high elevations compared to the rest of the Sierra Nevada (Null et al., 2010), supply of water to the San Joaquin Valley which produces over half of California's agricultural output (Escriva-Bou et al., 2023), increasing flood risk in the coming decades (Huang et al., 2018), large hydropower generating capacity (Null et al., 2010), and vulnerability to disturbances such as severe wildfire (Steel et al., 2023).

Snowpack is projected to become a less reliable water reservoir throughout the western US in the coming decades (Li et al., 2017; Pierce & Cayan, 2013; Siirila-Woodburn et al., 2021); this decline has already been observed in some locations (Berg & Hall, 2017; Mote et al., 2018; Pierce et al., 2008). In the Sierra Nevada, this is not driven by reduced precipitation, as average annual precipitation amounts are not expected to change significantly (Dettinger et al., 2018). California experiences larger interannual precipitation volatility than any other US state (Dettinger et al., 2011) and the state's precipitation extremes—both drought and flood—are expected to intensify under climate change (Berg & Hall, 2015; Swain et al., 2018). Delayed onset of the California's rainy season has also been projected (Swain et al., 2018) and observed (Lukovic et al., 2020). In contrast to average precipitation, average annual temperatures are projected to increase 3.3–5.0°C by end-of-century in the Sierra Nevada (Dettinger et al., 2018). Warming causes precipitation type change, as more precipitation falls as rain instead of snow, and earlier snowmelt (Pierce & Cayan, 2013; Reich et al., 2018; Shulgina et al., 2023). The southern Sierra is particularly vulnerable to earlier snowmelt (Liu et al., 2020; Musselman, Molotch, & Margulis, 2017; Null et al., 2010), as well as increased peak melt rates and consequently flood hazard (California Department of Water Resources, 2022; Huang et al., 2018; Musselman, Molotch, & Margulis, 2017). Projected temporal shifts across the Sierra Nevada also include earlier peak snowpack (Marshall et al., 2019), earlier runoff (Schwartz et al., 2017), and diminished runoff efficiency (Rhoades et al., 2022). Temporal shifts make management more challenging, even if the same or a greater amount of precipitation than historically is received.

To maximize utility of climate model data for managers, models must be able to resolve mountain topography at local scales and accurately represent future earth systems. Fine spatial resolution (<10-km) output allows models to capture elevational variations which underpin representation of orographic uplift and snow-rain transitions. Dynamical downscaling is preferable to statistical or hybrid downscaling approaches to accurately represent future earth systems, in part because it does not explicitly assume stationarity. Previous generations of downscaled climate model data have had one of these two strengths—fine spatial resolution or dynamical downscaling—but not both. For instance, the NA-CORDEX (Mearns et al., 2017) data set provided dynamically downscaled projections, but was limited by its 25-km spatial resolution. The LOCA method (Pierce et al., 2014) was used to generate statistically downscaled projections for California's Fourth Climate Change Assessment. These data have a fine spatial resolution (6-km), but are limited by stationarity assumptions. Here, we leverage the Western United States Dynamically Downscaled Data set (WUS-D3, Rahimi et al., 2024a, 2024b), which provides 9-km, dynamically-downscaled output from 15 Coupled Model Intercomparison Project 6 (CMIP6) general circulation models (GCMs). Several studies have assessed WUS-D3 snowpack water storage skill, but no independent studies have quanitatively evaluated WUS-D3 skill at capturing intrannual snowpack processes, and none have explicity focused on the Sierra Nevada. Visual assessment of WUS-D3 intrannual skill suggests it is roughly skillful at capturing peak SWE timing (Adhikari et al., 2024; Cowherd et al., 2024), but quantitative assessment is required to measure utility of WUS-D3 data to managers. For example, if WUS-D3 is less skillful at capturing melt processes than timing and magnitude of snowpack peaks, this undermines the utility of WUS-D3 data for managers interested in reservoir inflow and outflow.

Here, we apply the SWE triangle methodology (Rhoades et al., 2018a) to WUS-D3 in order to provide a novel analysis of WUS-D3 skill at capturing intrannual snowpack processes in the southern Sierra Nevada. Our evaluation of WUS-D3 skill across a range of intrannual hydroclimate measures directly informs the data set's ongoing use in California's Fifth Climate Change Assessment, and provides insights for researchers about where downscaled climate models still need to be improved for relevance to natural resource managers. After assessing WUS-D3 skill, we leverage this data set to explore future changes in snowpack water storage median magnitude and timing. This section builds upon a robust literature with projections from a novel data set, providing updated findings that water managers can use in regional planning. We conclude with brief directions for future research.

2 Methods

2.1 Instrumental and Downscaled Climate Model Data Sets

Instrumental SWE data were obtained from the National Water and Climate Center Air and Water Database Report Generator for 34 California Cooperative Snow Surveys monitoring stations in the southern Sierra Nevada (Figure 1; NWCC, California Cooperative Snow Surveys, 2024). The elevations of these monitoring stations range from 2000 to 3500 m, with a mean elevation of 2600 m. The period of record for these daily instrumental SWE data begins in the 1980s for most monitoring stations.

SWE data from 15 CMIP6 GCMs and the ECMWF Reanalysis v5 (ERA5) data set covering the period 1980–2100 were obtained from WUS-D3 (Figure 3). These data were dynamically downscaled using the Weather Research and Forecasting (WRF) model to a 9-km grid. After 2015, GCM boundary conditions reflect Shared Socioeconomic Pathway 3–7.0. The reanalysis-driven experiment is hereafter referred to as WRF-ERA5 and the downscaled GCMs are referred to as WRF-GCMs. See SI and Rahimi et al. (2024a, 2024b) for more detail.

2.4 Methods to Evaluate Model Skill and Future Change

To evaluate WRF-GCM skill at capturing intrannual snowpack processes, we compare instrumental, WRF-ERA5, and WRF-GCM data in the historical period (Figure 2). In addition, we evaluate variability across the WRF-GCMs for each date and metric in the historical period (Figure 3). To do this, we compare WRF-GCM means to the instrumental mean for each date and metric using z-scores, which allow efficient comparison of model skill across a range of measures with different units (Figure 4). These z-scores are derived from the following equation:

$$$z=\frac{\overline{x}-\mu }{\sigma }$$$ (1)

where $$\overline{x}$$ = the mean value of a date or metric for an individual WRF-GCM in the historical period, μ = the instrumental mean of a date or metric, and σ = the instrumental standard deviation of a date or metric. Thus, a z-score compares a WRF-GCM mean to the instrumental mean in terms of instrumental standard deviations. This method allows assessment of whether the WRF-GCM mean falls within the range of natural variability observed in the instrumental record. To project future changes in southern Sierra Nevada snowpack, we analyze WRF-GCM ensemble values for each date and metric in the mid- and end-of-century periods (Figure 4).

3 WRF-GCM Skill at Capturing Intrannual Snowpack Processes

The WRF-GCM ensemble is highly skillful at capturing peak SWE timing, but overestimates peak SWE. The ensemble hindcasts a median peak date of $${178}_{175}^{181}$$ (median 95% confidence interval indicated with superscript and subscript). This is only 2 days later than the instrumental value of $${176}_{165\,}^{179}$$ and 6 days later than the WRF-ERA5 value of $${172}_{163}^{178}$$ (Figure 2). Additionally, each of the individual WRF-GCM mean peak date values are within 0.5 instrumental standard deviations (8.4 days) of the instrumental mean (Figure 3). Conversely, the WRF-GCM ensemble overestimates peak SWE relative to both WRF-ERA5 and instrumental; WRF-ERA5 overestimates peak SWE relative to instrumental. The WRF-GCM ensemble median peak SWE value of $${783}_{733}^{839}\,\text{mm}$$ is approximately 19% greater than the instrumental median peak SWE value of $${658}_{550}^{909}\,\text{mm}$$ and 9% greater than the WRF-ERA5 median of $${721}_{594}^{969}\,\text{mm}$$ (Figure 2). WRF-GCM ensemble overestimation of the peak SWE relative to WRF-ERA5 likely reflects internal climate variability, while WRF-ERA5 overestimation of peak SWE relative to instrumental likely reflects meterological driving data set and WRF solver choices (see SI). WRF-GCM overestimation of peak SWE in the historical period indicates that the peak SWE values projected in the mid- and end-of-century periods (Figure 4) are potentially overestimates as well.

Despite advances in techniques, melt processes remain difficult for downscaled climate models and coupled land surface models to capture, which undermines their utility for managers interested in projecting water supply and flood risk. The WRF-GCM ensemble hindcasts a median melt date of $${267}_{264}^{272}$$ that is 19 days later than the instrumental median melt date of $${248}_{239}^{254}$$ and 4 days later than the WRF-ERA5 median melt date of $${263}_{254}^{280}$$ (Figure 2). The WRF-GCMs also uniformly overestimate melt date (Figure 3). WRF-GCM overestimation of melt date timing is likely driven by WRF biases because the WRF-GCM ensemble and WRF-ERA5 both overestimate melt date by a similar amount relative to instrumental. WRF-GCM overestimation of melt date in the historical period indicates that snowmelt may occur earlier than the WRF-GCM ensemble projects in the mid- and end-of-century periods (Figure 4).

WRF-GCM/WRF-ERA5 overestimation of peak SWE is likely driven by WRF-GCM cold bias at high-elevations that persists despite bias-correction, while WRF-GCM/WRF-ERA5 overestimation of melt date is likely driven by melt rate bias. Given that monitoring stations are slightly higher than corresponding WRF gridpoints on average, elevational differences do not explain this overestimation. The mean state bias adjustment applied prior to downscaling (Rahimi et al., 2024b) is intended to correct for some cold bias at high-elevations (Rudisill et al., 2024). The WRF-GCM ensemble ($${8.3}_{8.0\,}^{9.0}\text{mm}/\text{day}$$) and WRF-ERA5 ($${7.1}_{6.2}^{9.1}\,\text{mm}/\text{day}$$) both underestimate melt rate compared to the instrumental value of $${8.7}_{8.1}^{11.2}\,\text{mm}/\text{day}$$ (Figure 2), a characteristic shared across all WRF-GCMs (Figure 3). This stands in contrast to the findings of Rhoades et al. (2018b), which demonstrated regional climate model overestimation of melt rates. However, an assessment of Noah-MP—a key land surface model input to WUS-D3—found that the base case Noah-MP parameterization consistently underestimates daily melt rate across SNOTEL stations—a network of snow monitoring sites across the western United States (Von Kaenel & Margulis, 2024).

4 Future Changes in Snowpack Water Storage Magnitude and Timing

Despite unchanging average annual precipitation amounts, California may lose some of its water supply due to a confluence of three factors: (a) diminished peak SWE values, (b) increasing interannual precipitation variability and decreasing interannual peak SWE variability, and (c) a water management system that is not able to handle these increasing hydroclimate extremes without the assistance it historically received from snowpack. For (a), we find that the median peak SWE value over the southern Sierra declines to 74% of its historical value from $${783}_{753}^{839}$$ mm to $${581}_{538}^{653}$$ mm by mid-century, and to 48% of its historical value $$\left({377}_{337}^{409}\right.$$ mm) by end-of-century (Figure 4). This SWE loss is linked to $${2.33}_{1.94}^{2.85}$$ °C of warming by mid-century and $${4.57}_{4.29}^{5.06}$$ °C of warming by end-of-century across the 34 WRF gridpoints, which indicates a SWE loss rate of about 88 mm per °C of regional warming. Diminished peak SWE values indicate more water will run off of southern Sierra Nevada watersheds during the wettest times of year. While we focus on climate-changed average WY conditions here, our findings suggest that future wet years will likely be subject to the same diminished snowpack conditions. Given WRF-GCM overestimation of peak SWE (Figure 2), water managers should plan for at least the amount of SWE loss that we project here. For (2), the frequency of extreme wet years (exceeds historical 95th percentile annual precipitation) is projected to double by mid-century (Berg & Hall, 2015), and potentially triple by end-of-century (Berg & Hall, 2015; Swain et al., 2018). The frequency of extreme dry years (less than historical 5th percentile annual precipitation) is projected to nearly double across the state by mid-century (Berg & Hall, 2015), with smaller-magnitude (near doubling), near-term (present to next few decades) increases in Northern California and larger-magntitude (more than doubling), long-term (after mid-century) increases in Southern California (Swain et al., 2018). Increasing interannual precipitation variability is accompanied by decreasing interannual peak SWE variability (Table S1 in Supporting Information S1; Marshall et al., 2019 for WUS), which indicates that snowpack will not buffer increasing precipitation extremes. For (3), when more runoff arrives during the wettest times of year and reservoirs do not have space for this runoff—a condition that occurs in extreme or consecutive wet years—more of the year's surface water supply must be released beyond the reservoir when there is little demand for the water (United States Bureau of Reclamation, 2016). Despite this condition only occurring in some years, it is problematic because California already receives most of its precipitation in wet years (Dettinger et al., 2011) and this trend is expected to become more pronounced in the coming decades (Berg & Hall, 2015; Swain et al., 2018). As California receives a greater share of its precipitation in WY types that are likely to max out reservoir storage and snowpack does not buffer these changes, reservoir operators may be compelled to release more water at times when there is no demand for that water, nor sufficient downstream storage ready to capture these additional releases.

We find that southern Sierra Nevada snowmelt seasons will become progressively earlier and shorter, which indicates that southern Sierra snowpack will store less water into the dry times of year in the coming decades. Less water stored into the dry times of year is problematic because it contributes to dry fuel conditions and landscape vulnerability to disturbances such as severe wildfire, longer periods of low flow in rivers and streams which strains aquatic ecosystems, and provides less water for human consumptive use during the times of year when supplementation of reservoir storage is most critical. The snowpack peak date shifts 11 days earlier by mid-century ($${167}_{163}^{172}$$) and 20 days earlier by end-of-century ($${158}_{164}^{161}$$) (Figure 4). The peak date was previously projected to advance up to 1 month across the Sierra Nevada (Rhoades et al., 2018b). The melt date shifts 14 days earlier by mid-century ($${253}_{250}^{256}$$) and 33 days earlier by end-of-century ($${234}_{231}^{237}$$) (Figure 4). Peak runoff timing—the closest, commonly-used analog to melt date timing in the literature—is projected to advance by 2–4 months in the Sierra Nevada (Dettinger et al., 2018; Schwartz et al., 2017). The melt date shift is mirrored by a shift in the start date, which moves 5 days later by mid-century ($${66}_{63}^{70}$$) and 11 days later by end-of-century ($${72}_{69}^{77}$$) (Figure 4). The melt season is projected to become 7 days shorter by mid-century ($${86}_{83}^{90}\,\text{days}$$) and 18 days shorter by end-of-century ($${75}_{72}^{81}\,\text{days}$$) (Figure 4). Additionally, we find melt rate decreases as the century progresses (Figure 4). This is consistent with prior findings of slower snowmelt as warming progresses (Musselman, Clark, et al., 2017; Rhoades et al., 2018b). Water managers should plan for an earlier end to the snowmelt season than projected here because the WRF-GCM ensemble overestimates melt date (Figure 2). Additionally, WRF-GCM ensemble start date, peak date, and melt date interannual variability increases as the century progresses (Table S1 in Supporting Information S1; Marshall et al., 2019 for WUS). Increasing interannual variability in timing metrics points to a widening range of WY outcomes for managers to incorporate in planning.

Future period, emission scenario, and study area differences drive underestimation of SWE loss here compared to comparable literature. Siirila-Woodburn et al. (2021) provides a comprehensive review of 13 studies that project SWE loss in the Sierra Nevada by mid- and end-of-century. Across the studies included in this review, 55% of historical peak SWE is projected to remain by mid-century, which they define as 2050–74, and 35% by end-of-century, which they define as 2075–99. They also show that peak SWE losses of 20% (80% of historical remaining) are expected in the near future, which they define as 2025–2049. Regarding future period differences, our mid-century value underestimates SWE loss more than our end-of-century value. While our end-of-century period is fully contained within their end-of-century period, 10 years of our mid-century period is contained within their mid-century period, and 10 years is contained within their near-future period. Regarding emission scenario differences, the majority of studies included in Siirila-Woodburn et al., 2021 use RCP8.5. While both SSP3-7.0 and RCP8.5 are considered high-emissions scenarios, they are an imperfect match. Finally, our higher-elevation southern Sierra study area likely explains some SWE loss underestimation here compared to Siirila-Woodburn et al. (2021).

5 Directions for Future Research

Application of the SWE triangle methodology to WUS-D3 provides novel insights into intrannual snowpack processes in the southern Sierra Nevada, which are particularly important to water managers and informs the data set's ongoing use in California's Fifth Climate Change Assessment. For WRF-GCM skill at capturing intrannual snowpack processes, we found the ensemble is skillful at capturing peak SWE timing, but overestimates peak SWE and melt date. Melt rate remains difficult for models to capture and bias-correction does not resolve all cold bias at high elevations. For future changes in snowpack water storage magnitude and timing, the WRF-GCM ensemble projects diminished peak SWE and earlier, shorter snowmelt seasons. This indicates more water will run off of southern Sierra watersheds during the wettest times of year and southern Sierra snowpack will store less water into the driest times of year in the coming decades, which will have deep impacts across California's water sectors, from human consumptive use to ecosystem water needs.

We suggest three directions for future research. First, efforts to continue to improve melt rate skill of downscaled climate models and coupled land surface models would bolster the utility of these data for managers. Second, future research could use WUS-D3 to explore future interannual variability in snowpack water storage in greater detail. Future wet years have been explored with this data set (Marshall et al., 2024), but what about snow drought conditions? Finally, presenting results normalized by °C of warming will better enable planning with adaptation pathways, in which management changes are linked to environmental change thresholds (Grenier et al., 2024). By improving melt rate skill, further exploring hydroclimate extremes in the western US with this novel data set, and presenting results in an actionable format for decision-makers, we will be better prepared to adapt to the impacts of anthropogenic climate change on water resources in California.