1. INTRODUCTION
Snow is Earth’s most climatically sensitive land cover type and is continuing to decline as temperatures warm and precipitation patterns shift (IPCC et al. , 2019). Air temperature and moisture availability are first-order controls on snowfall. Maximum snowfall occurs at temperatures very near 0°C, so even a slight increase in temperature will shift a snowy winter to one with midseason rainfall and melt events (Jennings et al. , 2018). Snow cover variability affects multiple sectors, from water, energy, and forest management to transportation planning to outdoor recreation and tourism (Hagenstadet al. , 2018). Historically, snow monitoring has focused on ground-based measurements, with satellite remote sensing adding more capacity in the past several decades. While ground-based monitoring networks remain valuable, they do not have the spatial coverage, representative geospatial characterization, trend analysis, and predictive capabilities needed by researchers and water managers today (Milly et al. , 2008). Snow cover products derived from remote sensing require technical expertise and computing resources to produce and interpret, thereby limiting their broader usefulness. Here, we present a suite of snow metrics that may serve as key climate indicators to support the goals and needs of snow hydrologists, water managers, climate researchers, and various stakeholder groups.
The objectives of this research were to:
  1. Produce new snow metrics for assessing past and future changes in snow cover and snowpacks;
  2. Demonstrate how these snow-related climate indicators can be used to identify regions that may be vulnerable to future reductions in snow cover and warmer winters;
  3. Provide web-based visualization tools to communicate these snow metrics.
This suite of snow-related climate indicators is created to address sector concerns and stakeholder questions such as: Where and how have the extent/duration/timing of snow cover changed in the past, and how might it change in the future? How might snowpack storage change over the next several decades in a particular watershed? Are winter storm temperatures different from aggregated temperature data, and have winter storm temperatures in the western US increased in recent years? What has been the climatological frequency of warm winters for vulnerable economic sectors such as ski areas? How might the frequency of warm winters change in mid-century and late-century?
The specific snow metrics include Snow Cover Frequency (SCF), Snow Disappearance Date (SDD), snowstorm temperature (ST), At-Risk Snow (ARS), and Frequency of a Warm Winter (FWW). This suite of metrics aims to provide information to a wide range of researchers and stakeholders, including hydrologists, climatologists, water and land managers, snow sports enthusiasts, climate adaptation planners, and others for whom snow plays a role in decision making.

1.1 Previously developed snow metrics and their limitations

1.1.1 April 1 Snow Water Equivalent

The National Resources Conservation Service (NRCS) measures snow water equivalent along snow courses (survey transects) monthly and at automated SNOw TELemetry (SNOTEL) stations daily across the western United States. Snow water equivalent (SWE) is the amount of water represented by the snowpack (Serreze et al. , 1999). These NRCS sites were designed to serve as indices for seasonal streamflow forecasts based on regression-based relationships (Garen, 1992).
Although snow course and SNOTEL data were not intended as climate indicators, the data have been available for decades allowing researchers to incorporate SWE measurements into climate research. In particular, climatologists and hydrologists have used April 1 SWE as an annual measure of maximum snow accumulation and snow’s total contribution to the water budget of a watershed. However, of concern is that April 1 SWE cannot express the effects of midwinter melt events, which are now common in the maritime snowpacks of Washington, Oregon, and California (Mote, 2003; Sproles et al. , 2013, 2017; Knowles, 2015; Cooper et al. , 2016; Mote et al. , 2018). At lower elevation sites, maximum snow accumulation often occurs prior to April 1, thereby introducing a bias to watershed budget calculations (Montoyaet al. , 2014).
The spatial representativeness of SNOTEL and snow course sites is also a concern when used for purposes outside of streamflow forecasting. For logistical reasons, snow course and SNOTEL sites occupy a relatively limited elevation range; thus, high elevation snow and rain-snow transition zones are under-sampled (Molotch and Bales, 2006; Nolin, 2012; Gleason et al. , 2017). Even with over 800 measurement sites across the western United States (NRCS https://www.wcc.nrcs.usda.gov/about/mon_automate.html), the monitoring network is sparse, with many watersheds having only one or no sites.
Moreover, as the climate continues to warm, these sites, most of which were installed in the early 1980s, may become unrepresentative of watershed-scale SWE. This is because the snowline in the mountains increases during years of warm snow drought and dry snow drought (Cooperet al. , 2016; Sproles et al. , 2017). As such, they may underestimate trends in snow cover and changes in interannual variability across the seasonal snow zone. Seasonal drought outlooks that use April 1 SWE as a key indicator may miss key precipitation processes leading up to that date. For instance, anomalously low winter precipitation called a ’dry snow drought’ (Harpold et al. , 2017) is due to the natural variability of synoptic-scale atmospheric circulation and has regional impacts across all elevations. This is important in continental mountain regions where a shift in the storm track can lead to early and mid-winter dry conditions and low SWE.
In contrast, warmer than average winter temperatures that lead to a shift from snowfall to rainfall is termed a ’warm snow drought’ (Harpoldet al. , 2017) and may be caused by overall increases in winter storm temperatures. Warm snow droughts are most pronounced at low elevations and in maritime snow regions such as the Oregon and Washington Cascades and the California Sierra Nevada, where storm temperatures are close to the melting point. These examples demonstrate that different mechanisms drive warm and dry snow droughts, but there is no way to know this from April 1 SWE alone.

1.1.2 Snow Disappearance Date

A second previously developed snow metric is the Snow Disappearance Date (SDD). The presence or absence of snow affects landscape albedo, which in turn controls the energy balance. In snowy climates, SDD signifies the onset of spring and the start of the growing season. The elevational progression of snow disappearance in spring affects Arctic wildlife, with late spring, low elevation snow having a negative impact on populations of caribou and Dall sheep (Mahoney et al. , 2018; Boelman et al. , 2019). In the western United States, SDD is associated with wildfire activity (Westerling et al. , 2006a) in the sense that SDD reflects the end of snowmelt and the onset of seasonal declines in soil moisture and fine fuel moisture content. Lundquist et al. (2013) used SDD in their comparison of forested and open sites where they showed that, in locations with relatively warmer winters, forests lose their snow cover earlier than in open areas.
Snow disappearance date has been measured using ground-based and remotely-sensed measurements. In their study on spring snow disappearance, Foster et al. (1992) used satellite remote sensing and station-based data to map snow cover and snow disappearance for locations in Alaska, Canada, Scandinavia, and Siberia. In the studies using station data, the manner in which SDD was determined is not always clear. In his Arctic tundra studies, Foster (1992) defined SDD as the first day of the calendar year when station-based snow depth measurements dropped below 1” (2.5cm). In their 2013 meta-analysis, Lundquist et al. (2013) listed SDD as the first day of the calendar year with no snow as reported from station data (though they didn’t provide a measurement threshold). Lundquist and Lott (2010) used near-surface soil temperature measurements to detect the presence and absence of snow cover, which can be used to identify the SDD.
Such varied uses and ways of recording SDD indicate both the importance of this snow metric and a need for a spatially consistent approach to its measurement. Station-based data are inherently limited in their spatial representation of SDD, though they are critical for calibration and validation of remote sensing measurements of SDD. Temporal factors affecting SDD are also essential to consider. Transient snowfall events can influence SDD detection and can be important for hydrology and wildlife. The variable nature of spring meteorology means that it is not uncommon for spring melt to be followed by spring snowstorms that can drive the actual snow disappearance date to be days or even weeks later.

1.1.3 Snow cover Absence and Snow cover Persistence

Several researchers have used remotely sensed snow cover to create metrics such as snow cover absence, SA (Wayand et al. , 2018), snow persistence, SP, and snow season, SS (Hammond et al., 2018). Wayand et al. (2018) developed SA using high-resolution remote sensing data from Landsat-8 and Sentinel-2 to map snow-free areas during winter as a way to detect snow removal by wind erosion and avalanches in mountainous areas. The authors also created a snow cover persistence index that mapped snow-covered areas during summer as a way to identify wind- and avalanche deposited snow. Hammond et al. (2018) developed a remote sensing-based snow cover persistence metric though their definition of SP is ”the fraction of time that snow is present on the ground.” The advantage of SP, SS, and SA is that they can be computed using remote sensing data and produced at 10-30 m spatial scales, thus capturing high-resolution snow processes such as the effects of redistribution and boundaries between intermittent and seasonal snow zones (Moore et al. , 2015). However, to date, these metrics have not been produced daily or globally. There is a trade-off between temporal and spatial resolution satellite data. Finer spatial scale data such from Landsat 8 have a 16-day revisit time so even with two satellites there are long gaps in coverage. Such gaps can miss important events such as snowfall and snowmelt events.

1.1.4 | Climatologically-based Snow Metrics

At-risk Snow
First defined by Nolin and Daly (2006), “At-risk” snow (ARS) is when snowfall is at risk of turning to rainfall under climate warming conditions. Nolin and Daly used the PRISM 4-km gridded climate data (Daly et al. , 1993) and a decision tree approach to classify grid cells as either at-risk or not at-risk, based on monthly mean air temperature for the months of December, January, and February. The study’s geographic scope covered only the Pacific Northwest (Washington, Oregon, Idaho, and western Montana), where snowfall is often close to the melting point. Their approach used the 0°C monthly mean temperature threshold to partition between rain and snow. The monthly mean 0°C threshold was selected because, though spatial and temporal differences exist, both thermodynamically and practically, it represents a shift from snowfall and snow accumulation to rainfall and declining snowpacks. Their methodology assumed that locations that might warm to that 0°C monthly mean temperature threshold would be at-risk of converting from snowfall to rainfall. For instance, a grid cell with a climatological mean monthly temperature of -2°C for any one of the core winter months (Dec-Feb) was classified as ARS for a +2°C warming scenario. This data-driven approach was simple but did not consider possible changes in atmospheric circulation and storm patterns. Moreover, the 4-km grid-scale was too coarse to address changes at more local scales, especially in mountain regions where a 4-km grid cell can span a wide range of elevations and temperatures.
Frequency of a Warm Winter
Nolin and Daly (2006) also developed the Frequency of a Warm Winter (FWW) metric and as with ARS, Nolin and Daly used the PRISM gridded climate data product and monthly mean temperatures for December, January, and February. FWW is computed as the number of winters out of 30 consecutive winters that meet the ”warm winter” criterion. A warm winter is defined as one in which the mean monthly temperature exceeds 0°C during any one of the three core winter months (December-February). Thus if three winters out of 30 winters are classified as warm in a given grid cell, the frequency is 0.1 or 10%. To demonstrate possible impacts of changes in FWW, Nolin and Daly (2006) tabulated the historical and future FWW values for selected ski areas in the study region. All ski areas in the region showed increases in FWW for incremental temperature increases, but the 4-km spatial resolution introduced uncertainty due to the range of elevations in a grid cell.