Monitoring the Daily Evolution and Extent of Snow Drought

Snow droughts are commonly defined as below average snowpack at a point in time, typically 1 April in the western United States (wUS). This definition is valuable for interpreting the state of the snowpack but obscures the temporal evolution of snow drought. Borrowing from dynamical systems theory, we applied phase diagrams to visually examine the evolution of snow water equivalent (SWE) and accumulated precipitation conditions in maritime, intermountain, and continental snow climates in the wUS using station observations as well as spatially distributed estimates of SWE and precipitation. Using 5 a percentile-based drought definition phase diagrams of daily observed SWE and precipitation highlighted decision-relevant aspects of snow drought such as onset, evolution, and termination. The phase diagram approach can be used in tandem with spatially distributed estimates of daily SWE and precipitation to reveal variability in snow drought type and extent. When combined streamflow or other data, phase diagrams and spatial estimates of snow drought conditions can help inform drought monitoring and early warning and help link snow drought type and evolution impacts on ecosystems, water resources, and 10 recreation. A web tool is introduced allowing users to create real-time or historic snow drought phase diagrams.

diagram trajectories during the 21 st century. During the melt season, persistent cold and dry conditions can drive trajectories 125 upwards into the first or second quadrants as snow melts slower than expected historically, as occurred during May 2020. Dry snow drought conditions (meteorological drought) are identified in the lower left (third quadrant) when SWE falls into the D0-D4 range (i.e., less or equal to the 30 th percentile) and accumulated precipitation is below the median. We defined warm snow drought when satisfies snow drought conditions and accumulated precipitation is greater than the median (lower right, or fourth quadrant). To facilitate connecting various trajectories of phase diagrams with driving processes, the annotated 130 figure is paired with a conceptual diagram showing potential physical interpretations (Figure 2b).
The start of WY2020 was characterized by bottom 3 rd percentile precipitation conditions with low (bottom 20 th percentile) snowpack at the CSS Lab ( Figure 2a). Precipitation falling as snow led to rapid improvement from snow drought into the "Dry But Snowy" quadrant during late November into December, with precipitation recovering to near-normal by mid-December.
Persistent dry conditions lasting from late December through mid-March, driven by a blocking ridge west of North America 135 (Gibson et al., 2020), yielded snow drought onset in late January. Above-normal temperatures, dry conditions, and seasonallyinduced shifts in solar insolation in late February and early March caused snowpack declines to accelerate, reaching a minimum value in the 5 th percentile. Given that California receives the majority of its annual precipitation between December and March, dry spells will quickly lead to declines in precipitation percentile (trajectories move leftward; Figure 2a). WY2020, like other abrupt upwards and/or rightwards trajectories in the phase diagram via heavy precipitation (Guan et al., 2010) enhanced by orography (Huning et al., 2017). Snow drought amelioration in late March occurred when heavy snowfall resulted from a slowmoving cutoff low pressure system (O'Hara et al., 2009). By 1 April, the historically assumed peak timing of snowpack in the wUS (e.g., Huning and AghaKouchak (2020b)), snow drought conditions remained but had improved from the 5 th to nearly the 30 th percentile, though precipitation remained in the bottom 15 th percentile. Another cutoff low in early April provided 145 additional snow that terminated snow drought conditions, however accumulated precipitation remained below the median. This further highlights the importance of late spring (i.e., post-1 April) meteorological events in improving hydroclimatic conditions and a potential pitfall of assessing drought conditions at a single point in time. The remainder of April and May were drier-thannormal, but snowmelt occurred slower than climatology, with above-median snowpack observed in mid-May. By annotating the phase diagram, the story of the cool season can be expressed to show the key events producing observed outcomes.

Snow Drought Variation in Time and Space
Weather events drive elevation-dependent changes in snowpack and snow drought conditions (Hatchett and McEvoy, 2018). In regions located near climatological expected rain-snow transition elevations (Jennings et al., 2018), such as the Sierra Nevada, individual storms can produce dramatically different responses in snowpack spatial variability and magnitude. ARs are a common type of storm event yielding variable snowpack and hydrologic responses as a result of heavy precipitation with high 155 snow line elevations (Hatchett et al., 2017;Hatchett, 2018;Henn et al., 2020) or with snow line elevations that vary widely over the duration of the storm (Lundquist et al., 2008;Hatchett et al., 2020).
WY2018 was emblematic of the aforementioned variation in rain and snow transition elevations as both elevation-and spatially-dependent responses to storms and dry spells occurred in the Sierra Nevada ( Figure 3). WY2018 began with varying precipitation and SWE percentiles between three stations, again in the "Dry But Snowy" quadrant at the lower elevation stations 160 (CSS Lab and Tahoe City Cross) and near climatology for the high elevation station (Mount Rose Ski Area). A late November AR event was followed by a multi-month dry spell that terminated in late February. Snowpack and precipitation conditions improved markedly in March, (colloquially termed a "Miracle March"), due to persistent stormy conditions associated with multiple landfalling ARs and/or midlatitude cyclones.
To highlight the elevation-induced heterogeneity of snowpack response within WY2018, we investigate three different sta-165 tions situated along a similar longitude (Figure 1c). The late November warm and wet storm caused the CSS Lab and Tahoe City Cross (Figure 3a-b); both maritime snow climates) to shift rightwards and then downwards into the warm snow drought quadrant because much of the precipitation fell as rain. The CSS Lab is located along the Sierra Nevada crest while Tahoe City Cross is located further east in the rain shadow of the Sierra Nevada crest. Unlike the other two stations, the higher elevation Mount Rose Ski Area (hereafter "Mount Rose"), located further east in the Carson Range in a more intermountain snow cli-  City Cross from the 2 nd percentile to above the 40 th percentile while precipitation also improved from the 26 th to the 60 th percentile between late February and early April ( Figure 3b). As a result of "Miracle March", 1 April SWE conditions were closer to median than reflected by the majority of the winter, similar to WY2020 (Figure 2a). The record to near-record low, late winter SWE at the lower elevation CSS Lab and Tahoe City Cross are thus hidden by a single point-in-time perspective.
WY2018 and WY2020 demonstrate the importance of a complete WY perspective regarding the assessment of evolving snow 190 drought conditions, namely the importance of a few large precipitation events.

Warm Versus Dry Snow Drought: Implications for Runoff Timing
The warming-induced shift in precipitation phase from snow to rain has been shown in historical trends in the wUS (Lynn et al., 2020) and is projected to continue in a warmer world (Klos et al., 2014;Rhoades et al., 2018c;Musselman et al., 2018). Precipitation phase transition from snow to rain will result in more frequent warm snow droughts (Marshall et al.,195 2019; Huning and AghaKouchak, 2020a). This increase will disproportionately impact climatologically warmer maritime  Figure 4(a-b) shows the entire WY phase diagrams and a SWE spatial extent snapshot at various times (d-k), relative to median climatology, for WY2001 and WY2015. We also highlight the differences in hydrologic outcomes between these dry and warm snow drought years (Figure 4c). WY2001 had the second lowest cumulative flows for the Nisqually River in the 215 period studied (WY1943-2019), but 50% of the cumulative WY2001 flow occurred 30 days later than the median date (3 April) at which half the Nisqually flow occurs. In contrast, WY2015 demonstrated middle-of-the-range total WY flow (48 th of 77 years) but achieved 50% of the WY flow 56 days earlier than average. This indicates a key difference between the WYs. WY2015 had less snow water stored later into the season, which markedly influenced streamflow. During both seasons, despite the vastly different precipitation regimes, spatial SWE anomalies are not markedly different during mid-December ( Figures   220   4d and 4h), mid-January (Figures 4e and 4i), or late February (Figures 4f and 4j). Consistent with lower SWE percentiles at Paradise during WY2015 compared to WY2001, as shown on the phase diagrams, SWE anomalies are modestly more negative.
The lack of mountain snowpack during WY2015 was more notable than WY2001 (Figures 3g and 4k). The comparatively better spring snowpack in WY2001 likely helped maintain flows later into the year despite an otherwise dry year.
SNOTELs increased in SWE percentile through May while the basin-wide percentiles declined. This result may stem from the inclusion of lower elevation terrain in the basin aggregations whose snow rapidly melts out during drought years (potentially accelerated by snow-albedo feedbacks, e.g., Groisman et al. (1994)). Despite these differences, the basin-aggregated phase diagrams appear reasonably representative in capturing the broader hydroclimate conditions interpreted from phase trajectories.

270
We applied SNOTEL station data to create the phase diagrams, but a challenge in mountain environments is the lack of reliable, well-distributed, long-term observations. In lieu of station data, gridded observational products commonly inform natural resource decision-making and research efforts. The necessary components exist to create phase diagrams using gridded meteorological products (Daly et al., 2008;Abatzoglou, 2013), observation-based snow datasets (Zeng et al., 2018;Margulis et al., 2016), and output from hydrological simulations (Livneh et al., 2015). The challenge is how to aggregate spatial information 275 to become meaningful in complex terrain. We have performed a first step towards this goal. Initial methods to broaden the approach could be performed by: (1) binning regions by similar elevation, watershed, slope, aspect, and/or land cover type; (2) identifying areas that co-vary together in time and space using techniques such as principal component or cluster analysis; and (3) subjective grouping based on anecdotal information from managers. Creating meaningful phase diagrams using spatially distributed information is the primary goal of our ongoing research. This will allow evaluation of snow drought in regions 280 without long-term snow-observing networks such as in the northeastern U.S. or other high mountain areas worldwide. Towards this end, we next provide examples of how spatially distributed products can visualize snow drought.

Snapshots from Water Year 2015
Using WY2015 as an example, gridded SWE and precipitation allow visualizing the spatial extent and type of snow drought.  outcomes (e.g., runoff and spatial patterns of snowpack anomalies) to be explained with more nuance. Such explanation is important as similar end-of-season SWE anomalies in space (compare Figures 4g, 4k, and 7d) demonstrate markedly different hydrologic outcomes (Figure 4c). Last, Figure 7e highlights an example of the differing elevational response of snow drought in the Sierra Nevada (percentiles increase with increasing elevation) and the Uinta Mountains (percentiles decrease with increasing elevation) for the same midwinter time. This example shows how sub-seasonal snowpack heterogeneity could create differing melt-season responses (i.e., earlier snow loss at lower elevations with increasing radiation and springtime warming) or ongoing avalanche hazards (i.e., higher elevation snowpacks are more prone to weakening when shallow). In all cases, interannual variability ranges from less than 10% to more than 60% in December ( Figure 8a) and March ( Figure   8b). The maximum amount of area in snow drought on 1 April is less than 50%. The WY2015 snow drought stands out (  D0 snow drought onset in early December (Figure 10a and (yellow arrow in Figure 10b)). Note snow drought onset occurs, perhaps non-intuitively, at approximately 85% of median snowpack. This indicates limited variability in SWE at this time and station: large deviations from the median value are relatively infrequent. The late November and early December dry spells led to snowpack accumulation falling behind the climatological average (Figure 10b). While some accumulation occurred during 385 mid-December into early January, the rate of accumulation was less than climatology (Figure 10b), leading to a continued decline into D1 snow drought (Figure 10a). Percent of median SWE hovered around 65% leading up to the avalanche ( Figure   10b), though SWE percentiles fell into the D2 category. A transition to more active weather in late January into February followed with gains in SWE that mirrored climatology with little change in SWE percentile or percent of median (Figure 10b).
The presence of a shallow snowpack during the dry, low radiation periods in November and December promoted the formation 390 of a persistent weak layer with striated, 3-6 mm faceted grains buried 90 cm deep in the snowpack. According to the UAC, this is the layer where failure occurred on 11 February.
The snowpack conditions leading up to the Wilson Glade avalanche show the potential disconnect between percent of median and percentile. Prior to the loading events, the percent of median values (65%) between December and early February do not directly convey the infrequency of these values as percentiles can. Percentiles show that such conditions occurred only 10-395 20% of the time. The user's familiarity with a location will govern the meaningfulness of percent of medians through prior experience. On the other hand, percentiles provide perspective for less-familiar users to understand the distribution and state of the snowpack. Percentiles also allow comparisons between locations in terms of snow drought severity. By recognizing both as valuable, the option to view either on the snow drought tracker webtool is a planned improvement. Last, we recommend incorporation of percentiles into accident write-ups, such as provided by the UAC, to give this additional statistical perspective.

Conclusions
Our primary goal was to demonstrate a visualization approach to show the temporal evolution of snow drought conditions, and more broadly mountain hydroclimatic conditions, through the cool season. When annotated, phase diagrams help "tell the story" of a snow season and can help communicate the weather and climate events that shaped the outcome of peak snowpack and lifecycle of the snowpack. We provided examples showing a range of applications in various snow climates for extreme 405 years and how additional data such as spatially distributed SWE and precipitation as well as river discharge can further enhance the utility of information provided by phase diagrams. The spatial snow drought maps and basin-aggregated phase diagrams generated using gridded data products demonstrate an approach evaluating snow drought patterns across the landscape or in sparsely observed regions.
Our approach can be extended beyond addressing the noted limitations. While our primary purpose was to show the evolution 410 of conditions in the current year, phase diagrams are easily produced for all previous years to allow comparisons of trajectories at seasonal or monthly timescales. These diagrams can incorporate forecasts of precipitation and SWE to show how snow drought conditions may evolve at subseasonal-to-seasonal timescales. For example, inclusion of bias corrected ensembles of medium range to subseasonal forecasts of precipitation and SWE from various forecasting center model(s) can create