Rain-on-snow events in Alaska, their frequency and distribution from satellite observations

Wet snow and the icing events that frequently follow wintertime rain-on-snow (ROS) affect high latitude ecosystems at multiple spatial and temporal scales, including hydrology, carbon cycle, wildlife, and human development. However, the distribution of ROS events and their response to climatic changes are uncertain. In this study, we quantified ROS spatiotemporal variability across Alaska during the cold season (November to March) and clarified the influence of precipitation and temperature variations on these patterns. A satellite-based daily ROS geospatial classification was derived for the region by combining remote sensing information from overlapping MODIS and AMSR sensor records. The ROS record extended over the recent satellite record (water years 2003–2011 and 2013–2016) and was derived at a daily time step and 6 km grid, benefiting from finer (500 m) resolution MODIS snow cover observations and coarser (12.5 km) AMSR microwave brightness temperature-based freeze–thaw retrievals. The classification showed favorable ROS detection accuracy (75%–100%) against in situ climate observations across Alaska. Pixel-wise correlation analysis was used to clarify relationships between the ROS patterns and underlying physiography and climatic influences. Our findings indicate that cold season ROS events are most common during autumn and spring months along the maritime Bering Sea coast and boreal interior regions, but are infrequent on the colder arctic North Slope. The frequency and extent of ROS events coincided with warm temperature anomalies (p < 0.1), but showed a generally weaker relationship with precipitation. The weaker precipitation relationship was attributed to several factors, including large uncertainty in cold season precipitation measurements, and the important contribution of humidity and turbulent energy transfer in driving snowmelt and icing events independent of rainfall. Our results suggest that as high latitude temperatures increase, wet snow and ROS events will also increase in frequency and extent, particularly in the southwestern and interior regions of Alaska.


Introduction
The atmospheric conditions typically associated with high latitude winter rainfall affect the physical properties of the snowpack, including energy content, water content, depth, density and grain size, frequently resulting in a wet snow surface (Singh et al 1997). These effects are due to the associated transfers of latent and sensible heat, either directly or through turbu-lent exchanges that hasten snow melt (Marks et al 1998). Whenever snow surface layers reach 0 • C, additional energy flux to the snow surface contributes to melt and rising water content; snowmelt will continue to occur whenever the cold content of the snowpack exceeds 0 • C or until the snow has completely melted (Dingman 2015). Thus, wintertime rain events can be a major driver of wet surface snow conditions indicated from satellite observations (Frei et al 2012).
However, rain is not required for wet snow to exist; nor do wet snow conditions always follow rainfall events. Different science communities have used the term 'rain-on-snow' to collectively refer to wet surface snow conditions and the many physical processes that lead to their occurrence. We recognize that rain-onsnow is not necessarily synonymous with wet snow, but that the occurrence of rain on a winter snowpack frequently precedes the presence of wet snow conditions at high latitudes. We therefore retain the usage of the term rain-on-snow (ROS) in this investigation to collectively describe these processes.
Wet snow, and the icing events that frequently follow ROS, affect several ecosystem processes including hydrology, carbon cycling, wildlife movement and human transportation, at multiple spatial and temporal scales Roe 2003, McCabe et al 2007). ROS events, and the positive heat flux to the snowpack often associated with them, are one of the dominant drivers of winter and springtime flooding in mountainous regions and at higher latitudes (Marks et al 1998, Guan et al 2016, Jeong and Shushama 2018. Enhanced liquid water content (LWC) to the snowpack, whether by ROS or melt events, can also reduce a snowpack's insulating effect on the soil (Lafrenière et al 2013, Kim et al 2015. Furthermore, accumulated water at the soil surface from ROS-driven snowmelt can release latent heat into the soil horizon, and in turn result in accelerated thawing of frozen ground Roe 2003, Rennert et al 2009). These thawing processes ultimately hasten the release of soil carbon to the watershed and atmosphere in the form of dissolved organic matter or gasses (Hobbie et al 2000). Further, accumulated water between the soil surface and snowpack also has the potential to freeze, forming a significant ice barrier to browsing ungulates, which can contribute to large wintertime die-offs (Grenfell and Putkonen 2008, Riseth et al 2011, Loe et al 2016, Berger et al 2018. As intensified warming of the high latitudes, known as 'Arctic Amplification', continues Francis 2006, Cohen et al 2014), an increase in the frequency, distribution, and intensity of ROS events is predicted (Jeong and Shushama 2018), with potentially adverse impacts to Arctic ecosystems and the communities that depend on them.
The Arctic Boreal Vulnerability Experiment (ABoVE) is a broad-scale international and interdisciplinary field campaign initiated by NASA to understand environmental change in the Arctic and boreal region (ABR) of western North America and associated linkages to social-ecological systems (Kasischke et al 2014). The science objectives of ABoVE include quantifying changes in the condition and distribution of snow, and its impact on ecosystem structure and function. A key limitation to the quantifying and understanding of ROS in the region is a general lack of available observations, which are constrained by its remoteness, severe climate, and sparse regional weather station networks. However, satellite remote sensing methods have been developed for detecting and mapping ROS in the ABR. Successful approaches include the use of active and passive microwave sensors from polar-orbiting satellites that provide frequent observations and enhanced sensitivity to landscape freeze-thaw dynamics (Kimball et al 2004, Bartsch 2010b, Semmens et al 2013, Wilson et al 2013. However, these approaches have generally involved only limited areas and periods, or have relatively coarse (∼10-25 km resolution) retrievals.
The objectives of this study were to quantify spatiotemporal variability in ROS across Alaska during the winter season (November-March) and to clarify the influence of precipitation and temperature anomalies on ROS frequency and distribution. The domain for this study is the state of Alaska, which has a long snow season and faces challenges to both natural resources management and socio-economic structure, due to changing snow conditions caused by regional warming trends (Bokhorst et al 2016, Kontar et al 2018. Much of Alaska is in the ABoVE domain, where a better understanding of the distribution and underlying drivers of ROS will contribute to the ABoVE science objectives and provide critical information to Alaskan land managers. To address the study objectives, we generated a daily ROS geospatial classification across Alaska by combining synergistic remote sensing information from overlapping MODIS (moderate resolution imaging spectroradiometer) and AMSR (advanced microwave scanning radiometer) sensors. Here, MODIS provides eight-day repeat coverage and relatively fine scale information (500 m resolution) on snow cover extent, while AMSR provides daily microwave brightness temperature (T ) retrievals sensitive to landscape freeze-thaw dynamics, but within a relatively coarse (∼12.5 km) sensor footprint. The combined information from these sensors provides a means for ROS mapping, with enhanced gridding (∼6 km resolution) suitable for resolving regional ROS patterns and underlying physiographic and climate drivers.
The application of satellite remote sensing to detect ROS events has progressed in recent years through the development of new data sources and techniques, including both radar (Kimball et al 2004, Bartsch 2010b, Bartsch et al 2010a, and passive microwave (PM) sensors (Grenfell and Putkonen 2008, Wang et al 2013, Wang et al 2016. In Alaska, these sensors have been applied to detect ROS using different classification algorithms, including backscatter change detection (Kimball et al 2001, Bartsch 2010b, Wilson et al 2013, diurnal amplitude variation from PM T retrievals (Semmens et al 2013), and a T differencing approach (Wang et al 2016). More recently, spectral gradient ratios, including the Gradient Ratio (GR, Grenfell and Putkonen 2008)   from different microwave frequencies and polarizations for ROS detection. The PM-based GRP approach was also observed to be effective in detecting ROS and associated winter melt events within the ABR (Dolant et al 2016, Langlois et al 2017. However, to our knowledge, this study provides the only available ROS satellite record for Alaska that provides 6 km daily resolution from current operational satellites that overlaps with the timing of the ABoVE campaign. In this study, we used the GRP approach with PM observations from the Advanced Microwave Scanning Radiometer sensors AMSR-E and AMSR2 (hereafter denoted as AMSR) for daily classification of ROS events across Alaska. The AMSR GRP-based ROS classification was conducted over snow-covered areas defined from MODIS. The study period for this investigation encompassed water years (WYs) 2003-2016 and the available AMSR record, which overlapped with the first phase of the ABoVE campaign (Kasischke et al 2014). The daily ROS record encompassed the months of November through March, when snowmelt from solar irradiance is minimal and snow cover is widespread and relatively consistent throughout the region (Lindsay et al 2015). The ROS classification was mapped to a 6 km resolution grid and used to quantify and understand ROS spatiotemporal variability and underlying drivers across Alaska.

Spatial domain
The state of Alaska spans approximately 20 • of latitude and 50 • of longitude, encompassing the North Pacific and Arctic Boreal regions of the northern hemisphere. Within the region, many gradients influence the climate, including: latitude, distance from large water bodies, the relative thermal mass and circulation of coastal waters, terrain, and elevation. Alaska is a peninsula with over 10 000 km of coastline, bounded by the Pacific Ocean to the south and the shallower, seasonally ice-covered Bering, Chukchi and Beaufort seas to the west and north. The eastern border of Alaska runs through boreal forest characterized by a cold interior continental climate. Thirteen different climate divisions have been described for the state of Alaska (Bieniek et al 2012). For the spatial analysis of ROS distributions, we aggregated the 13 climate divisions into four larger regions delineated by National Hydrography Hydrologic Unit Code (HUC 8) watersheds (USGS 2017) (figure 1). The aggregated Alaska HUC 8 divisions examined for this study include the Alaska Gulf Coast (AGC), Interior (INT), Bering Sea Coast (BSC) and North Slope (NS). These areas reflect Alaska's major climatic regions, of the relatively moderate Pacific maritime, cold-dry boreal interior, and polar arctic northwest coast and North Slope regions. The high latitude ecosystems found in these climate divisions play an important role in the Earth's energy, water and carbon cycles, and are some of the most vulnerable to recent climate warming (Chapin et al 2014, O'Neel et al 2015.

Satellite data used for ROS classification
The AMSR-E sensor was launched in 2002 on board the NASA Aqua satellite, and operated until 2011 (Kawanishi et al 2003). The AMSR2 follow-on mission was successfully launched in 2012 on board the JAXA GCOM-W satellite and continues normal operations (Imaoka et al 2010, Du et al 2014. We used combined calibrated T records from AMSR-E for WY 2003-2011 and AMSR2 for WY 2013-2016. The AMSR record was derived using an empirical calibration of similar frequency T retrievals from overlapping FY3B Microwave Radiation Imager (MWRI) observations (Du et al 2014). The AMSR record has twice-daily, vertical (V) and horizontal (H) polarization T retrievals acquired from ascending and descending polar orbital equatorial crossings at 1:30 pm and 1:30 am, which is suitable for detecting ROS (Dolant et al 2016, Du et al 2016. Lower-frequency T retrievals (18.7 GHz and 36.5 GHz, henceforth rounded to 19 and 37 GHz) from the AMSR record were used for ROS detection in this study, as they are sensitive to snow cover properties and landscape freeze-thaw dynamics  but insensitive to potential signal degradation from polar darkness, low solar illumination, cloud cover, and atmospheric aerosol contamination effects (Rees et al 2010, Tedesco et al 2015. The native AMSR T footprints are relatively coarse at 19 GHz (27 km × 16 km for AMSR-E and 22 km × 14 km for AMSR2) and at 37 GHz (14 km × 8 km and 12 km × 7 km) due to naturally low PM earth emissions (Kawanishi et al 2003, Imaoka et al 2010, Frei et al 2012. In this study, we used spatially resampled ascending orbit T retrievals from the calibrated AMSR record in conjunction with MOD10A2 eight-day maximum snow cover extent (SCE) derived from MODIS (Hall et al 2002, Hall andRiggs 2007).

Spatially resampled AMSR
The AMSR orbital swath T data were spatially resampled to a 6 km resolution polar EASE-Grid (version 2) geographic projection, using an inverse distance squared weighting method (Brodzik et al 2012, Du et al 2017a. To ensure cross-sensor consistency, the gridded AMSR2 T data were empirically calibrated against the same AMSR-E frequencies using a double-differencing method and similar overlapping observations from the FY3B MWRI sensor record (Du et al 2017b, Du et al 2014. The new 6 km grid provided an intermediate resolution between the finer scale (500 m) MODIS SCE and the coarser resolution (∼12.5 km) AMSR T observations, while enabling enhanced assessment of terrain and land cover spatial heterogeneity.

Theoretical approach to the ROS classification
We operationally defined ROS days as the satellite PM detection of abrupt changes in surface snow wetness and isothermal states induced by physical processes, such as sensible, latent and turbulent heat exchange that are often associated with winter rainfall. The physical basis of the PM ROS algorithm is the differential response in microwave emissions at 19 (V, H) GHz and 37 (V, H) GHz frequencies to changes in snow cover density and LWC within the snowpack surface. As relatively dry snow initially transitions to wet snow with increasing LWC, T increases due to absorption by wet snow (Tedesco et al 2015). Yet the interaction between T and snow wetness varies over different regions of the microwave spectrum. T at 19 (V and H) GHz will change with LWC to a lesser degree than at 37 (V and H) GHz (Rees et al 2010, Vuyovich et al 2017. Grenfell and Putkonen (2008) found distinct patterns in dielectric properties at 19 and 37 GHz in response to ROS events, leading to their application of a spectral GR that portrays larger differences between V and H polarized (pol) T retrievals at these frequencies following ROS events equation (1): Dolant et al (2016) found that during ROS events, the GR derived from H pol T (GR-h) returned negative values, while the GR derived from V pol T (GR-v) returned positive values. This inverse relationship led to the development of the gradient ratio polarization (GRP) between GR-v and GR-h, allowing for the ability to set designated thresholds to classify ROS events. Dolant et al (2016) and Langlois et al (2017) applied the GRP equation (2) to single-pixel T time series from SMMR, SSM/I, and AMSR-E to detect ROS in areas of Quebec and the Canadian Arctic Archipelago. (2)

ROS workflow
In the current study, we applied a similar GRP approach developed from previous studies at point locations (Grenfell and Putkonen 2008, Dolant et al 2016, Langlois et al 2017 for mapping daily ROS patterns across Alaska. The Alaska regional classification was derived using daily ascending V and H pol T retrievals at 19 and 37 GHz from the 6 km resolution polar EASE-grid AMSR record. The created workflow is summarized in figure S1 available at stacks.iop. org/ERL/13/075004/mmedia and described below. We masked water-contaminated pixels induced by the conical scanning of AMSR sensor records , Du et al 2016 using a 24 km (∼4 pixel) shoreline and water-body buffer created from the 2011, 30 m resolution National Land Cover Database (Homer et al 2015). We then used the MODIS SCE record to identify snow-covered areas after screening out low-quality pixels, including missing or degraded snow cover observations, identified by the MOD10A2 product quality flags. The AMSR 37 GHz V pol T record was analyzed separately, to identify potential snow-covered pixels outside the water body buffer, where T < 265 K (Vuyovich et al 2017). Pixels identified as being snow-covered  year (2003-2011, 2013-2016) study period defined by the AMSR record. We applied two different GRP thresholds to classify ROS events for different elevation zones: GRP < 1 was used to identify ROS events below 900 m, while GRP <−5 was used for elevations above 900 m; a more detailed description of the GRP threshold selection is given in the supplementary section (S1). A spatial connectivity threshold of > 10 pixels was then used as a designated size threshold to isolate and analyze more regionally extensive ROS events (Wilson et al 2013).

Two-tiered validation 2.6.1. Tier 1-empirical in-situ ROS observations
The Tier-1 validation coupled empirical ROS observations by an observer at the National Weather Service (NWS) field office in Fairbanks, Alaska (table 1), with in situ weather station measurements acquired from Fairbanks International Airport (MesoWest ID-PAFA, 134 m above sea level). To determine the agreement between the in situ observations and the satellite PM-derived ROS classification, daily mean GRP values were created from 6 km pixels located within a 50 km radius around the Fairbanks station location. Next, the occurrence of the daily mean GRP values < 1 were examined in conjunction with: (1) ROS events empirically observed by the NWS observer; (2) precipitation and fog observations (Wang et al 2016) at the station, and; (3) measured precipitation the day before or the day when GRP was < 1.

Tier 2-climate observation network
The Tier 2 validation involved an expanded spatial domain employing climate observations across Alaska acquired through the MesoWest and SynopticLabs API (https://synopticlabs.org/api/). The API provided data from several regional weather station networks, including the NWS, remote automated weather stations, and snow telemetry network stations. The climate data were assembled from 235 individual stations to acquire daily surface meteorological parameters, including minimum and maximum air temperatures (T min , T max ), average (24-hour) air temperature (T avg ), 24-hour accumulated precipitation (prcp), dew point temperature (T dew ), and relative humidity (RH). For the study period, 53 of the 235 stations had all of the requested climate variables. The developed Tier 2 workflow is shown in figure S3 and described below. ROS days classified from the satellite record were validated using in situ weather observations from collocated climate stations to identify if rain occurred either the day of or the day before (i−1) a classified ROS event (Obs rain ), or if the observed station precipitation was null (Obs null ). Given the limitations of wintertime precipitation measurements (Merenti-Valimaki 2001, Martinaitis et al 2015, Grossi et al 2017, Obs null included conditions where either no precipitation was measured or there was no effective precipitation measurement. Three temperature-driven variables were therefore created and used as a proxy for the rainfall observations (Obs rain ), which can have large measurement uncertainty during freeze-thaw transitions (Martinaitis et al 2015). The temperature-based Obs rain metrics included wet bulb temperature (T w ), the ratio between daily T dew and T avg (T dew /T avg ), and the ratio between daily T max and T min (T max /T min ), which were used as indicators for atmospheric moisture and energy flux to surface snow. A more detailed description of the temperature-based precipitation metrics can be found in the supplementary section (S2). We then constrained Obs null by the mean and standard deviations derived from temperature-based Obs rain metrics. ROS days which met all the constraining conditions set by Obs rain were classified as commission, whereas all Obs null observations that did not meet the defined conditions were classified as omission.

Statistical methods and climate anomalies
Due to the relatively short study period and a data gap between the AMSR-E and AMSR2 records in 2012, we did not attempt to perform a temporal trend analysis of the ROS results. However, we did calculate the mean, standard deviation and coefficient of variation (C ) of monthly total ROS days. The C was used to characterize the relative dispersion of ROS days (Sugg et al 2017), with higher values equating to high variability and low predictability, with the inverse being true as the C approaches zero (Frei et al 2012).
Pixel-wise correlations were performed to determine the sign and strength of relationships between monthly total ROS days and respective climate anomalies. Climate anomalies were derived using 1 km resolution gridded daily surface weather station observations from the North America Daymet record (Thornton et al 1997). Daymet was one of the few products available for Alaska with a spatial (1 km 2 ) and temporal resolution similar to the AMSR-derived ROS record. Daymet daily T min , T max , and prcp for the period from 1980 to 2016 were acquired from the DOE ORNL data portal (www.daymet.ornl.gov/dataaccess.html). The 1 km Daymet data were resampled to the 6 km study grid and used to create a baseline monthly mean climatology and standard deviation for each pixel and climate parameter. Here, the baseline climatology was assembled from a 23-year Daymet historical record . Monthly mean annual values of each climate parameter were then derived for each year of record from 2003 to 2016 (excluding 2012) and used with the respective climatology for each pixel to create gridded normalized anomalies (figure 2).

Tier-1 validation
The Tier 1 validation indicated strong agreement between PM-observed ROS and the occurrence of liquid precipitation from both direct measurements and empirical observations. Of the three types of precipitation validation measurements and observations, no single type showed consistent agreement with the PM-observed ROS record (table 2). Of the 11 ROS event observations made by the NWS observer over the 13-year record, ten were detected as PM-observed ROS days. The precipitation type observations (rain, fog) from the PAFA station provided more overall observations than the direct precipitation measurements, but were unable to identify all ROS events consistently (figures 3(a) and (b)). Yet, for the years of interest, ROS omission errors indicated from all three types of validation observations ranged from 0 to 5 events, with associated accuracies ranging from 75% to 100%.
The results summarized in figures 3(a) and (b) also show several days in early November when the GRP was < 1, suggesting early wintertime freeze-thaw transitions due to sensible and latent heat flux, as  there was limited precipitation measured during this period. Also, significant snowfall is reported in late February (figure 3(a)) and early December ( figure  3(b)). During these snowfall events, the GRP remained above the detection threshold, which suggested confidence in the identification of ROS rather than snowfall events when GRP was < 1 and measured precipitation was > 0.   (table 3). Further discussion on the limitations and caveats of validating the PM-derived ROS events is presented in the supplement (S4). However, results must be seen as a relative comparison as Wilson et al (2013) included October and April in their annual summations. Both studies indicated a higher occurrence of freeze-rethaw or ROS days in the southwestern portion of Alaska. Bartsch (2010b) also detected melt events across Alaska using daily 13.4 GHz (Ku-band) radar backscatter retrievals from QuickSCAT during the same period and as Wilson et al (2013) and found similar results. But more interestingly, PM-derived melt events (Semmens et al 2013, Wang et al 2016 and active microwave melt events (Bartsch 2010b, Wilson et al 2013 demonstrated similar results to this study, an increasing trend in events moving from the central interior region and into southwest Alaska. The PM-observed ROS days showed the greatest occurrence in the southwest and central portions of Alaska, including the BSC, AGC, and INT regions, but the frequency and intensity of these events showed large year-to-year variability. The temporal variation by WY and month in PM-detected ROS days for each Alaskan sub-region is shown in figure 5; these results indicate that the BSC and north central portions of the AGC consistently possessed the highest mean  An inter-annual comparison between the freezerethaw record of Wilson et al (2013) derived from QuickSCAT and our calculated ROS events showed strong agreement from November to March (2003March ( -2008. Both studies indicated that the largest spatial coverage of such events occurred in November, while the NS experienced no events during March in either study. Wilson et al (2013) reported that the NS did not experience any form of melt events until April. The results of the analysis of melt events of Wang et al (2016) from 25 km PM retrievals also supported the temporal and geographical pattern in the NS. Summary statistics in table 4 show that the C during November and December in the NS is quite high at 0.91 and 1, respectively. These values indicate that even during November and December, ROS days are still an uncommon event across the NS.

Correspondence between ROS events, temperature, and precipitation
Linear regressions between temperature departure data, provided by the Alaska Climate Data Center (figure S4), and the PM-derived mean seasonal ROS events indicated that temperature had the greatest explanatory power for predicting ROS events within the BSC (p < 0.001) followed by the INT (p < 0.01) and AGC (p < 0.01) regions ( figure 6). In the NS, the temperature departure was insensitive to ROS occurrence (p < 0.9), likely due to colder climate conditions and a lower overall number of ROS events detected in the region. However, ROS events are sensitive not just to temperature, but rather the interactions between temperature, humidity, and precipitation. A correlation analysis between PM-derived ROS and Daymet-derived climate anomalies aided in exploring these interactions.
The correlations between monthly (November-March) climate anomalies and days with ROS were statistically significant (p ≤ 0.1) in many locations ( figure 7). Overall, days with ROS coincided with above-normal precipitation and temperature, with notable temporal and spatial variability in both the sign and strength of the relationships. The relationships between days with ROS and temperature, and precipitation showed greater spatial heterogeneity in November than in December, as both positive and negative relationships are observed. However, from January through March, ROS days and climate anomaly correlations became positive, with the strongest of these positive correlations occurring in February. Correlation patterns were similar for both daily minimum and maximum air temperatures and ROS events over all months represented.
Aggregated correlations for ROS days and associated climate anomalies by Alaska climate regions showed consistently positive correlations from January through March for all regions except the NS (figure 8). In November and December, the mean temperature correlations remained positive but had a large spread in the correlation distribution, including both positive and negative relationships. The mean precipitation correlations in November and December were negative in the BSC and INT regions, but as with temperature, also showed a large range of variability. The negative correlations may be a consequence of the uncertainty introduced by the Daymet model (Daly et al 2008, Oyler et al 2015, but could also be a consequence of using a maximum snow cover extent product (MOD10A2) during periods of intermittent snow. The NS region showed a predominantly positive relationship with temperature, but a more variable relationship with minimum temperature in March. Precipitation correlations in the NS were sporadic and weak due to the characteristic colder and drier Arctic climate of the region, which showed a paucity of PM-derived ROS events during the December-March period, when seasonal temperatures are generally well below freezing and the cold Arctic air mass holds little moisture. Also shown in figure 2 are below normal temperature anomalies across the NS during the study period, which also likely contributed to the infrequent occurrence of ROS days across the region.

Sources of error, limitations, and advances
The PM ROS events over Alaska showed variable correlations with the selected climate anomalies during November and December, particularly in regions with low elevation (< 200 m) pixels. Inconsistent correlations may also indicate the occurrence of misclassified pixels at lower elevations. The combination of the high variability in correlations and greater occurrence of PM-derived ROS events indicates that these areas are influenced by wet snow and by shallow, transient snowpack conditions frequently found in low-elevation landscapes (Rees et al 2010). The high-density network of tundra lakes at low elevations in the BSC region, in addition to large proglacial lakes in southwest Alaska, may also contribute to the higher number of PM-derived ROS observations in the region (Wilson et al 2013, Wang et al 2016. The generally greater occurrence of freeze-thaw events in early November and late March in these regions may contribute additional uncertainty, such that the ROS algorithm may be detecting increased LWC introduced by snowmelt in the absence of rainfall or atmospheric condensation (Dolant et al 2016).
Our validation results from Fairbanks, AK, showed that most ROS events occurred in early November. The timing of ROS events coincided with many fog observations, indicating that latent and turbulent heat flux driven snowmelt may have contributed to the ROS detection during periods with no measured precipitation (Semmens et al 2013, Wang et al 2016. These results are also consistent with a prior study indicating that fog and positive temperatures are a primary driver of melt events in the Yukon River Basin (YRB), and that that the presence of fog is an effective indicator for warm air intrusions (Semmens et al 2013).
The results from this study were similar to the ROS spatial and seasonal patterns reported from previous studies involving different satellite microwave sensors, classification algorithms and study periods (Bartsch 2010b, Semmens et al 2013, Wilson et al 2013and Wang et al 2016. These similar findings include a generally greater occurrence of melt events in the southwestern part of Alaska. The studies show that melt events are very infrequent from November to March, but increase dramatically further into spring. While the combined results from these studies indicated multiple effective methods for classifying and documenting ROS and associated melt events from different algorithms and sensors, they do not utilize ongoing sensor missions and/or report ROS events at a 6 km resolution. In this study we address this by: (1) creating a new ROS record over Alaska using synergistic MODIS and AMSR-E/2 observations that overlap the NASA ABoVE campaign, while enabling continuity of the ROS record through continuing satellite operations; (2) using an algorithm approach that requires only limited inputs emphasizing MODIS SCE and AMSR T retrievals, while the combined information from these sensors supported finer (6 km) resolution delineations of ROS patterns and environmental gradients; (3) providing a regional application of the GRP algorithm, which extended previous localized GRP applications involving in situ field sites (Grenfell and Putkonen 2008, Dolant et al 2016, Langlois et al 2017.
Future ROS record versions will continue to focus on refining the GRP threshold to more effectively account for variations in snow conditions and constraining uncertainties over different land cover types. Yet, regardless of the threshold challenges, this study demonstrated the utility and effectiveness in using the GRP to detect ROS and associated melt events across a large boreal-Arctic landscape. Potential applications of the GRP to detect other snow processes including snow onset, melt onset, and duration remain to be explored.

Consequences of climate change and ROS events in Alaska
Studies examining projected temperature and precipitation trends over Alaska in the latter part of the 21st century indicated future warmer winter and annual temperature conditions across the state (Stafford et al 2000, Serreze and Francis 2006, Bieniek et al 2014; and historically, over the past 60 years, Alaska has experienced almost double the rate of warming relative to other regions in the United States (Chapin et al 2014). While our results from both temperature departures and climate anomalies indicated that ROS frequency is intensified in years with anomalously high temperatures, these temperature anomalies are often driven by large-scale atmospheric circulation patterns that have been found to be highly correlated with ROS (Cohen et al 2015). Such warm events in Alaska are associated with southwesterly flows and Pacific-North American (PNA) pressure systems that promote ROS and melt events from October to December (Rennert et al 2009, Semmens et al 2013. More recent studies indicated that stratospheric circulations (i.e. polar vortex) strongly influence Alaskan winter temperatures. Specifically, during periods of a weak polar vortex, cold polar air masses are replaced by warmer conditions known as 'warm Arctic cold continents' (Overland and Wang 2010, Cohen et al 2014, Kretschmer et al 2018, which may promote ROS and associated melt events. The atmospheric blocking and enhanced winter temperatures are also purported to be a major driver of recent record warm Arctic temperatures and record low sea ice extents (Cohen 2016).
Projected warming trends across Alaska and the Arctic (Chapin et al 2005) are expected to increase variability in regional snow cover conditions. Model simulations projected a 10%-20% decrease in SCE across the Arctic by 2050, with the greatest losses over Alaska (Callaghan et al 2011). These warming trends may also increase the frequency, duration and extent of surface thawing and refreezing, rainfall and mixed precipitation events, altering snowpack structure and decreasing snow-covered area and duration (Chapin et al 2005, Callaghan et al 2011, Cohen et al 2015, Kim et al 2015. All these factors are expected to contribute to the polar amplification of global warming due to the important role of snow cover on surface albedo and the terrestrial energy budget Francis 2006, Derksen andBrown 2012). The changing snow cover conditions are also expected to impact regional hydrology and ecosystem processes, due to the role of snow cover as an important water storage and thermal buffer influencing underlying soil active layer temperature and moisture constraints on ecosystem processes, and permafrost stability (Cohen et al 2012, Yi et al 2015. Some studies suggested that ROS events will become more common in a warming climate across the ABR (Semmens et al 2013, Jeong andShushama 2018), which is consistent with an analysis of the recent historical record reporting an annual increase of about seven melt event days per year from 1998 to 2013 over the pan-Arctic (Wang et al 2016). However, the long-term influence of enhanced ROS and melt events within Alaska and the associated impacts of these changes on the regional hydrology, ecosystems and human populations remain uncertain.

Summary
This paper presented a new satellite-derived ROS dataset derived from MODIS snow cover observations and a passive microwave spectral gradient ratio-based classification (Dolant et al 2016) derived using calibrated 6 km AMSR T records at 19 GHz and 37 GHz frequencies. The daily ROS classification was conducted over Alaska for the winter months (November-March) from WYs 2003WYs -2016WYs (excluding 2012. A two-tiered validation approach using regional weather station observations indicated favorable ROS classification accuracies ranging from 75% to 100%. The resulting multi-year satellite record revealed markedly higher ROS frequencies in the southwest and central portions of Alaska. The ROS days also occurred most frequently in November and December and coincided with warm temperature anomalies. ROS events were consistently observed in the BSC and AGC during all months of the year, and often occurred during periods of above-normal temperatures in the INT and NS regions. These results were similar to previous remote sensing-based ROS studies derived over different periods and using different classification algorithms; together, these results indicate strong sensitivity of satellite microwave remote sensing to related ROS processes. The northern boreal and Arctic regions are characterized by an extended period of seasonal snow cover, which strongly influence regional ecosystems, hydrological processes, the surface energy budget and global climate. As the northern latitudes continue to experience accelerated warming at roughly twice the mean global rate, ROS is expected to play a more significant role in both ecological and hydrological processes. To understand future implications of enhanced ROS events, we presented a ROS algorithm that utilized satellite observations from current operational satellites (AMSR2, MODIS), enabling ROS retrievals over Alaska that overlap with recent extensive and planned field campaigns from the NASA ABoVE. Thus, the data record developed in this study, when synthesized with other biophysical observations, are expected to contribute to addressing several data gaps and ABoVE science objectives pertaining to climaterelated impacts on boreal and Arctic ecosystems, wildlife, permafrost hydrology and snow processes, and associated climate impacts on human-natural systems.