Present-day and future Greenland Ice Sheet precipitation frequency from satellite observations and an Earth System Model

The dominant mass input component of the Greenland Ice Sheet (GrIS) is precipitation, whose amounts and phase are poorly constrained by observations. Here we use spaceborne radar observations from CloudSat to map the precipitation frequency and phase on the GrIS, and use those observations, in combination with a satellite simulator to enable direct comparison between observations and model, to evaluate present-day precipitation frequency in the Community Earth System Model (CESM). The observations show that substantial variability of snowfall frequency over the GrIS exists, that snowfall occurs 5 throughout the year, and snowfall frequency peaks in Spring and Fall. Rainfall is rare over the GrIS, and only occurs in regions under 2000 m elevation and to the peak summer season. Although CESM overestimates the rainfall frequency, it reproduces the spatial and seasonal variability of precipitation frequency reasonably well. Driven by a high-emission, worst-case RCP8.5 scenario, CESM indicates that rainfall frequency will increase considerably across the GrIS, and will occur at higher elevations, potentially exposing a much larger GrIS area to rain and associated meltwater refreezing, firn warming, and reduced storage 10 capacity. This technique can be applied to evaluate precipitation frequency in other climate models, and can aid in planning future satellite campaigns.


Introduction
The Greenland Ice Sheet (GrIS) contains the largest volume of ice on the Northern Hemisphere, equivalent to 7.3 meter sea level equivalent. While the GrIS has been losing mass since the 1970s, and likely also in earlier episodes in the 20th century to CloudSat's near-surface radar reflectivity to estimate near-surface precipitation frequency (Haynes et al., 2009;Ellis et al., 2009;Smalley and L'Ecuyer, 2015). Recently, these near-surface radar reflectivity derived precipitation frequencies have been compared to climate model output in a scale-aware and definition-aware framework (Kay et al., 2018). Here, we use this framework to compare present-day GrIS precipitation frequency between observations (CloudSat) and an Earth System Model (CESM). After understanding present-day biases, we assess future 21st century changes in precipitation frequency over the 5 GrIS, and discuss the implications for future radar missions. We start this paper with a presentation of our framework for comparing models and observations and a description of the model simulations (Section 2), followed by results (Section 3).
Section 4 presents a discussion and conclusions.
2 Data and methods 2.1 Scale-aware and definition-aware framework for evaluating simulated precipitation frequency 10 Evaluating precipitation simulated by Earth System Models with satellite observations is challenged by the scale differences (model grids are ∼100 km, while CloudSat footprints are ∼1 km), and because of inherent differences in the definition of precipitation between models and observations. To address these challenges, the science community has developed a software package called Cloud Feedbacks Model Intercomparison Project (CFMIP) Observational Simulator Package (COSP, Bodas-Salcedo et al. (2011)). COSP contains a sub-column generator and instrument forward models, called simulators, to convert 15 raw model output at the model grid scale into pseudo-satellite observations at the satellite footprint. As such, COSP outputs can be directly compared to equivalent satellite observations in a scale-aware and definition-aware framework. For this study, we use the Quickbeam radar simulator (Haynes et al., 2007) to simulate modelled CloudSat reflectivity profiles. Subsequently, following Kay et al. (2018), we calculate near-surface precipitation frequency based on thresholding the modeled near-surface CloudSat reflectivity. Using this framework, we are able to directly compare modelled and observed CloudSat near-surface 20 precipitation frequency. The observations we use are gridded observations of 2CPC CloudSat near-surface precipitation frequency (Ellis et al., 2009) during 11 years (June 2006-May 2016. The model and the observations use the same reflectivity thresholds (see Table 2 in Kay et al. (2018)) for assessing near-surface precipitation frequency. Here we use the 'light snow', 'snow', 'light rain', and 'rain' categories, as heavy precipitation (as defined by Kay et al. (2018)) does not occur on the GrIS.
For more details regarding the methodology, refer to Kay et al. (2018).

Model simulations with CloudSat near-surface precipitation frequency diagnostics
We assess GrIS precipitation simulated by the Community Earth System Model (CESM) version 1 with the Community Atmosphere Model version 5 (CESM1-CAM5, CESM hereafter, Hurrell et al. (2013)). CloudSat near-surface precipitation frequency diagnostics were implemented in COSP version 1.4 (Kay et al., 2016b(Kay et al., , 2018. While this study uses COSP1.4, the CloudSat-based diagnostics described here are also available for the broader scientific community within the latest COSP 30 version, COSP 2 (Swales et al., 2018). In order to evaluate present-day GrIS precipitation and to assess GrIS precipitation in a warmer future world, we ran CESM using the Representative Concentration Pathway (RCP) 8.5 scenario. The simulations span 90 years (2006 to 2095) and was initialized in 2006 from member 1 of the CESM1 Large Ensemble (Kay et al., 2015). The same simulation has been used to assess the influence of global warming on rising cloud heights (Takahashi et al., 2019).

Present-day precipitation from CloudSat
First we explore the present-day spatial and temporal precipitation frequency patterns that have been observed by CloudSat from 2006 to 2016. Figure 1 shows the observed annual and seasonal mean spatial patterns of total snowfall (i.e. the sum of 'light snow' and 'snow') and rainfall (the sum of 'light rain' and 'rain') frequencies on the GrIS. The annual mean snowfall on the GrIS varies from ∼10% in the dry, high-elevation northern GrIS, to >30% over Southeast Greenland. The interior 10 experiences snowfall most frequently in the summer (JJA, >20%), whereas most snow in the coastal regions falls in winter (DJF), and to a less extent in Spring (MAM) and Fall (SON). Observed snowfall frequency over the oceans surrounding the GrIS is highest in the winter, particularly in the Labrador Sea (southwest of the GrIS), where winter snowfall frequency exceeds 50%. In summer, snowfall does not occur over the oceans around the GrIS. Rainfall over the interior of the GrIS is negligible throughout the entire year. Rain occurs in summer, albeit rarely (< 10%),over the marginal, low-elevation zones of the GrIS.  A unique perspective on the CloudSat precipitation frequency climatology across the GrIS can be offered by analyzing their gradients with respect to surface elevation ( Figure 3). Snow frequency varies moderately with elevation, and the highest snow frequencies (>20%) are found at the lowest elevations (< 200 m above sea level (a.s.l.)) as well as between 1500 and 2000 m a.s.l.. The latter maximum can be explained by the strong topographically forced snowfall in Southeast GrIS, where the maximum snowfall occurs at these elevations. When classifying the snow frequency, heavier snow peaks at these elevations 5 and otherwise decreases with elevation, while light snow frequency clearly increases with height and dominates heavy snow above 2000 m a.s.l. (not shown). Rainfall frequency (which is dominated by rain, as light rain is almost zero everywhere (not shown)) does not exceed 2% anywhere on the GrIS, and rain is never observed above 2000 m a.s.l.. Note that, due to the hyperbolic shape of the GrIS and steep surface slopes along the margins, low-elevation areas occupy a very small fraction of the ice sheet, while higher-elevation areas occupy a much larger fraction. This implies that, although all areas below 2000 m 10 a.s.l. experience rain, all these elevation bands combined only occupy ≈ 38% of the ice sheet area.

Present-day precipitation from CESM
The patterns of precipitation frequency across the GrIS as simulated by CESM are mostly consistent with those derived by CloudSat ( Figure 4). The highest snowfall frequencies are found in the oceanic regions neighbouring the GrIS, the North Atlantic and Baffin Bay along the southwest GrIS coast. On the ice sheet, snowfall frequency is highest in the south (>40%), 15 and decreases northward to low values of <20% in the high-elevation interior. CESM simulates a clear seasonal cycle in snowfall frequency, with highest frequency in winter and lowest in summer. CESM produces rainfall on the oceans around Greenland during most of the year, while GrIS rainfall is constrained to the summer season, and limited to the coastal regions.
Next, we compare the CESM simulated precipitation frequency on the GrIS to the frequencies derived by CloudSat ( Figure   5). Snowfall frequency over the GrIS is generally overestimated by CESM, especially in winter and fall (>15%). Over the 20 surrounding oceans, CESM clearly produces more frequent snowfall than CloudSat, with up to 75% more frequent snowfall in 5 https://doi.org/10.5194/tc-2020-31 Preprint. Discussion started: 11 February 2020 c Author(s) 2020. CC BY 4.0 License. the North Atlantic in winter. In contrast, interior GrIS summer snowfall frequency is slightly lower in CESM than in CloudSat.
In contrast with CloudSat, CESM only produces rainfall in the low-elevation GrIS coastal zones, and in summer, but the rain frequencies are clearly overestimated, especially over the western GrIS ablation zone and the oceans. CESM produces slightly lower rain frequencies in the North Atlantic compared to CloudSat.
6 https://doi.org/10.5194/tc-2020-31 Preprint. Discussion started: 11 February 2020 c Author(s) 2020. CC BY 4.0 License.  The seasonal cycle of precipitation frequency averaged over the GrIS, as shown in Figure 6, highlights seasonal variations in light snow and light rain frequencies as simulated by CESM. In summer, the only season in which light rain occurs according to CESM, the simulated light snow frequency is smaller than in the other seasons. Throughout most of the year, the simulated light snow contributes more to the total snowfall frequency than the heavier snow. This heavier snow also exhibits less of a seasonal variability than the light snow. Similarly, light rain dominates the total rainfall across the Greenland ice sheet, as 5 heavier rain does not occur.
Analyzing the differences between CESM and CloudSat with respect to elevation across the GrIS (Figure 7), we see that CESM overestimates snowfall frequencies with 5 to 10% at all elevations. CESM also produces an increase of snow frequency with elevations from the coast to 2000 m a.s.l., which is not confirmed by CloudSat. With regards to rain, CESM clearly produces too high frequencies at lower elevations (double to triple the CloudSat frequency). On the other hand, the model  correctly simulates the clear decrease in rain frequency above 1500 m a.s.l., and agrees with CloudSat in that it simulates no rain above 2000 m a.s.l..
A part of these discrepancies between CESM and CloudSat may be ascribed to CESM (at its horizontal resolution of 1 degree) not resolving the steep topography and related surface climate and precipitation gradients of the marginal GrIS. This is illustrated in Figure A1, which shows that the original CESM grid overestimates the extent of low-elevation areas and under-5 estimates the extent of high-elevation areas of the GrIS. While we have attempted to correct for this by regridding the CESM results to the GIMP grid, which virtually removes this bias (green line in Figure A1), this implies that the CESM atmospheric model 'feels' a lower topography of the coastal GrIS than in reality, enhancing atmospheric and surface temperatures and rain in these elevations. However, since the model also produces too much snow at these elevations, we conclude that CESM tends to exaggerate the precipitation frequency of both snow and rain across the GrIS, rather than attributing the incorrect phase to 10 precipitation. While acknowledging these model biases in absolute precipitation frequencies, we argue that, overall, CESM reproduces the spatial patterns and seasonal cycle of snow and rain frequency satisfactorily well. This allows us to use CESM to analyze future changes in precipitation frequency on the GrIS.
The 21st century changes in precipitation frequency, as depicted in Figure 8, are substantial over the entire ice sheet. Across 5 the south and much of the coast of the GrIS, annual snowfall frequency decreases by up to 10%. This contrasts the interior of the ice sheet, where annual snowfall frequency increases by up to 10%. This coastal decrease and interior increase is most clearly present in the summer (JJA), when coastal decreases in snow frequency exceed 20% to up to 40% in the southern GrIS. The increase in GrIS interior snow frequency is consistent throughout all seasons, and most prominent in winter (DJF). Snowfall frequency over the oceanic regions surrounding the GrIS decreases throughout much of the year, although strong increases to 10 the north are noted in winter, and to a lesser extent, in spring and fall. This snowfall increase is potentially associated with sea ice loss in these regions in the 21st century. More open water leads to enhanced atmospheric instability, condensation, and precipitation.
Rain frequency change shows a much more homogeneous signal across the GrIS and neighboring oceans Figure 8. Annual rain frequency increases with 5-15% across the entirety of coastal GrIS, which essentially leads to a doubling of the present-day 15 CESM rain frequency in these regions. While the winter season is still too cold for any rain on the GrIS at the end of the 21st century, rainfall occurs more frequently in spring, summer, and fall, and this frequency increase peaks in summer.
Averaged over the GrIS, the change in heavier snow frequency is slightly positive (0 to 2%) in winter, and negative in summer (down to -4% in August). Light snow frequency only changes substantially from June to October, with a decrease that also peaks in August (-4%). While heavier rain still doesn't occur on the GrIS at the end of the 21st century, light rain clearly 20 increases, and dominates the change in snow frequency. At the end of the 21 century, light rain occurs in all months outside 9 https://doi.org/10.5194/tc-2020-31 Preprint. Discussion started: 11 February 2020 c Author(s) 2020. CC BY 4.0 License.  the core winter (November to March), which suggests that the rain-occurring season is extended with about 4 months relative to the present ( Figure 6). In summer, the increase of light frequency peaks at almost 10%, which implies that rain frequency more than triples in summer relative to the present.
The increase in (light) rain frequency is apparent over most of the GrIS (Figure 10), with roughly a tripling of rain frequency at all elevations below 2500 m a.s.l.. End of the 21 st century rain frequency varies between 7 to 13% at elevations between 5 10 https://doi.org/10.5194/tc-2020-31 Preprint. Discussion started: 11 February 2020 c Author(s) 2020. CC BY 4.0 License. 0 and 1500 m a.s.l., and decreases sharply above that elevation. However, the area of the GrIS that experiences at least some rain clearly extends inward and to higher elevations. Rain is projected to occur at elevations up to 2500 m a.s.l., in comparison to <2000 m a.s.l. in the present-day period. This exposes an additional area of >250,000 km 2 (>15%) of the GrIS to liquid precipitation. In comparison to the rain changes, the changes in snow frequency are relatively small, with a small (0-2 %) decrease in snow frequency below 2500 m a.s.l., and a small increase (up to 2%) above that elevation, on the high GrIS interior.

5
The relative minor change in snow frequency indicates that the increase in rain frequency is not completely compensated for a decrease in snow frequency. This finding signals that overall precipitation frequency is increasing over the GrIS, with an increase of rain dominating over the entire ice sheet but the highest elevations, where rain does not occur and snow frequency increases.

10
In this paper, we used observations derived from active radar remote sensing (CloudSAT) and simulations with the Community Earth System Model to characterize precipitation frequency over the Greenland Ice Sheet. For the present-day climate, the observations show that snowfall occurs frequently over the GrIS, with variations (1) in snowfall classification (light and heavier snow occur approximately equally frequently), (2) temporally throughout the year, and (3) spatially across the ice sheet.
Rainfall, on the other hand, is rare, and only occurs in summer and at elevations below 2000 m a.s.l..

15
These observations were subsequently used to evaluate precipitation frequency output generated by CESM. The model is equipped with a satellite simulator, which allows for a consistent 'apples-to-apples' comparison with the observations. The results showed that, while CESM overestimates precipitation frequency on the GrIS overall, the model shows a realistic seasonal cycle and spatial gradients.
To then analyze future changes in GrIS precipitation frequency, we analyzed CESM output for the end of the 21st century.

20
The model suggests dramatic changes in the occurrence of rainfall, with rain occurrence extending in time (from April to October) and at much higher elevations (up to 2500 m a.s.l.). In contrast, snow frequency changes only marginally, and only increases across the high-elevation GrIS.
The comparison between CESM and CloudSat revealed clear biases in the simulated snow and rain frequency. This result is consistent with the work of McIlhattan et al. (2017), who showed that the overestimated CESM snowfall frequency is 25 potentially related to a exaggerated growth of cloud ice in expense of supercooled cloud liquid water in the model. The lack of supercooled liquid in polar clouds in CESM has been reported on previously (Miller et al., 2018;Kay et al., 2016a), and leads to substantial biases in surface downwelling longwave radiation and surface temperature (Kay et al., 2016a), and GrIS surface melting.
Carefully recognizing these CESM biases, caution is warranted when quantitatively assessing simulated changes in the 30 precipitation frequency throughout the 21st century. Doing so, we suggest to focus particularly on relative changes simulated by CESM, which are likely more robust than the absolute changes. For example, while the absolute change in rainfall frequency is likely biased because the present-day rainfall frequency is overestimated, the simulated tripling of GrIS rainfall frequency is potentially a more robust change. In addition, this study only uses one climate model and one climate change scenario; to further test the robustness of our findings, future work should focus on using other models, with a satellite simulator embedded, and apply various climate change scenarios.
Interpreting the relevance of these 21st century changes in precipitation frequency for the GrIS climate and mass balance, an outstanding question is how frequency relates to mass. For example, as rainfall frequency increases, does that imply that 5 there is more mass of rain added to the GrIS surface? As we do not have reliable observations of precipitation fluxes from CloudSat, we use CESM to analyze the relation between snow and rain frequency and the representative precipitation fluxes ( Figure 11). Interestingly, for both snow and rain, the relation between precipitation frequency and rate apparent, with a nearlinear increase in flux with frequency at low frequencies, and a much larger increase of flux with frequency as frequencies are higher. This relation, as suggested by CESM, indicates that even for small changes in precipitation frequency, precipitation 10 rates change considerably; for example, an increase of snow frequency from 10 to 15% is associated with an approximate doubling in snowfall rate (200 to 400 mm per year). That implies that a dramatic increase in rainfall frequency, as suggested by CESM, will be associated with much more rain on the GrIS. This has potential dramatic consequences for the GrIS surface conditions. In the GrIS ablation zone, slightly less snow in winter, and more rain in the transition seasons, will lead to more rapid degradation of the winter snowpack, expediting exposure of bare ice in Spring and delaying ice burial in Fall. Rain falling 15 on ice will decrease surface albedo, further enhancing melt, and the rain water will collect in surface lakes and streams that eventually end up in the ocean. In the percolation zone, less snow and much more rain will affect the storage capacity of the