CloudSat estimates that 1773 km
Falling snow transfers moisture and latent energy between the atmosphere and the surface. Snow impacts the surface radiant energy transfer by modifying albedo and emissivity. Accumulated snow can also act as a thermal insulator that modifies sensible heat fluxes and how surface temperature responds to changes in atmospheric conditions. Furthermore, it acts as a surface water storage reservoir (Rodell et al., 2018), providing seasonal runoff that provides freshwater supplies for both human populations and water-dependent ecosystems. Billions of people around the world depend on these resources. These water supplies are recognized as being at risk from climate change and rising global temperatures (Barnett et al., 2005; Mankin et al., 2015).
The advent of satellite-borne instruments capable of detecting falling snow and of reanalysis products that diagnose snowfall have made possible a global examination of how snowfall is distributed and its contribution to atmospheric and surface processes. Precipitation gauge measurements of snowfall for meteorological and hydrological purposes provide valuable data but have historically suffered shortcomings related to spatial sampling and gauge performance (Kidd et al., 2017). Shortcomings in the accuracy of such measurements and methods to improve that accuracy have been the focus of a number of studies (Goodison et al., 1998; Kochendorfer et al., 2018). Beyond accuracy issues, these gauge networks are necessarily of limited spatial coverage, potentially biasing climatologies over large domains. Coverage of ocean regions is not possible. Over land, gauges tend to be located near inhabited areas, leading to sparse or nonexistent coverage in more remote locations (Groisman and Legates, 1994). These remote locations include areas such as the high latitudes and mountains, where snowfall can be the dominant form of precipitation. Even when these areas have relatively dense gauge networks, such as the CONUS (contiguous United States) mountains, gridded datasets have their limitations, most notably gauge under catchment issues and large snowfall accumulation gradients in complex terrain that are often insufficiently sampled by existing in situ networks (Henn et al., 2018).
Given these shortcomings in snowfall surface observations, studies on snowfall in remote locations commonly rely on reanalyses (e.g., Bromwich et al., 2011). Reanalyses utilize numerical weather prediction models to integrate observations of large-scale geophysical fields (e.g., temperature and water vapor). One strength of reanalysis datasets is their continuous spatial and temporal coverage. However, the veracity of reanalysis snowfall datasets depends strongly on the underlying model and the assimilated datasets, which often exhibits systematic and varied biases (Daloz et al., 2018). In addition, their low spatial resolutions can be a limitation especially in regions of complex topography, and reanalyses should therefore be used with caution. For example, Wrzesien et al. (2019) showed that reanalyses have large biases in terms of snow water equivalent (SWE) over North America. Wang et al. (2019) compared the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis 5th generation (ERA5) and ERA-Interim snowfall estimates over Arctic sea ice and showed higher snowfall in ERA5 compared to ERA-Interim, resulting in a thicker snowpack for ERA5. Orsolini et al. (2019) focused on the Tibetan Plateau and evaluated snow depth and snow cover estimates from reanalyses (ERA-5; ERA-Interim; Japanese 55-year Reanalysis, JRA-55; and Modern-Era Retrospective analysis for Research and Applications 2, MERRA-2), in situ observations and satellite remote sensing observations. They showed that reanalyses can represent the snowpack of the Tibetan Plateau but tend to overestimate snow depth or snow cover. Snow accumulation measurements from automatic weather stations are compared to reanalysis datasets (ERA-Interim and National Center for Environmental Prediction-2, NCEP-2) over the Ross Ice Shelf in Antarctica in Cohen and Dean (2013). While both reanalysis datasets miss a number of accumulation events, ERA-Interim is able to capture more events than NCEP-2. Liu and Magulis (2019) evaluated snowfall precipitation biases over high-mountain Asia in MERRA-2 and ERA-5. The results show that, at high altitudes, snowfall is underestimated in both reanalyses. In this current study, four reanalysis datasets will be examined: MERRA, MERRA-2, ERA_Interim and JRA-55.
As an alternative to reanalyses, snowfall rates can now be assessed using satellite observations (with sufficient spatiotemporal coverage) provided by CloudSat's Cloud Profiling Radar (CPR). CloudSat observations, nearly continuous since 2006 (Stephens et al., 2002, 2008), have been applied to produce near-global estimates of snowfall occurrence and intensity (Liu, 2008; Kulie and Bennartz, 2009; Wood and L'Ecuyer, 2018). The resulting datasets have been examined extensively from local to global scales (Liu, 2008; Kulie and Bennartz, 2009; Hiley et al., 2011; Palerme et al., 2014; Smalley et al., 2015; Chen et al., 2016; Behrangi et al., 2016; Norin et al., 2015; Milani et al., 2018; Lemonnier et al., 2019a, b). CloudSat has substantially extended the spatial extent of precipitation measurements compared to existing gauge or radar networks. In particular, these instruments have greatly enhanced the observations of light precipitation, including snowfall over oceans, over remote high-latitude regions and over inaccessible land areas (e.g., Behrangi et al., 2016; Milani et al., 2018; Smalley et al., 2015; Norin et al., 2017; Lemonnier et al., 2019a, b).
However, satellite-based retrievals also have inherent uncertainties
related, for example, to their limited temporal coverage. For instance, they
might miss some heavy events such as atmospheric rivers in western North America and
South America (Ralph et al., 2005; Neiman et al., 2008; Viale and Nunez,
2011). Therefore, CloudSat snowfall retrievals have been extensively
assessed against a wide range of independent ground-based measurements.
Hiley et al. (2011) seasonally compared CloudSat snowfall estimates with
Canadian surface gauge measurements, showing better results for higher
versus lower latitudes – especially lower-latitude coastal sites. They
speculated that the latitudinal differences might be due to CloudSat
sampling (more observations at higher latitudes), snow microphysical
differences associated with warmer snow events that could affect CloudSat
estimates (e.g., wetter snow, rimed snow and/or mixed phase precipitation),
or precipitation-phase identification issues associated with snow events in
the 0–4
In spite of the noted shortcomings in snowfall datasets from gauges, radar
and reanalyses, mountain snowfall has not yet been thoroughly studied using
multiple reanalyses and the CloudSat dataset. In this study, we derive
mountain snowfall from five datasets (CloudSat 2CSP, MERRA, MERRA-2,
ERA-Interim and JRA-55) to answer the following questions.
How much snow falls on the world's mountains? What percentage of continental snow falls on mountainous regions?
Given the challenges in retrieving snowfall from single-frequency radar observations, especially in complex terrain, the CloudSat estimates are not treated as the “reference” dataset, though we note that they are the only estimates derived directly from observations. All five sources are treated as providing valid independent estimates of the fraction of snow that falls in mountainous compared to all continental regions to document the current state of knowledge in this field. The next section presents the different datasets employed in this study, as well as methodological information such as the mountain and continental masks. Section 3 compares mountain snowfall fraction and magnitudes between the different datasets, while the following section (Sect. 4) discusses the differences in absolute magnitude of snowfall estimates. Finally, Sect. 5 summarizes the results of this study and offers concluding remarks.
For this work, the CloudSat data are spatially gridded onto a
CloudSat's 2CSP snowfall product, version R04 (Wood et al., 2013), provides
estimates of instantaneous surface snowfall rates (S) for each of these
pixels derived from the observed vertical profiles of radar reflectivity
(Z). A version R05 is now available; however, the snow retrieval status
variable is evaluated in the same way in the two versions of the product.
The global snowfall amount is very similar in R04 and R05; therefore, the results
should only differ slightly with the new version of CloudSat. The data are
spatially gridded onto a
Snow and rain are discriminated based on the CloudSat 2C-PRECIP-COLUMN
product (Haynes et al., 2013), which applies a melting-layer model driven by
the ECMWF analyses temperature profiles. Snow particles are assumed to melt
following the model of melted mass fraction described by Haynes et al. (2009). All profiles with melted fractions less than about 15 % at the
surface (
This study also considers four modern reanalyses: MERRA, MERRA-2,
ERA-Interim and JRA-55. MERRA (Rienecker et al., 2011;
CloudSat has not been assimilated in any of the four reanalyses, thus it can be
considered independent. All datasets used in this study are bilinearly
interpolated from their native resolution to match the
Spatial maps of the continental mask
The table summarizes the snowfall estimates of mountain and
non-mountain snowfall for MERRA, MERRA-2, ERA-Interim, JRA-55 and CloudSat,
for the time period 2007–2016, for Eurasia, North America, South America,
Africa and globally. Global snowfall is the cumulative snow falling over all
land in the world, which includes the four continents already cited plus
Greenland, Australia and Antarctica. For each area and dataset, a table cell
shows: the amount of mountain (top left), non-mountain snow (top right,
km
Snowfall estimates from all sources are partitioned between the different
continents using the “continental mask” shown in Fig. 1a. The
continental mask was first used in L'Ecuyer et al. (2015). Following this, the
mountain and non-mountain regions are separated using the “mountain mask”
presented in Fig. 1b. Based on the Kapos et al. (2000) definition, grid
cells are classified as mountainous based on elevation, slope and local
elevation range. They used the global digital elevation model GTOPO30 and
ARC-INFO to identify areas above particular altitudes and generate grids
containing the slope and the local elevation range, and then they combined
these variables, with adapted criteria, to define mountainous regions. The
original mask was produced using with a spatial resolution of 30 arcsec
(
In this article, total mountain snowfall is equal to the cumulative snow falling over North America, South America, Africa and Eurasia. Greenland and Antarctica are considered ice sheets and therefore do not qualify as continents with mountains. Global snowfall is the cumulative snow falling over all lands in the world, which includes the four continents already cited plus Greenland, Australia and Antarctica.
Table 1 shows the snowfall estimates for mountain and non-mountain snowfall
for CloudSat and the reanalyses, over each continent and globally. According
to CloudSat observations, 1773 km
Spatial maps of global cumulative mountain snowfall
(millimeters per month per grid box) for
To visualize where the snow is falling, Fig. 2 presents the geographical
distribution of the mountain snowfall estimates in CloudSat and the
reanalyses. As expected, in all datasets a majority of the mountain snow
falls in the Northern Hemisphere (Himalayas and Rockies; 95 %–99 %), with
little snowfall (
Spatial maps of the global cumulative mountain snowfall
(millimeters per month per grid box) over high-mountain Asia for
Table 1 also shows the contribution of mountain snowfall to total snowfall for CloudSat and each reanalysis over each continent. To get a better sense of the contribution of orography to snowfall, the percentage of mountainous grid points over each continent is provided in the last column of the table. Eurasia has the highest fraction of mountainous grid boxes with 33 % of its grid boxes considered mountains. North America and South America have a quarter of their grid boxes covered with mountains, and only 14 % of the African continent is considered mountainous. The contribution of mountain snowfall does not vary substantially between continents. For Eurasia, South America and Africa it is around 10 %, while for North America it represents around 5 % of the snow falling over the continent. Over all the continents, the agreement between the reanalyses and CloudSat observations is very good, with differences under 4 %.
Coherently with the previous section, the magnitude of mountain snowfall
estimates over the four continents varies a lot depending on the datasets
examined. MERRA's datasets and CloudSat present a similar magnitude in terms
of mountain and continental snowfall, while ERA-Interim and JRA-55 present
much lower estimates than the other datasets. For example, over Eurasia the
values for mountain snowfall vary between 379 and 1440 km
Snowfall estimates (millimeters per month per grid box) over
The previous section showed a very good agreement between all the datasets in terms of mountain snowfall fractions. However, the spatial maps presented in Fig. 2 and the absolute snowfall amounts in Table 1 showed substantial differences in magnitude between the different datasets. This is further demonstrated in Fig. 4, which summarizes the snowfall estimates in a millimeters per month per grid box over Eurasia, North America, South America and Africa and its partitioning between mountainous (blue) and non-mountainous areas (yellow) for the five datasets. To ease the comparison between the different datasets, here the snowfall amounts are normalized by the number of mountain and non-mountain grid boxes. There is some consistency in the relative behavior of the various datasets between the regions. Consistent with the results in Sect. 3, JRA-55 always has the lowest estimates of snowfall per grid box (see Table 1). For example, over North America and Eurasia, JRA-55 produces 68 % less snowfall than the average of the four other datasets (Fig. 4). Even so, when looking at Fig. 5, which presents the frequency of snowfall occurrences for each continent for all datasets, the frequency of snowfall occurrences for JRA-55 is very close to the other products. This indicates that JRA-55 underestimates the intensity of many snowfall events. ERA-Interim also tends to be on the lower end of the spectrum concerning snowfall compared to the other datasets (Fig. 4). This can be at least partly attributed to its systematic lower frequency of snowfall occurrences (see Fig. 5). With the exception of North America, MERRA-2 generally has the highest total snowfall compared to the other datasets (Fig. 4). Again, this is consistent with the results shown in the previous section. This overestimate is related to the way this dataset represents the frequency of snowfall events. MERRA-2 produces much more snowfall events than the other datasets (see Fig. 5). This bias might be similar to the bias identified for precipitation in climate models, producing too frequent and too lightly precipitating events, referred to as “perpetual drizzle” (Stephens et al., 2010). This could be happening for snowfall events in MERRA-2.
Frequency of occurrence of snowfall estimates over
The differences in snowfall among datasets is especially prominent over Africa and South America. Over Africa (Fig. 4d), both MERRA and MERRA-2 produce much more snow than the other datasets, with MERRA-2 producing nearly twice as much snowfall as MERRA. MERRA produces 75 % more snowfall than the average of the three remaining datasets (ERA-Interim, JRA-55 and CloudSat), while MERRA-2 produces 85 % more. For the same reasons, over South America MERRA-2 produces 73 % more snowfall than the average of the other datasets. Furthermore, it highly exceeds the mountain and non-mountain snowfall compared to the other datasets. However, as most of the snow over South America is mountainous, the excess in mountainous snowfall has a stronger impact on the differences in total accumulated snowfall. The seasonal cycle of mountain snowfall over South America (not shown) provides another interesting explanation for this specific bias. From January to December, MERRA-2 overestimates the other datasets but with a similar seasonal cycle in the first part of the year. However, during the second part (after June), the behavior of MERRA-2 is very different – instead of a decrease in mountain snowfall, snowfall accumulations remain very high and steady. This is clearly a major contributor to the high snowfall estimates of MERRA-2 over South America.
Overall, these results are coherent with previous studies comparing different reanalysis datasets (Daloz et al., 2018; Sebastian et al., 2016, Thorne and Vose, 2010). They all show that reanalyses are able to represent some general patterns but also show very important differences. For example, Sebastian et al. (2016) compared atmospheric budgets for the computation of water availability in different reanalyses. They showed considerable variations in the individual components of the different budgets and suggested that part of these variations could be attributed to differences in the representation of clouds and convective schemes for precipitation. Furthermore, Daloz et al. (2018) showed significant differences in the representation of clouds in the reanalyses examined in this article, confirming the hypothesis of Sebastian et al. (2016). More specifically, they showed that JRA-55 exhibits some strong deficiencies in the representation of clouds and that MERRA-2 introduces some biases that were not evident in MERRA. These results may partly explain the deficiencies observed for these two datasets.
Snowfall plays an important role in a number of atmospheric and surface
processes that impact energy and hydrological cycles and can influence the
Earth's climate. To understand these processes and how they will be
influenced by future climate change, it is imperative to have reliable
observations of present-day mountain snowfall. This study is a preliminary
step towards an estimate of mountain snowfall from CloudSat satellite
observations and four reanalyses (MERRA, MERRA-2, JRA-55 and ERA-Interim).
In this work we answer the following questions.
How much snow falls on the world's mountains? What percentage of continental snow falls on mountainous regions?
A total of 1773 km
A total of 4 % to 5 % of snow falls over the mountains (see Table 1).
One aim of this research is to provide context for researchers who want to use snowfall estimates globally or on specific continents from reanalyses and/or satellite observations. The results of the discussion clearly emphasize the necessity of using several datasets, including different platforms such as reanalyses and satellite observations. Results presented here can help future analyses select validation datasets for specific continents, since we show that some datasets behave differently to the others for continental snowfall estimates. For instance, modelers have difficulties accurately representing snowfall over South American mountains (Gelaro et al., 2017), and it is suspected that MERRA-2 is not the optimal dataset to use for this continent. However, this study and Wrzesien et al. (2019) showed that over North America MERRA-2 is certainly a realistic dataset with substantial skills. Generally, there is no good or bad dataset; however, some datasets may outperform others over certain continents. These different abilities in the reanalyses and satellite products can lead to issues when validating climate models, for example. We therefore recommend using an ensemble of the products in the same sense that it is recommended to use several models or simulations.
This study also suggests that estimates of the fraction of snow that falls in the mountains compared to all continental snowfall may be more reliable than estimates of the absolute magnitude of mountain snow accumulations. A hypothesis behind this result could be that the datasets presented here have a similar representation of the large-scale forcings but differences at local and smaller scales, which could be due to differences in the physical parameterizations of the models, subgrid-scale parameterizations of orographical effects. Indeed, even if the reanalyses are based on different models, they should simulate similar and realistic large-scale forcings. For CloudSat, its ability to capture these forcings would come from its relatively good level of temporal and spatial coverage. This could explain the consensus between the different datasets in terms of snowfall fractions. On the other hand, at smaller scales, both types of datasets experience different limitations, which would explain the dissimilarities in snowfall magnitude. For example, for CloudSat, its spatial coverage could lead it to miss some heavy snow events like atmospheric rivers.
In the future, this work will expand in several directions. First, a deeper and more process-oriented analysis of the differences observed during the different datasets should be done over each continent. While this study is confined to mountain snowfall produced by CloudSat and reanalysis datasets, it also serves as a foundation for studying cloud microphysical and dynamical processes operating within snow-producing clouds forced by orography. Because different modes of snowfall have varying impacts on the environment and potentially unique remote sensing fingerprints, identifying specific types of snowfall could lead to better measurements of snowfall. In addition, this could also improve forecasting by representing different snowfall modes more realistically within numerical weather models. Also, to evaluate the ability of climate models to represent snowfall estimates, this same analysis could be realized for climate models such as the Coupled Model Intercomparison Project 6 (CMIP6) ensemble.
CloudSat data used herein were acquired from the CloudSat Data Processing
Center (DPC) and at the time of writing can be accessed online at
MERRA and MERRA-2 data were provided by NASA's
Global Modeling and Assimilation Office (GMAO) and obtained through the
Goddard Earth Sciences Data and Information Services Center (GES DISC,
JRA-55 data was provided by the Japanese Meteorological Agency and obtained
through the National Center for Atmospheric Research's (NCAR) Research Data
Archive (
ERA-Interim data was provided by the European Centre for
Medium-Range Weather Forecasts (ECMWF) and obtained via the ECMWF WebAPI (
ASD and TL conceived of the presented idea. ASD developed the current work with support from MM, MK, NW, MD and MW. MM also performed the computations for preparing the datasets. MK and NW provided critical expertise on the analysis and preparation of the satellite observations. MD and MW provided the mask used for selecting mountainous regions. CS and AD provided critical feedbacks on the manuscript.
The authors declare that they have no conflict of interest.
The views, opinions and findings contained in this report are those of the author(s) and should not be construed as an official National Oceanic and Atmospheric Administration or U.S. Government position, policy or decision.
Parts of this work by Tristan L'Ecuyer were performed at the University of Wisconsin-Madison for the Jet Propulsion Laboratory, California Institute of Technology, sponsored by the National Aeronautics and Space Administration (NASA) CloudSat program. Parts of this research by Norm B. Wood were performed at the University of Wisconsin – Madison for the Jet Propulsion Laboratory, California Institute of Technology, sponsored by the National Aeronautics and Space Administration. We acknowledge the support of the DPC in providing their data.
This research has been supported by the NASA (grant no. NNX16AE21G) and the Center for Climatic Research (Seed grant).
This paper was edited by Guillaume Chambon and reviewed by Jean-Baptiste Madeleine and one anonymous referee.