**1. Introduction**

As a major part of the cryosphere, snow is an important component of the hydrological cycle and with its unique physical properties, it is an essential environmental variable directly affecting the Earth's energy balance. Proper description and assimilation of snow information into hydrological, land surface, meteorological and climate models is therefore important to address the impact of snow on various phenomena such as hydrological monitoring, avalanche forecast, and weather forecast, to predict snow water resources and to warn about snow-related natural hazards [1–9].

Understanding the microstructural, macrophysical, thermal and optical properties of snowpack is essential [10] and there is a grea<sup>t</sup> need for accurate snow data at different spatial and temporal resolutions to address the challenges of changing snow conditions.

Distinctive snow properties are currently determined by traditional ground-based measurements as well as remote sensing, over a range of scales, following considerable developments in instrument technology over recent years. Snow measurements are becoming increasingly important for freshwater management, mitigation of climate changes, adaptation to new climate conditions, and risk assessments such as avalanches, floods [11], and droughts [12].

At the present time in situ measurements of the snowpack state are performed on the ground at numerous stations (e.g., WMO synoptic stations) and during intensive field campaigns (e.g., [13–17]). Simultaneous measurements of snow properties and soil properties are of further advantage [18] but only available at selected stations. However, depending on the region, in situ measurements could have a relatively coarse spatial coverage and are only representative of a limited area due to spatial heterogeneities of snow [19–21]. An increase in the number of snow measurements from national high-resolution weather networks integrated into the WMO GTS (Global Telecommunication System) and WIS (WMO Information System) would thus provide a clear benefit [22]. With the implementation of the Global Cryosphere Watch (GCW) in 2011, the WMO established a program that considers the growing demand for authoritative information on past, present and future state of the world's snow and ice resources [23]. Although GCW is global in scope, the program needs activities at all scales, including regional, national and local levels [24] and recognizes the requirements for assimilation, model development and validation.

Space-born remote sensing data have the potential to provide estimates of certain snow properties [25]. In the visible (VIS) and near infrared (NIR) spectral range space-borne remote sensors (e.g., MODIS, AVHRR, Sentinel-2) can determine the snow cover extent (SCE) and snow cover fraction (SCF) at a high spatial resolution and long time-series of these data exist (e.g., [26–33]). The observation of snow cover area is of particular value in headwaters of mountainous regions [34–36] and one can expect to obtain volume information thanks to recent advances in photogrammetry and in the availability of stereo image [37]. In addition, remotely sensed daily SCE has been shown to improve performance of hydrological models applied to various catchments in Austria [38,39], Italy [40,41], Switzerland [8,42,43], Turkey [34,44–46], Iceland [47] and the USA [48].

Optically derived snow cover products are considerably limited by the presence of clouds [49], which results in spatiotemporal gaps [25]. This limitation was quantified by [38], who found that, on average, clouds obscured 63 % of Austria in MODIS daily snow maps, from February 2000 to December 2005. Similarly, [50] found that, on average, clouds obscured 55% (MODIS Aqua) and 50% (MODIS Terra) of Po river basin (Northern Italy), for the period 2003-2012. Interestingly, they have pointed out that in the Alpine region of the basin (>1000 m.a.s.l.), the presence of clouds increases during the melting season when SCE and SCF products are most relevant: on average the percentage of cloud obstruction is 70%. Thus, cloud removal procedures are necessary to mask clouds for the snow product to be used or assimilated in hydrologic and land surface models. In the literature, different cloud removal procedures were developed based on temporal and spatial filters, see, e.g., for MODIS products [39,50–55]. In addition, digital imagery for snow extent monitoring [56] are conducted with a newly developed system for acquisition, processing and visualization of image time series from multiple camera networks [57]. These systems could connect in situ measurements and remote-sensing products and could provide SCE information in overcast conditions.

Passive microwave sensors can measure snow mass (snow water equivalent, SWE) and are not affected by illumination (night, clouds), which limits optical data during much of the high latitude snow season [58–60]. However, the spatial resolution of passive microwave SWE data is currently too coarse for many watershed-scale hydrological applications in mountainous regions [30,61], and point gauge snow data have sparse and uneven spatial coverage [62] and their accuracy is sensitive to the assumptions used, the topography, and properties of the snow pack (e.g., [63–71]). Alternatively, active microwave sensors have the potential to determine snow depth or mass from space with higher resolution but require spaceborne measurements at appropriate frequencies (Ku-band) [25] and the time resolution is more limited than for passive microwave sensors. In addition, the signal is sensitive to the snow layer properties, which complicates direct estimation of SWE from the satellite signal. For example, based on the Synthetic Aperture Radar (SAR) Interferometry technique and Sentinel-1 data, Snow Water Equivalent (SWE) temporal variations with sub-centimeter measurement accuracy can be retrieved with up to 20 m spatial resolution in any weather and sun illumination condition [72].

In addition to the determination of snow characteristics by in situ and remotely sensed measurements, another approach to obtain snow properties is to use physical or conceptual snow evolution modeling. Three major classes of snowpack models are employed for various applications [73]: single-layer snow scheme (e.g., [74]), scheme of intermediate complexity (e.g., [75]) and detailed snowpack models, which differ in the description and the parameterization of the properties inside the snowpack and the related processes [76–78]. Single-layer representations of snow thermodynamics are still used in operational NWP models [20]. In more advanced land surface schemes employed by operational models multi-layer snow options with fixed or variable numbers of layers are available [20], e.g., HTESSEL at ECMWF [79], JULES at the Met Office [80], ISBA-ES in SURFEX [81], and TERRA in the ICON model at DWD [82]. Detailed snowpack models include in addition state variables for snow microstructure, namely grain sizes and shapes in layers [20,83]. However, continuous estimates of the snow state from numerical model predictions are still limited by scarcity and uncertainties in meteorological forcing data (see [84] for an example) and model structural problems for snow processes in land surface models [25,85–87].

Assimilation of remotely sensed snow-related observations and ground-based snow measurements has been proven to be an effective method to improve hydrological and snow model simulations [88–93]. Therefore, the potential of data assimilation methods to consistently improve model simulations of the snow state by assimilation of measurements from in situ as well as from remote sensing has received continuously increasing attention [25,59,88–102]. With data assimilation (DA hereinafter) techniques, an improvement of the simulated snow properties from numerical models can be obtained by the combination of observational datasets with numerical model predictions and consideration of the uncertainties of observed and modeled variables [103]. Several studies report the assimilation of in situ snow observations [62,101,103–105] even for operational applications [2,106,107]. A number of snow DA experiments taking into account remotely sensed snow properties have been performed with different observational datasets, including snow covered area and SWE [59,88,90,91,100,108,109].

However, visible or near-infrared observations do not allow assimilation updates under cloudy conditions and the updates depend on the snow depletion curve used to relate SCE or SCF to SWE [25]. Direct assimilation of SWE from passive microwave remote sensing data exists [88,100,110] but radiance assimilation may be more effective [25,111–118]. The latter approach is indeed able to overcome the difficulty arising from the non-unique and complex relationship linking the passive microwave signal and several snow properties (e.g., density, grain size/microstructure parameters, temperature and wetness). To assimilate radiance, radiative transfer modeling is needed to play the role of the 'observation operator' in the DA scheme, that is to relate the snow properties predicted by the dynamical snow scheme to the remotely-sensed observed variables as well as to provide the associated uncertainties [119]. Many such numerical models have been developed and evaluated over the last decade for the passive microwave such as HUT [120], MEMLS [121], DMRT-QMS [122], DMRT-ML [123]. Although their performances appear to be comparable (e.g., [124]), the formulations and parameterizations used in each model are diverse. This apparent paradox is only partially understood (e.g., [125,126]), which has motivated the development of a uniform, harmonized, modeling platform called the Snow Microwave Radiative Transfer Model (SMRT, [127]). This new model is also able to better represent the snow microstructure, which currently remains the main bottleneck for SWE estimation [114,128].

The European Cooperation in Science and Technology (COST) promoted and funded the Action ES1404 called "A European network for a harmonized monitoring of snow for the benefit of climate change scenarios, hydrology, and numerical weather prediction," or "HarmoSnow" for short. The HarmoSnow project (2014-2018) coordinates efforts towards harmonized snow data processing and handling practices by promoting new observing strategies, bringing together different communities, facilitating data transfer, upgrading and enlarging knowledge through networking, exchange and training, and linking them to activities in international agencies and global networks [129]. Due to the large heterogeneity of methods and tools for manual measurement of snow and their assimilation in numerical models one of the first activities of HarmoSnow was to carry out surveys to obtain an updated picture of the existing variety of a) snow measurement practices and instrumentations, and b) the data assimilation methods and snow data processing used in NWP, hydrology, and climate studies by the European institutions. The results of the first survey are published in [11]. This paper aims to assess the current situation and understand the diversity of usage of snow observations in DA, forcing, monitoring, validation, or verification within NWP, hydrology, snow and climate models. Our findings are based on the responses from the community to the second survey, on snow DA methods and processing, and on literature review.

#### **2. European Survey on Usage of Snow Observations in Data Assimilation, Forcing, Monitoring, Validation, or Verification**

The survey was conducted via an online questionnaire from September 2015 to December 2017 on the COST HarmoSnow website. This questionnaire (see Supplementary Material: COST ESSEM 1404 working group 3 survey: Questionnaire and results) was compiled by COST HarmoSnow experts in snow modeling and data assimilation and distributed across the COST, EUMETSAT H-SAF and GCW member networks. The questionnaire was answered by 51 participants from 31 countries. The survey consists of 32 questions in six sections and one text box for additional comments (see Supplementary Material: COST ESSEM 1404 working group 3 survey: Questionnaire and results), which are also available at the COST HarmoSnow website. Most questions used multiple choice answers. This procedure ensures clear answers and space for further explanations was provided. A weighting of the answers was not made, but we are aware of their heterogeneity in terms of institutional representativeness and implications for the representativeness of our derived conclusions. The evaluation was performed manually.
