**1. Introduction**

The seasonal snow cover has an important role as hydrological storage for the Earth's fresh water resources. The amount of water stored in the snowpack as snow and ice is expressed as snow water equivalent (SWE) and is a key variable in water resources management, which is an essential component within the Earth's climate system [1]. The amount of water, which is released seasonally (or in events) as snowmelt in the rivers, as well as the timing of the water release, mainly in spring, is relevant for numerous hydrological applications, such as hydropower production, irrigation, and fresh water supply. In addition, knowledge about the snow situation is a concern of many safety related institutions and businesses, such as avalanche warning centres and (re-)insurance companies. The onset of snowmelt and its intensity are major drivers for flood forecasting, especially in mountainous areas, and are, besides the knowledge on the total amount of water stored as snow, a very valuable information for hydropower companies.

Regarding in situ snow measurements, until now, SWE is mainly measured manually by weighing a given volume of snow, which is cut out of the snowpack with tubes [2]. This approach is reliable but can provide only a snapshot in time and in space, and in addition, it is labour-intense and destructive. Automatic SWE measurements are mainly performed by weighing systems like snow pillows and snow scales [3]. These methods provide time series, but the instruments and their installation and operation are quite costly, and their results might be affected by ice-bridging effects of thermal fluxes resulting in potential over- or underestimations of SWE [4]. Alternative SWE in situ observation systems make, e.g., use of cosmic rays or neutron rays [5] or apply a passive gamma monitoring sensor (GMON) [6]. However, the reliability of those sensors depends strongly on the underlying surface conditions, the measurements are limited to a certain amount of SWE and thus are only applicable at certain locations [7].

As an alternative to the standard in situ methods, L-band Global Navigation Satellite (GNSS) signals can be used to derive snow cover properties. Different methods were developed within the last years. As an advantage, it is possible to apply these methods globally as the GNSS signals can be tracked almost any place on Earth, they are non-destructive and can be used even for large amounts of snow [8]. Snow height (HS), for example, can be determined by the reflection of GNSS signals on the snow–air interface [9]. Liquid water content (LWC) can be derived by GNSS signal strength attenuation through a snowpack of a given volume [10]. Henkel et al. [11] presented for the first time the possibility to derive SWE for dry snow conditions using two low-cost GNSS sensors for a carrier phase-based approach to detect signal changes within the snowpack. Steiner et al. [12] confirmed this by using a similar technique with geodetic sensors and applying different ambiguity resolution strategies and wideline combinations. Finally, we developed a novel approach combining GNSS signal attenuation and time delay by combining information on GNSS carrier phases and signal strengths. We accomplished deriving the three snow cover parameters SWE, LWC, and snow height in parallel, as recently demonstrated in Koch et al. [8].

Spatially distributed snow information such as the snow cover extent, information weather the snow is dry or wet, and dry snow SWE or snow height, can be derived from Earth Observation (EO) data, based on different remote sensing techniques using active or passive microwaves, or optical, infrared or thermal approaches. An overview is given, for example, by Hall [13] and Tedesco [14]. In recent years, especially the freely accessible Sentinel-1, -2 and -3 data are a useful source for determining the above-mentioned snow parameters, like snow extent, or wet snow [15]. Besides 'raw' satellite data, also different, often project-based, internet portals like GlobSnow, CryoLand, Google Earth Engine, or EUMETSAT H-SAF are providing already processed satellite-based snow parameter products. However, in general, all remote sensing products are often not available in high-temporal resolution or may lack the required spatial resolution. This is especially the case for optical images, e.g., from MODIS, which face potential cloud cover issues [16–18]. Active microwave products are often restricted due to foreshortening or layover effects, in particular in mountain regions and passive microwave products are very coarse regarding their spatial resolution. Recent approaches tend to apply

more and more multi-sensor techniques to overcome some of these limits, which is, e.g., currently a big aim of the NASA SnowEx campaign [19]. Additionally, the combination of EO and hydrological model approaches helps to increase the temporal and spatial availability of snow or run-off information as e.g., presented by Cline et al. [20] and Immerzeel et al. [21].

In the current study, we present a comprehensive overview of a combined approach on using in situ measurements, EO, and hydrological modelling to derive continuous information on snow parameters and run-off. The applied methods and sensors of the in situ component as well as hydrological services were designed, developed and demonstrated in the framework of the business applications demonstration project SnowSense (2015–2018), which was co-funded by the European Space Agency (ESA). The SnowSense service mainly targets snow hydrological applications and is based on three pillars, including (i) a newly developed SnowSense in situ snow monitoring stations based on GNSS signals, (ii) EO products of the snow cover extent and information if and where the snow is dry or wet, and (iii) an integrated physically-based hydrological model.

The in situ and EO information are used to assimilate the input and the parameters of the applied hydrological model PROMET (Processes of Mass and Energy Transfer) [22] to calculate SWE, snowmelt onset, and river run-off in catchments as spatial layers. Those data layers contain the relevant information for flood forecasts and hydropower plant management, particularly for so far non- or sparsely equipped catchments in remote areas. Within the project demonstration phase, we validated the GNSS in situ snow stations and the first run-off results of the combined approach were already provided as an operational service for a commercial hydropower plant company and the administration of the island of Newfoundland, Canada, being our first demo users.

#### **2. Materials and Methods**

#### *2.1. The SnowSense Concept*

The overall aim of the demonstration project SnowSense was to build up an integrated service for run-off and hydropower assessment and forecast related to snow cover dynamics in remote areas. The idea of the service concept encloses the integration of in total three components for the provision of snow hydrological information. The service is based on the integration of (i) GNSS-based in situ measurements, (ii) EO monitoring and (iii) a hydrological model, to generate SWE, snowmelt and run-off products for remote areas in a temporal and spatial resolution, related to user's operation defaults. We chose one day as temporal and 1 km as spatial resolution. Furthermore, it is possible to provide estimations and forecasts on hydropower generation (Figure 1).

**Figure 1.** The modular concept of SnowSense, integrating satellite technologies like Global Navigation Satellite System (GNSS), satellite communication and Earth Observation (EO) as the basis for a service.

In the following subchapters, the applied Canadian test sites and the description, development and assimilation of the three basic components of the snow cover sub-system as well as the entire service are presented.

#### *2.2. SnowSense Testsites in Canada*

#### 2.2.1. The Island of Newfoundland

The main region of interest for the development of the SnowSense service was the remote location of the island of Newfoundland (97,000 km2), which is a province of Newfoundland and Labrador, Canada (Figure 2). The altitudinal gradient of this island ranges between 0 m and 800 m a.s.l. Newfoundland is primarily characterized by having a subarctic and a humid continental climate, with an annual precipitation ranging from 900 mm on the Northern Peninsula to 1700 mm in the southwestern region of the island [23,24]. This environment, together with a low population, provides the island of Newfoundland ideal conditions for taking advantage of using run-off for hydropower and building up reservoirs. However, the entire region has so far an insufficient instrumentation with only two functioning SWE measurement stations [25].

**Figure 2.** Overview on locations of SnowSense in situ stations (red triangles) and run-off stations (purple diamonds), within the Humber, Exploits, and Gander catchments in Newfoundland, Canada. Catchments are outlined with a red line. The locations of the two existing snow water equvalent (SWE) stations are indicated as black dots. The background shows the modelled SWE, which was assimilated by in situ stations and Earth Observation (EO) for the 15 March 2018.

As shown in Figure 2, we had access to several run-off gauges to validate our combined and assimilated hydrological service. In this study, we focus on the Humber and the Exploits catchments (Figure 2), which are of of most interest for the operations of the users. As marked in Figure 2, we set-up seven SnowSense in situ stations in Newfoundland, where two are located in the Exploits catchment and two in the Humber catchment. The hydrology of the catchment of the Humber River is influenced by the large lakes Deer Lake, Grand Lake, Sandy Lake and Hindis Lake. The water flow from Hindis to Grand Lake and to Deer Lake is regulated including the Deer Lake Power Plant (130 MW). Water flow from the Upper Humber area to Reidville is more or less natural.

The hydrology of the Exploits River catchment is strongly influenced by the Red Indian Lake (537 km2), dividing the catchment into an upper and a lower part. The release of water from the lake at the Millertown Dam is very dynamic (55–700 m3/s) and can be assessed from a downstream station.

In the framework of the demonstration project, we provided information gained from the SnowSense in situ stations and the combined EO and modelling approach as service to two dedicated demo users. Both demo users, namely Nalcor/NL Hydro and the Water Resources Management Division in the Department of Municipal Affairs and Environment (WRMD), are active in hydropower and flood forecast. During the project demonstration phase, they accessed, monitored and investigated the SnowSense products based on their specific requirements, whereof we present a selection of the results in this paper. For the two catchments as test sites, we set up the combined approach of in situ measurements, EO, and modelling.

#### 2.2.2. The Forêt Montmorency NEIGE Site near Quebec

The Forêt Montmorency is located 70 km north of the city of Quebec and is the largest teaching and research forest in the world with an area of 412 km2. It is open to the public and the University Laval operates in this forest the NEIGE test site (673 m a.s.l.). The Forêt Montmorency is drained by the Montmorency River and one of its tributaries, the Black River. The annual rainfall exceeds 1500 mm and in winter average snowfall is above 6 m.

As we had no possibility to validate SWE and LWC in the Humber and Exploit catchments in Newfoundland, we additionally had the opportunity to install a SnowSense in situ station at the study site NEIGE at Forêt Montmorency. At this location, numerous snow and meteorological instruments are tested and in operation. This encompasses, for example, the CS725 sensor, which was developed by Hydro Québec (Canada) in collaboration with Campbell Scientific (Logan, UT, USA) and which was used for comparison with the GNSS in situ measurements in this study [6]. Moreover, manual snow pit measurements, taken on a weekly basis, were available as further validation information on SWE. The CS725/GMON sensor was located 25 m next to the GNSS sensors, and the manual measurement are performed within a distance of less than 150 m. The site is a flat, sand/gravel surface area without significant vegetation. The close-by meteorological station site is in a distance of less than 500 m with an elevation difference less than 25 m. Since 2014, the University Laval is managing the NEIGE site, which is following the World Meteorological Organisation's (WMO) Canadian Solid Precipitation Experiment's (C-SPICE) objectives for solid precipitation measurements and comparisons of automatic instruments. During the winter season 2017/2018, several sensors and manual measurements from nine partners altogether took place at this location, which was a profound basis for comparing the results of our GNSS in situ station with other sensors.

#### *2.3. In Situ Station Design and Setup*

For the in situ component, we made use of information gained by signal attenuation and signal delay of GNSS signals passing through the snowpack to derive the snow cover properties SWE and LWC [8]. In general, it is possible to use any kind of GNSS signals such as the signals of the Global Positioning System (GPS), Galileo, GLONASS or Beidou for this approach. For this study, we used the freely-available GPS L1-band signals at a frequency of 1.57542 GHz. Recently, the additional usage of Galileo signals was integrated and will be applied in further studies [26]. The in situ GNSS sensor setup is in general composed of two static antennas as presented in Henkel et al. [11] and Koch et al. [8], whereof one antenna (GPS1) was mounted on the top of a mast system to be permanently above the snow cover, and the other one (GPS2) was placed on the ground before the first snow fall. In winter, when GPS2 was covered by a layer of snow, the received GPS signals were markedly influenced by signal attenuation and time delay within the snowpack. For further information on the setup, the algorithms, the detailed GNSS signal processing steps as well as their validation at a high-alpine study site in Switzerland, we refer to [8,10,11].

A large focus regarding the development and the design of the SnowSense in situ sensor system for the study presented in this paper was to provide a practical and lightweight solution for the retrieval of continuous snow cover properties like SWE and LWC. The setup is in particular suitable for locations, which were so far not equipped with measurement sensors or which were difficult to access and not regularly visited by field observations. Having this purpose in mind, we designed, produced and tested an easily transportable solution consisting of a mast system with rigging. Regarding the deployment of the sensor system, it is practically possible to carry all sensor components in a large backpack, and two persons can install it within approx. two hours. Furthermore, the sensor system consists of two multi-GNSS sensors (u-blox, Thalwil, Switzerland), which receive GPS and Galileo signals, a processing unit capable for on-board processing of SWE and LWC as well as a power and a communication managemen<sup>t</sup> unit. The latter two units are described in detail in [26]. The on-board processing is more power consuming than simply storing and transmitting raw data measurements. However, we promote the on-board processing to reduce the amount of data to be transmitted and to provide real-time data directly from the station to the client. The autonomous power supply for the operation and processing is provided by a combination of a solar panel and a battery pack and is controlled by the integrated power unit. As in many remote regions like in Newfoundland, the mobile network is not available, and we used an Iridium satellite communication (SatCom) module to transmit a daily or even sub-daily overview of the processed data. However, for regions with mobile networks, the communication unit can also be replaced by this technique. Figure 3a,b shows examples of some station components like the low-cost GNSS antennas and receivers as well as the microcontroller used for processing. Figure 3 gives an example of an installed sensor system and the solar panels ona3m high stable aluminum mast consisting of three pluggable segments.

**Figure 3.** (**a**) demo suitcase with the applied low-cost LEA-M8T GPS receivers (u-blox, Thalwil, Switzerland), u-blox patch antennas and a small microcontroller and data storage; (**b**) examples of GPS antennas of GPS1 mounted at a pole and of GPS2 laid out on the ground before snowfall; (**c**) example of a SnowSense station including a mast system, a self-supplied power managemen<sup>t</sup> (solar panels on top), a processing unit (including battery pack), and a communication module installed near Millertown, Newfoundland, Canada. GPS1 and the communication antenna are mounted at the top of the mast and GPS2 is covered by snow.

#### *2.4. EO and Data Processing*

For wet-snow detection from space, we used Sentinel-1A and -1B EO data in interferometric wide (IW) swath mode. IW mode is the main acquisition mode over land and satisfies the majority of service requirements. It acquires data with a 250 km swath at 5 m by 20 m spatial resolution [27]. The satellite data is automatically downloaded from Copernicus Data Hub (https://scihub.copernicus.eu/), and pre-processed using ESA Sentinel Toolbox for geometry and terrain corrections and it is spatially resampled on a 100 m by 100 m grid [28]. Afterwards, the data is radiometrically calibrated and corrected with regard to the elevation angle [29]. The wet-snow mapping is performed as proposed by Nagler and Rott [30] with a fixed threshold of -3 dB from a reference scene. Here, we are using a predefined reference data set with averaged standard backscatter values for different satellite geometries according to Appel et al. [29].

The results are binary wet snow maps with a resolution of 1 km by 1 km, containing three classes: wet snow, no information (which includes dry snow or snow free conditions), and no data. The maps were used in a further step for the assimilation process (see Section 2.6) and as additional visualization product, i.e., provided for the demo users in a password protected online portal (see Figure 4).

**Figure 4.** (**a**) Sentinel-1 IW mode composite for the Island of Newfoundland for 30 April 2017; (**b**) derived map of wet snow area (blue) from Sentinel-1 data.

#### *2.5. Processes of Mass and Energy Transfer—PROMET Model Component*

For the retrieval of the spatially distributed snow information and forecast capabilities, the hydrological part of the coupled land surface process model PROMET [22] was chosen. It is a raster-based model that has been developed to spatially simulate the elements of the land surface water balance including vegetation, soil moisture, snow cover, reservoirs, and river discharge. PROMET incorporates a four-layer soil model and considers lateral flows along hill slopes. Due to its raster-based architecture, the model allows for the assimilation of remote sensing data for distributed hydrological applications [31]. PROMET solves the water and energy balance for hourly time steps and calculates the run-off of river basins, while it strictly conserves mass and energy. It thus allows for the validation of the complete process chain, from rainfall over soil-moisture dynamics to vegetation controlled evapotranspiration and finally routed run-off, against measured discharges on the basin and sub-basin scale.

PROMET has successfully been applied for a variety of hydrological studies in medium-to large-sized watersheds [22,31–33]. Required input data consist of raster-based GIS information, characterizing the spatial distribution of land use patterns, soil types and topographic features. In addition, parameters describing the characteristics of the soil and land use or crop categories are supplied through tabular input.

For the application on the island of Newfoundland, we set up a basic model environment based on a freely available digital elevation model (DEM), from which the river network was automatically derived and lakes or reservoirs were inserted as structures. Land use information was derived from EO data sources (GlobCover), as soil map the Harmonized Soil Map of the World [34] is used. All spatial input data as well as the model grid itself have a resolution of 0.00833 degrees (<1 km2). The parameterization of the catchment characteristics followed the internal standardized procedure developed by VISTA, purely based on GIS and EO data. As the PROMET model is physically based, it does not require a calibration. Further fine-tuning of parameters would be possible, but were, however, not conducted due to missing information e.g., on reservoir managemen<sup>t</sup> and limited validation data.

PROMET can be driven by high resolution (1 h) time series of meteorological station network records, using the measured meteorological parameters. However, due to the specific characterization of our study area, only a sparse meteorological station network was available, which was not sufficient for our service target. Moreover, even in case of a sufficient number of meteorological stations, the amount of continuously provided information, e.g., on radiation and humidity or dew point are often missing or inaccurate. To overcome the lack of meteorological station data, we used global and regional climate model data as input, similar as presented by Zabel [35]. For the application in Newfoundland, we used data provided by Government of Canada as part of their High Resolution Deterministic Prediction System (HRDPS). This system is a set of nested limited-area model (LAM) forecast grids from the non-hydrostatic version of the Global Environmental Multiscale (GEM) model with a 2.5 km horizontal grid spacing.

The fields in the HRDPS high resolution data set are made available as GRIB2 for four times a day for the Pan-Canadian domain for a 48-h forecast period [36]. Those data sets are consistent in temporal (1 h) and spatial (2.5 km) resolution. We adapted the import routines for the meteorology and selected the appropriate parameters for the model forcing to perform calculations of the snow and run-off situation with hourly temporal resolution. Due to the architecture of PROMET, we are able to calculate any periods, starting from previously stored status and recovery files, and provide all information on the current snow and hydrological situation as spatial data sets in a 1 km by 1 km resolution with an hourly timestep and as tabled point data.
