*2.6. Assimilation*

In general, and as, e.g., presented in [37–42], data assimilation is likely to improve the accuracy of the modelled snow parameters in a way that the modelled snow parameter or run-off results agree better with the real situation. For the SnowSense data assimilation, we used information gathered by the SnowSense GNSS in situ stations and the already existing online SWE measurement sites, as well as information from EO data. Regarding the in situ locations, we used the SWE and LWC (were available), regarding the EO sources, we used binary information whether the snow is dry or wet from Sentinel-1. All information was then compared twice a week with the non-assimilated output of SWE, LWC and the snow cover extent simulated by the PROMET model. In case of significant differences between simulated and observed values, assimilation runs for the last calculation period had been triggered. To adjust the model output to the observed in situ and EO information, the following options were available [43]: Adjustments of (i) the precipitation gain for the amount of snow/solid precipitation, (ii) the critical temperature for the transition of liquid and solid precipitation, and (iii) the albedo parameter for the short wave energy fluxes controlling the ablation dynamic. The sequence of the assimilation steps is following a decision tree based work-flow. After adjusting the model parameterization, the model is re-calculating the required output parameters for an improved agreemen<sup>t</sup> between model and reality. The adjustment of i,ii and iii is performed spatially. During the winter 2017/2018 phase, the assimilation process was performed partly manual. For a future service application, we will establish a more automated process.

**Table 1.** Statistical comparison of SWE derived by GNSS, two GMON sensors (CS725\_K and CS725\_TL),

and

#### **3. Results and Discussion**

In the first part of this section, we present the results and the validation of the GNSS in situ station for the well-equipped study site NEIGE at Forêt Montmorency. In the second part, we focus on the results and the validation of the hydrological service of the combined and assimilated run-off at, in total four, gauges of the Humber and Exploit catchments in Newfoundland.

#### *3.1. Station Performance at the Forêt Montmorency NEIGE Site near Quebec*

The SnowSense station was installed in October 2017 at the NEIGE site at Forêt Montmorency and was operational during the entire snow-covered winter season 2017/18. The power supply and the communication unit resisted the cold and windy environment without damage or failure. The station delivered continuous daily SWE and LWC measurements without any significant discontinuities (Figure 5a).

In general, the SWE values derived by the SnowSense GNSS in situ station are in good agreemen<sup>t</sup> with the provided reference measurements by the two GMON CS725 sensors and the manual snow pit measurements (Figure 5b).

The GMON sensor as well as the GNSS sensor are both non-destructive measurement methods and are largely capable of deriving SWE. Both sensors were already validated against other sensors like snow pillows and manual measurements (e.g., [7,11,44]). As stated by Choquette et al. [6], at the observed study site NEIGE, an average error of 18% between manual measurements and the GMON sensor is reasonable, and, for SWE levels less than 400 mm w.e., the estimation is inside the 5–10% range.

At the time the GNSS measurements started in early winter 2017, the GNSS-derived SWE as well as the measurements of the two GMON sensors lay in a similar range at approximately 100 mm w.e. At the end of April 2018, the maximum amount of SWE with approximately 500 mm w.e. was reached, which was indicated by the GNSS solution as well as the manual measurements and one GMON sensor. Comparing the two GMON sensors (blue solid and blue dashed line, Figure 5b), however, an offset of the SWE measurement of up to 30% occurred between them. This offset was low in the beginning of the time series and increased during the winter continuously. The reasons for this might originate in different sensor locations e.g., with slightly different wind conditions, and are still under discussion, but are out of the scope of this paper. The manual measurements were performed 16 times on a weekly basis during the snow season. On each day, three to four snow pits were analyzed. The resulting averaged SWE measurements from the snow pits lie well in between the range of the two GMON sensors. The SWE results of the SnowSense GNSS station (black solid line) follows in the beginning of the season the lower GMON sensor (CS725\_TL, blue dashed line). Since 20 February, after a heavy precipitation event, the GNSS derived SWE jumps to the level of the GMON sensor with the higher SWE values (CS725\_K, blue solid line). The coefficient of determination (**R**2) between the GNSS data and the data from the GMON sensors is 0.53 for CS725\_TL and 0.93 for CS725\_K, respectively.

Throughout the entire season, the GNSS measurements agree very well with the manual measurements from snow pits including their minimum and maximum values. Here, the coefficient of determination (**R**2) is 0.64. For all **R**<sup>2</sup> and root mean square errors (RMSE) errors of the validation study, we refer to Table 1.



**Figure 5.** (**a**) daily GNSS-derived SWE and LWC values at the Forêt Montmorency NEIGE site 2017/2018; (**b**) daily GNSS-derived SWE values and reference measurements from two GMON CS725 sensors and manual snow pit measurements of mean, minima and maxima SWE at the Forêt Montmorency NEIGE site.

The range of the SWE validation results presented in this study for the Forêt Montmorency are in good accordance with previously conducted studies validating the GNSS in situ SWE component successfully at the high-alpine study site Weissfluhjoch in the Swiss Alps as presented in Henkel et al. [11] and Koch et al. [8]. One key result of their validation is a very good performance of the SWE determination capabilities of the in situ GNSS approach: the inter-comparison of the SnowSense GNSS station with a snow pillow and manual measurements show a very high agreemen<sup>t</sup> indicated by the values of **R**<sup>2</sup> close to 1 [11]. The root mean square errors (RMSE) between the measurements from GNSS, snow pillow, and from snow pits fit very well; they lay in the range of 11–24 mm w.e. [11] but might be higher regarding wet snow conditions, e.g., due to an increase in GNSS signal attenuation (approx. 45 mm w.e.) [8].

Until mid-April 2018, the snowpack was predominantly dry at the study site Forêt Montmorency. In January, February and March, only single wet-snow events of up to two to three days occurred. The occurrence of wet snow presented by the GNSS-derived LWC is shown in Figure 5a. As no further sensors for a comparison with the GNSS-derived LWC were available at this site, we compared the GNSS-derived LWC with meteorological parameters. In general, the measured LWC is in a good temporal correspondence with rainfall events and goes along with warm air temperatures (Figure 6), which was also demonstrated in Koch et al. [10]. However, as shown in other previous studies, we were able to successfully validate the GNSS LWC measurements with other sensors like an upward-looking ground-penetrating radar and capacity probes at the study site Weissfluhjoch [45].

**Figure 6.** Daily GNSS-derived LWC values at the Forêt Montmorency NEIGE site for the winter season 2017/2018 related to (**a**) rain and snowfall events and (**b**) air temperature development. Temperature and precipitation measurements were conducted at the close-by provincial and federal station sites.

#### *3.2. SnowSense Service for the Island of Newfoundland*

For the entire SnowSense service based on the GNSS in situ measurements, EO and modelling, spatially-distributed maps of the SWE were generated and run-off estimates were derived for several river locations in the two main catchments Humber River and Exploit River. Figure 2 gives an example of a SWE map, which was modelled and assimilated for the entire island of Newfoundland for the 15 March 2018. Such a SWE map was provided for each day.

The simulated and assimilated run-off results were validated against four run-off gauge measurements in total. Measurements (raw data) are made available by the Water Resources Management Divisions. As for the catchments of the Humber River and the Exploits River, a strong interest in flood forecast was predominating; the run-off results are presented for the gauge stations at the lower parts of the two catchments.
