Accurate and reliable solid precipitation estimates for high mountain regions are crucial for many research applications. Yet, measuring snowfall at high elevation remains a major challenge. In consequence, observational coverage is typically sparse, and the validation of spatially distributed precipitation products is complicated. This study presents a novel approach using reliable daily snow water equivalent (SWE) estimates by a cosmic ray sensor on two Swiss glacier sites to assess the performance of various gridded precipitation products. The ground observations are available during two and four winter seasons. The performance of three readily-available precipitation data products based on different data sources (gauge-based, remotely-sensed, and re-analysed) is assessed in terms of their accuracy compared to the ground reference. Furthermore, we include a data set, which corresponds to the remotely-sensed product with a local adjustment to independent SWE measurements. We find a large bias of all precipitation products at a monthly and seasonal resolution, which also shows a seasonal trend. Moreover, the performance of the precipitation products largely depends on in situ wind direction during snowfall events. The varying performance of the three precipitation products can be partly explained with their compilation background and underlying data basis.
Observations of solid precipitation (snowfall) at high elevations are important for many fundamental and applied scientific questions in, for example, glaciology, hydrology and water resource management
This study presents a novel approach using temporally continuous and reliable ground observations taken on two alpine glaciers in Switzerland. With daily snow water equivalent (SWE) observations by a cosmic ray sensor (CRS), which has previously been evaluated with promising results
The investigated precipitation products are based on (i) a dense gauge network (RhiresD), (ii) a ground-based weather radar network combined with the gauge network (CombiPrecip) and (iii) the analysis of the high resolution non-hydrostatic numerical weather prediction model COSMO-1. They are all operationally compiled by MeteoSwiss, readily available and have a high temporal (daily) and spatial (
Gridded precipitation products are evaluated by their agreement of snowfall amounts compared to the in situ measured new SWE amounts at a seasonal and monthly resolution. Moreover, prevailing in situ meteorological conditions such as wind direction, wind speed and air temperature are used to characterize snowfall events, which are over- or underestimated by the precipitation products.
The two glacier sites of this study are located in different parts of the Swiss Alps (Fig.
Shown are the location and outlines of Plaine Morte and Findelen glaciers
On Plaine Morte (46.38
Some technical issues on Plaine Morte resulted in several data gaps of temperature, humidity, wind speed and snow depth, which were substituted with measurements from nearby highly correlated stations
The gridded precipitation data (RhiresD) is compiled using approximately 420 quality-controlled gauges of the Swiss Federal observation network. The gauge network consists of heated pluviometers with a rocking mechanism (1518 H3 and 15188 Lambrecht) and with a weighing mechanism (Pluvio1 and Pluvio2 by Ott). The pluviometers are not surrounded with a wind shield and measure with a 10 min frequency and a minimum amount of 0.1 mm
The gridded precipitation product RhiresD is a result of an interpolation based on a climatological mean and a regionally varying precipitation-topography
Switzerland has an operational ground-based weather radar network consisting of five polarimetric C-band Doppler radars
As shown in
The numerical weather prediction model COSMO has been developed by the consortium for small-scale modeling, a collaboration of national weather services of several countries
From each precipitation product, precipitation estimates of the grid cell closest to the two automatic weather stations is extracted over time
The CRS provides continuous hourly observations of SWE. To increase the precision of the instrument and to reduce the noise, the hourly observations are averaged over 24 h centered around 06:00 UTC. The time of 06:00 UTC was chosen to be coherent with the RhiresD data set. Because the CRS provides continuous data on the snowpack, the time series needs to be broken down to obtain daily changes of SWE. Therefore, the difference between 06:00 UTC of
In further processing (see Sect.
The precision is derived by error propagation. All constant parameters and continuous observations (incoming cosmic ray flux, air pressure and neutron counts) are assigned with an estimated uncertainty, which is propagated through the non-linear equations that correct the raw neutron counts and convert them to SWE. The dominating source of uncertainty arises from the statistical error of the neutron count itself. The statistical error follows a Poisson distribution, where lower neutron counts (higher SWE) result in a lower precision
The precision of the CRS allows excluding noise. However, true events can also be lost, especially in deep snowpacks. In addition, random false signals may not be excluded by the precision estimate. To best-possibly assure that we only include daily changes of SWE that are related to precipitation events, a double conditional bias is introduced in Sect.
An integrated end-of-season bias is derived using cumulative precipitation and SWE observations at the end of the accumulation season, typically around 30 April. The integrated bias is motivated by the fact that reliable SWE observations in high mountain regions are often limited to a few measurements in time, sometimes one manual observation per season only. These observations are temporally limited because of the logistical and financial efforts needed to obtain them in high mountain regions. The temporal evolution of SWE is then inferred by using precipitation amounts from nearby stations at lower elevations or from precipitation data sets. Nonetheless, the end-of-season measurement of the snowpack is an integrated observation, i.e., apart from precipitation, it also includes effects of wind redistribution, sublimation, and potentially mass loss by early snowmelt. In consequence, several types of errors are introduced into such a time series. The comparison of an integrated end-of-season bias (
A seasonal end-of-season integrated bias is computed as
In consequence of the time period definition, each winter season is based on a different number of days (Table
On Plaine Morte, no observations were available in October and November 2017. Hence, the onset of the snowpack could not be estimated and is defined as the date when measurements restarted. Another technical issue interrupted SWE observations on 26 April 2018, further shortening the 2017/18 winter season on Plaine Morte.
Onset of the snowpack, number of days from the onset until 30 April (26 April for winter 2017/18 on Plaine Morte) and end-of-season amounts for the stations on Plaine Morte and Findelen. The end-of-season amounts are based on in situ SWE observations (SWE
Additionally, we derive an overall integrated bias by dividing the sum of all end-of-season cumulative precipitation amounts by the sum of all SWE measurements obtained at the end of the season as
To identify environmental conditions that result in a smaller or larger integrated end-of-season bias, daily air temperature (mean, minimum and maximum), wind speed (mean), wind gust (mean and maximum) and relative humidity (mean, minimum and maximum) are averaged over the same accumulation season (Table S1 in the Supplement). Seasonal averages of both glaciers and all winter seasons (six observations) are correlated with the end-of-season integrated bias of each winter season with a Pearson correlation and a significance level of 0.05. Moreover, these seasonal averages are also correlated with the relative difference between the integrated and conditional biases (Sect.
A conditional bias is calculated for each winter seasons and month during the whole time period with available in situ observations on the two glacier sites. To analyse the seasonality of the bias, all months of all winter seasons are aggregated to derive the conditional seasonal monthly bias. Moreover, all snowfall event-days are categorized by wind direction measured on the glacier site. For each wind direction category, a bias is calculated.
The conditional bias (
The threshold applied to daily SWE amounts (
An event-day is defined with
The event-days are always defined in the same way for all conditional bias, independent of temporal resolution. Nonetheless, they vary among the precipitation data sets. The difference between the conditional biases lies within the aggregation of these event-days, which is either by the accumulation season, the individual months, the month of the year (aggregated) or the wind direction.
For the end-of-season conditional bias, event-days within the accumulation season, i.e., from onset of the snowpack to 30 April of each season (Table
In the monthly conditional bias, the number of contributing event-days varies with month, season and glacier. On Plaine Morte (Findelen), 25 (14) months with measurements are available. To render the correlations of the monthly bias with daily air temperature, air humidity and wind speeds more robust, only months with more than two event-days are considered. This excludes December 2016 and April 2018 on Plaine Morte and January 2020 on Findelen. The conditional seasonal monthly bias is derived by aggregating all event-days by the month of the season. For example, the seasonal monthly bias of December is derived by all event-days from December 2016–2019 on Plaine Morte. It is a more robust estimate of the bias compared to averaging the monthly bias over all years.
In addition, we analyse the influence of wind direction by categorizing wind direction into four main sectors; North, East, South and West. Each wind sector has a size of 90
The uncertainty of all conditional biases are estimated by a leave-one-out cross validation with the mean absolute difference to the reference value. This uncertainty estimate, however, does not include a constant bias of the CRS nor the error arising from comparing a point estimate to a grid cell estimate.
The end-of-season precipitation biases vary over the different winter seasons, the two glacier sites and the analysed precipitation products (Fig.
End-of-season bias derived with the accumulated (integrated) SWE amounts and cumulative precipitation during the accumulation season (
The integrated and conditional biases are highly correlated for
The relative difference between the integrated and conditional bias is significantly correlated with air humidity for
The difference between the integrated and conditional overall end-of-season bias is sensitive to the threshold applied. In the following, we refer to the reference threshold as the one in the main analysis (
With a high threshold of
In conclusion, if only one observation in time is available and the goal is to infer snow accumulation at a higher temporal resolution, adjusting precipitation amounts with the integrated end-of-season bias may be justified and has shown good results
Panels
At monthly resolution, precipitation data sets and in situ SWE observations on Plaine Morte agree fairly well with
In general, the performance on Plaine Morte is better than on Findelen, partly due to the following reasons. A weather radar is located next to Plaine Morte and it is surrounded by several precipitation gauges in all four cardinal points. On Findelen, in contrast, precipitation gauges are only located in the west of the glacier. In addition, the quality of weather radar estimates is lower due to the following reasons. The larger distance between radar and target site results in more residual ground clutter and a poorer resolution caused by beam broadening. Most echoes on Findelen are rejected because they are considered to be ground clutter contaminated
All precipitation data sets have a common trend throughout the accumulation season on both glaciers (Fig.
On Plaine Morte, more event-days are recorded from December to March than in October, November and April (Fig.
An important aspect to note is the correlation of the precision of the CRS with the depth of the snowpack (Sect.
Figure
Conditional bias based on event-days and derived for the four main wind sectors on
Re-analysed precipitation (COSMO-1) has most inaccurate precipitation estimates with easterly winds on Plaine Morte (Fig.
On Findelen, the performance of each precipitation data set is also related to the wind sectors (Fig.
It is also interesting to note that
The bias aggregated by wind direction shows some sensitivity to the applied threshold (see Sect.
Correlating the bias based on monthly estimates (Fig.
Generally, the dependence of the performance on synoptic situations indicated by in situ wind direction during snowfall events is in line with
A recent study by
Given the challenging environmental conditions for direct observations of SWE and/or snowfall on the two glacier sites of this study, the CRS provides the highest possible data quality and reliability. Limitations of the CRS observations are given by the noise, which depends on the depth of the snowpack, and the continuous cumulative time series, which needs to be transformed to an instantaneous measurement coherent with the precipitation estimates. This transformation is not straightforward because of the noise included in the observations. To increase confidence within the daily continuous estimates before the daily changes are calculated, we derive a daily average centered around 06:00 UTC (Sect.
An aspect that needs to be addressed is the sensitivity of these results to the thresholds used to calculate the double conditional bias. The threshold applied for the main result (
Another important limitation is the spatial representativeness of these results. Only two sites with point observations are available for this study and they significantly differ with respect to topography and the regional climatic conditions. Furthermore, we compare point estimates to a grid cell of at least
Moreover, the assessment of the bias between in situ SWE observations and precipitation will most likely not apply to non-glacierzied mountain sites, in particular to wind-blown crests. Glaciers typically form in places that are either snow-rich or cold throughout the year, and snow accumulation lasts several summer seasons. In addition, CombiPrecip-adj only exists for selected glacier sites
A wide range of studies focus on improving precipitation estimates in high mountain regions using various approaches, in situ observations and/or geostatistical interpolation methods. However, validating solid precipitation estimates in high mountain regions remains a challenge and typically includes large uncertainties
From this study, we draw the following conclusions.
The end-of-season integrated and conditional bias (based on event-days) differ from each other, underlining the influence of other processes on the snowpack. It also implies that on event basis, the bias is generally higher when also including light precipitation events. The bias has a large variability at a monthly resolution with a strong seasonal trend. Depending on the data product, i.e., whether it is based on interpolated direct measurements, remotely-sensed or modelled, and its underlying observations, under- or overestimation of in situ SWE observations can be partly explained by air temperature, wind speeds and wind directions during snowfall events.
Despite the spatial limitations of this study, the variation of the bias at different temporal resolutions has important implications for hydrologists, glaciologists and meteorologists. It shows that the compilation background and data source of precipitation data sets play a major role in how the data performs. Hence, no single best precipitation data set can be announced. Yet, depending on the application and the target site, one or the other data set may be more accurate or suitable.
In future applications, it needs to be noted that precipitation data sets perform differently depending on their data source. Having a ground reference is crucial to adjust precipitation estimates, but adjusting precipitation amounts with a temporally constant factor may also introduce large uncertainties in the evolution of cumulative precipitation in high mountain regions. Nevertheless, the data products presented in this study have great potential for application in hydrological and/or glaciological studies. In particular, algorithms to process radar-gauge data and numerical weather modelling may benefit from the findings of this study. Further studies are needed to investigate the link between snowfall and snow accumulation in more depth, especially with more in situ observations.
The precipitation products are readily-available from MeteoSwiss. The data of the two automatic weather stations will be available in a future repository.
The supplement related to this article is available online at:
RG prepared the manuscript and performed data analysis with contributions from all co-authors. MGu helped with data processing. MGa, MH and NS contributed to the design and execution of the study.
The authors declare that they have no conflict of interest.
This article is part of the special issue “Applied Meteorology and Climatology Proceedings 2020: contributions in the pandemic year”.
The Federal Office for Meteorology and Climatology MeteoSwiss kindly provided the RhiresD, CombiPrecip and COSMO-1 data. We are grateful to all who helped with installation and maintenance of the automatic weather stations. Moreover, we acknowledge the NMDB database (
This research has been supported by the Swiss National Science Foundation (SNSF) (grant no. 200021_178963).
This paper was edited by Renato R. Colucci and reviewed by two anonymous referees.