Forecasting avalanche danger at a regional scale is a largely data-driven yet also experience-based decision-making process by human experts. In the case of public avalanche forecasts, this assessment process terminates in an expert judgment concerning summarizing avalanche conditions by using one of five danger levels. This strong simplification of the continuous, multi-dimensional nature of avalanche hazard allows for efficient communication but inevitably leads to a loss of information when summarizing the severity of avalanche hazard. Intending to overcome the discrepancy between determining the final target output in higher resolution while maintaining the well-established standard of assessing and communicating avalanche hazard using the avalanche danger scale, avalanche forecasters at the national avalanche warning service in Switzerland used an approach that combines absolute and relative judgments. First, forecasters make an absolute judgment using the five-level danger scale. In a second step, a relative judgment is made by specifying a sub-level describing the avalanche conditions relative to the chosen danger level. This approach takes into account the human ability to reliably estimate only a certain number of classes. Here, we analyze these (yet unpublished) sub-levels, comparing them with data representing the three contributing factors of avalanche hazard: snowpack stability, the frequency distribution of snowpack stability, and avalanche size. We analyze both data used in operational avalanche forecasting and data independent of the forecast, going back 5 years. Using a sequential analysis, we first establish which data are suitable and in which part of the danger scale they belong by comparing their distributions at consecutive danger levels. In a second step, integrating these findings, we compare the frequency of locations with poor snowpack stability and the number and size of avalanches with the forecast sub-level. Overall, we find good agreement: a higher sub-level is generally related to more locations with poor snowpack stability and more avalanches of larger size. These results suggest that on average avalanche forecasters can make avalanche danger assessments with higher resolution than the five-level danger scale. Our findings are specific to the current forecast set-up in Switzerland. However, we believe that avalanche warning services making a hazard assessment using a similar temporal and spatial scale as currently used in Switzerland should also be able to refine their assessments if (1) relevant data are sufficiently available in time and space and (2) a similar approach combining absolute and relative judgment is used. The sub-levels show a rank-order correlation with data related to the three contributing factors of avalanche hazard. Hence, they increase the predictive value of the forecast, opening the discussion on how this information could be provided to forecast users.
In many snow-covered mountain regions, avalanche forecasts are disseminated to the public to inform and warn about avalanche conditions. The provision of these warnings to the public consists of two steps: first, a prediction of the avalanche hazard is made, and, second, the prediction is communicated in a forecast product.
Assessing and forecasting avalanche hazard is a largely empirical process in which a human forecaster analyzes and interprets data to make an informed judgment regarding current or expected avalanche conditions The
Once all relevant avalanche problems have been identified, their location and temporal occurrence specified, and their character described, avalanche hazard is summarized in regional avalanche forecasts using one of five danger levels (see Fig.
An increased level of detail may include, for instance, decomposing the judgmental forecasting process and specifying each of the individual components relevant for the final hazard assessment
Unfortunately, addressing this question is not straightforward as avalanche danger and, hence, the sub-levels cannot be measured. However, since the danger levels represent a rank order in terms of the severity of the avalanche conditions, we tackle this question using a comparative approach testing whether there is a positive monotonic correlation between the sub-levels assigned to danger levels and data describing the three contributing factors of avalanche hazard. Specifically, we investigate whether there is a rank order relationship between the data and the sub-levels. For this, we make use of both observational data collected for the purpose of avalanche forecasting in Switzerland and independent data sources not used in the forecasting process: the output from two recently developed models
We first determine for each parameter in what range of the danger scale it correlates with the forecast danger levels ( Does a data source representing a contributing factor of avalanche hazard correlate with the danger level For the range in the danger scale determined in (1), is there a monotonically increasing correlation between the parameter representing a contributing factor and the sub-levels
The Swiss avalanche forecast has previously been described in several publications
Maps of Switzerland showing
During winter, the national avalanche warning service at the WSL Institute for Snow and Avalanche Research SLF (SLF) publishes an avalanche forecast at 17:00 LT (local time), valid until 17:00 LT the following day (see example in Fig.
The production of the forecast always starts with the assessment of the current avalanche conditions. Numerous data are used in this process. These include measurements from automated weather stations located at the elevation of potential avalanche starting zones
The Swiss avalanche forecast describes regional avalanche conditions. The average size of the almost 150 warning regions, the smallest spatial units used in the forecast, is about 200
In the following, we refer to the danger levels (
Note that the criteria to distinguish between sub-levels and the range covered by a sub-level within a danger level remained undefined. Furthermore, forecasters made no such differentiation for 1 (low), as a further distinction within this level seemed impossible. In addition, an internal analysis of qualifiers assigned to danger levels describing wet-snow conditions showed that forecasters primarily assigned the sub-level
Overview showing the analyzed data sources and the contributing factors of avalanche hazard (snowpack stability, the frequency of snowpack stability, avalanche size) for which we consider the respective data sources to be a proxy.
We analyzed observational data collected as part of our operational avalanche forecasting (Sect.
In the following, we describe the data and their preparation for this analysis.
We extracted the forecast danger level, the unpublished sub-level, and the critical aspects and elevations, referred to as the
The occurrence of avalanches directly indicates instability
In Switzerland, about 80 “stationary” observers report avalanches in their region on a daily basis. Observers report avalanches either individually or by aggregating avalanches into an avalanche summary report. In addition to avalanches regularly reported by these observers, field observers, who are also part of the observer network, and the public may report avalanches. Reported avalanche properties include the location and the estimated time of the release, the avalanche size (size classes 1 to 5 according to
Whumpfs, a sudden, collapse-type failure of a weak layer due to rapid localized loading
When reporting their observations after a day in the field, observers also report whether they observed human-triggered whumpfs and shooting cracks and how frequent these danger signs occurred using three classes (DS.class):
We extracted all observations which were reported after a day in the field. This resulted in 5996 observations.
Information on snowpack stability can also be obtained by digging a snow pit and performing a stability test. These tests primarily provide very localized information on snowpack stability. Therefore, to obtain information on the frequency distribution of snowpack stability, numerous tests must be performed on the same day and in the same region
Classification of stability tests:
In Switzerland, two stability tests are performed regularly by observers to assess snowpack stability: the Rutschblock test
In total, 2201 RB and 2261 ECT were available. Their spatial distribution is shown in Fig.
Recently,
We relied on an updated version of the data set used by
In addition to observational data, we analyzed the output of two recently developed random forest classifiers predicting the danger level
The first model, which we refer to as the
The second model developed by
Workflow: preparatory steps (steps 1 to 3) and analysis to answer research question 1 (step 4) and 2 (steps 5 and 6).
We linked the forecast with the observations and the model output by their location and calendar day.
For this analysis, we distinguished between
data sources which mostly included only a single data point or even no data at all per forecast danger region (Fig. data which allowed the calculation of a proportion or a mean for each forecast danger region (Fig.
The first group included observations of danger signs, stability test results, and the accident and movement points, while the second group contained observations of natural and human-triggered avalanches and the predictions of the two models.
For the data sources with sufficient data points per danger region, we defined the following parameters that summarize the observations or modeled output for a given danger region (i.e., for the danger region A in Fig.
Furthermore, for each danger region, we derived an avalanche activity index (AAI) relative to
In a second step, for each danger region with the same forecast
Danger levels are rank ordered. The absolute increase in danger from one danger level to the next is unknown. To derive the expected danger rating
the proportion the accident–movement point ratio ( The occurrence of natural avalanches of increasing size is a key criterion defining the higher danger levels in the avalanche danger scale For data which rely on a human being present in avalanche terrain, we combined the (few) cases at 4 (high) and 5 (very high). At these danger levels, travel in avalanche terrain is strongly reduced due to dangerous conditions leading to a strong reduction in observational data. For each of the two models, we combined the predictions at 4 (high) and 5 (very high) as the models relied on training data merging these two danger levels
Not all the data sources describing the contributing factors are equally suitable to explore differences between all the danger levels or sub-levels in the entire range of the danger scale.
Graphical representation of the critical aspects (colored black in the aspect rose, here W–N–SE) and the critical threshold elevation (here 2000
To answer research question 1 (does a data source representing a contributing factor correlate with the danger levels
To obtain a better understanding of the distribution of the samples, we calculated the bootstrap-sampled median Whenever we refer to median values, we add a tilde above the variable name, i.e.,
Finally, we calculated a factor
Three data sources (accident and movement points, danger-level model, instability model; marked with an * in Fig.
The entire analysis was performed using the software
Natural avalanche activity increased with increasing danger level (Fig.
Avalanche activity index (AAI) for natural avalanches per 1000
The increasing frequency of natural avalanche occurrence of increasing size with increasing danger level, as seen for the danger level
Spatial density of natural avalanches
Spatial density of natural avalanches
The number of human-triggered avalanches per 10 000 km
At the resolution of the forecast sub-levels, the number of human-triggered avalanches
Human-triggered avalanches are comparably rare events. This means that
Observers seldom reported human-triggered danger signs at 1 (low) with less than 1 in 22 observations. In contrast, danger signs were rather common at 3 (considerable) and 4 (high) when
As can be seen in Fig.
The accident–movement point ratio (
As shown in Fig.
The density of human-triggered avalanches (or the number relative to the area of PRA) (
In summary, a positive monotonic relationship between data related to the frequency of locations where human triggering is possible and
Proportion of Rutschblock test results (
The median proportion of Rutschblock (RB) test results related to instability,
Similar findings can be noted when analyzing the relationship between
The median proportion of ECT results related to instability increased with increasing danger level from 1 (low) to 3 (considerable) (
Analyzing the correlation between
In summary, we observed an increasing proportion of stability tests related to instability with increasing
Output from random forest models predicting the danger level
The danger rating predicted by the danger-level model showed a strong increase from 1 (low) (
Turning to
The median proportion of simulated profiles classified as unstable (
Findings were similar when exploring the correlation between
The overarching research question we explored was as follows: given the daily observations and measurements, often still incomplete at the time when avalanche forecasters in Switzerland meet for their afternoon forecaster briefing, and a numerical weather prediction model, can human avalanche forecasters forecast avalanche hazard for the following day with higher resolution than the five danger levels? To this end, we analyzed a wide variety of data related to the contributing factors of avalanche hazard and investigated their relationship with sub-levels assigned to danger levels in Switzerland. The specific question we had was therefore, given the current forecasting set-up in Switzerland, whether the sub-levels were assigned in a way that they express the expected rank-order relationship between the three contributing factors of avalanche hazard and the sub-levels. As we could not rely on a clear definition of the sub-levels, we split the analysis into two steps: first, we determined the range of the danger scale in which a given data source was valuable to distinguish between danger levels. Second, we analyzed whether a monotonic correlation between sub-levels and the data source existed.
For the first research question, we determined in which range of the danger scale a data source was suitable for our analysis. As summarized in Table
Turning to our second research question, we summarize an increase in the value of the analyzed parameters for most of the sub-level pairs (
These findings represent the average. Of course, there will be errors in both the forecast danger level (absolute judgment) and the forecast sub-level (comparative judgment). For instance, a recent study explored the agreement between danger-level assessments provided by specifically trained observers after a day in the field (local nowcasts) and the forecast regional danger level
Table summarizing whether an increase (light blue,
Forecasters felt generally comfortable assigning a sub-level in dry-snow conditions. We attribute this to the fact that forecasters must characterize the severity of the avalanche conditions to accurately describe the situation in the forecast, regardless of whether a sub-level is assigned or not. However, assigning a sub-level makes this evaluation more systematic and facilitates communication with other forecasters on duty. Forecaster feedback suggests that the additional mental effort required for the assignment of a sub-level is small and that discussing the sub-level at the forecaster briefing does not take more time than discussing any of the other elements which are communicated in the forecast products.
Our analysis showed that forecasters can estimate sub-levels in dry-snow conditions based on the available data, thus providing a way of increasing the resolution of the forecast danger level while maintaining the well-established standard of assessing and communicating avalanche hazard using the five danger levels. Moreover, the comparison with the two models not used in the forecast production process indicated that the sub-level forecasts were reasonably consistent. The models mirrored differences in the forecast danger level and the sub-level, as well as concerning aspects and elevations where the danger prevailed.
Refined danger ratings allow forecasters to express a more natural and gradual change in avalanche danger compared to the five danger levels. While models have the potential to provide continuous output, such an approach is not possible for humans. Therefore, the experts assessed avalanche danger in two stages combining an absolute and a relative judgment
While we have shown that the method of combining absolute and relative judgments can result in avalanche danger assessments with finer granularity, it might still be advantageous to describe typical characteristics for each sub-level. This may not only help forecasters when deciding on a sub-level but may potentially also be useful for users of this information. Therefore, we envision that by using the presented data, but also the actual descriptions of avalanche danger in the avalanche forecast
We have demonstrated that, on average, the forecast sub-levels have predictive value; that is, they correlate with the three contributing factors of avalanche hazard. Therefore, we argue that the sub-levels should be provided in a suitable form to forecast users as they may support the decision-making process.
We see two potential use cases. The first, more traditional use case is the provision of the sub-levels as part of the avalanche forecast product, permitting a direct interpretation of the sub-level by the human forecast user. However, as several studies have shown, the comprehension of the information communicated in the bulletin is strongly related to the education of the user and to the complexity of the avalanche situation
Second, the sub-levels could also be used for the development and validation of models. These may, in turn, improve avalanche forecasting. One such example is the danger-level model, which was trained and validated with the defined danger levels
We aimed at exploring the correlation between
Our study was set in Switzerland. While the results can therefore not readily be applied to other countries, we believe that the more general finding, namely the approach of combining absolute and relative judgments, should be applicable in other forecast settings as well.
The distribution of the data was not uniform over the entire forecast domain. For instance, hardly any data were available for the Jura or middle and southern Ticino regions (region B in Fig.
Can forecasts of avalanche danger be refined by using a combination of absolute and comparative judgments? We addressed this question by comparing 5 years of Swiss avalanche forecasts including a sub-level qualifier (comparative judgment) assigned to the danger level (absolute judgment) with several data sources considered a proxy for the three contributing factors of avalanche hazard. We have shown that, on average, these sub-levels reflect the expected increase in the number of locations with poor snowpack stability and in the number and size of avalanches with increasing forecast sub-level.
Our findings are specific to the current forecast set-up in Switzerland. However, we surmise that avalanche warning services whose hazard assessment is based on a similar temporal and spatial scale as is used in Switzerland should also be able to refine their assessments if (1) enough relevant data in time and space are available and (2) a similar approach combining absolute and relative judgments is used. We like to emphasize that warning services, which intend to assign sub-levels to a danger level, should make an effort to explore their quality, particularly if their communication in forecast products is envisioned. Such quality assessments, however, should not only be made for sub-levels but for any information conveyed in forecast products.
The sub-levels clearly increase the predictive value of the forecast, opening the discussion on how this information could be provided to forecast users.
Maps of Switzerland showing
The density of human-triggered avalanches (or the number relative to the surface area) compared to
The data collected as part of operational avalanche forecasting are available at the data repository
The authors contributed as follows: FT (study design, data curation and extraction, analysis, models, manuscript writing), SM (models, manuscript reviewing), CP (models, manuscript reviewing), GS (data curation and extraction, manuscript reviewing), and KW (study design, manuscript reviewing).
Two of the authors (Frank Techel, Kurt Winkler) are avalanche forecasters, directly involved in the production of the forecast. Günter Schmudlach is developer of the
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We thank the reviewers Rune Engeset and Karl Birkeland and the editor Pascal Haegeli for their valuable, constructive feedback, which helped to improve the manuscript.
This paper was edited by Pascal Haegeli and reviewed by Rune Engeset and Karl W. Birkeland.