Fig 2 - uploaded by Frank Techel
Content may be subject to copyright.
Distribution of the forecast danger level D RF (a) and the sub-levels D RF. sub (b) during the four winters 2016/2017 to 2019/2020 for dry-snow avalanches, as issued in the morning forecast.
Source publication
In public avalanche forecasts, avalanche danger is summarized using a five-level ordinal danger scale. However, in Switzerland-but also in other countries-on about 75% of the forecasting days, only two of the five danger levels are actually used, indicating a lack of refinement in the forecast danger level. A refined classification requires the for...
Contexts in source publication
Context 1
... the goal to explore whether D RF. sub was better than a random . This random assignment of sub-levels, however, was not fully random as we sampled according to the distributions of the forecast sub-levels for each of the danger levels (as shown in Fig. 2b). This approach already introduces some skill in the random assignment of sub-levels. Proceeding as described before, we obtained the difference in sub-level ranks and thus the misclassification cost for D RF.sub.random according to Tables 1 and ...
Context 2
... by exploring overall distributions, temporal changes and spatial gradients in danger ratings between immediately neighboring warning regions. And secondly, we focus on the quality of the forecast sub-levels, that is, the agreement between forecast and local estimate and whether forecast sub-levels were better than random (Sections 4.2 and 4.3). Fig. 2a shows the distribution of forecast danger levels D RF for drysnow conditions in the Swiss Alps during the 4-year period. 2-Low and 3-Considerable were forecast about 80% of the time. Avalanche danger was not explicitly communicated for each of the more than 100 warning regions in the Alpine forecast area, but warning regions were ...
Context 3
... proportion of the forecast sub-levels D RF. sub decreased monotonically from 3-minus to 5-minus (Fig. 2b). At 2-Moderate, no such pattern showed. Of note was the comparably low proportion of 2-plus (10%), used less often than both the immediately lower 2-neutral (18%) and higher (3-minus) sub-levels (17%). Expecting an approximately similar usage of these D RF. sub , these proportions were significantly different (proportion test (R Core ...
Context 4
... to 2-Moderate, often during periods when D decreased rather gradually. On these days, D RF.sub decreased more often by two levels than on the days immediately before (Section 4.1.2). ii. 2-plus was used significantly less often in the forecasts than would be expected, when compared to the frequency of the respective lower and higher sub-levels (Fig. ...
Similar publications
This article presents research on the detection of pathologies affecting speech through automatic analysis. Voice processing has indeed been used for evaluating several diseases such as Parkinson, Alzheimer or depression. If some studies present results that seem sufficient for clinical applications, this is not the case for the detection of sleepi...
Community detection is a key task to further understand the function and the structure of complex networks. Therefore, a strategy used to assess this task must be able to avoid biased and incorrect results that might
invalidate further analyses or applications that rely on such communities. Two widely used strategies to assess this task are general...
A dialogue policy module is an essential part of task-completion dialogue systems. Recently, increasing interest has focused on reinforcement learning (RL)-based dialogue policy. Its favorable performance and wise action decisions rely on an accurate estimation of action values. The overestimation problem is a widely known issue of RL since its est...
Discourse signals are often implicit, leaving it up to the interpreter to draw the required inferences. At the same time, discourse is embedded in a social context, meaning that interpreters apply their own assumptions and beliefs when resolving these inferences, leading to multiple, valid interpretations. However, current discourse data and framew...
This article presents research on the detection of pathologies affecting speech through automatic analysis. Voice processing has indeed been used for evaluating several diseases such as Parkinson, Alzheimer, or depression. If some studies present results that seem sufficient for clinical applications, this is not the case for the detection of sleep...
Citations
... In contrast to Brabec and Meister and Schirmer et al. (2009), Pérez-Guillén et al. (2022a) trained a model not only using data from a single station describing meteorological variables but by also including snow stratigraphy information simulated with SNOWPACK on more than 120 AWS located at the elevation of avalanche-prone areas in all regions of Switzerland. Their standard version of the classifier exhibited an accuracy of 74%, which is remarkably good considering that the accuracy of human-made avalanche forecasts in Switzerland is estimated to be in the range between 75 75% and 81% (Techel and Schweizer, 2017;Techel et al., 2020). Since the winter season 2021/2022, several machine learning models are operationally tested by avalanche forecasters in Switzerland, including the model by (Pérez-Guillén et al., 2022a), with generally positive feedback from the forecasters regarding model performance and usefulness (van Herwijnen et al., 2023). ...
Despite the increasing use of physical snow-cover simulations in regional avalanche forecasting, avalanche forecasting is still an expert-based decision-making process. However, recently, it has become possible to obtain fully automated avalanche danger level predictions with satisfying accuracy by combining physically-based snow-cover models with machine learning approaches. These predictions are made at the location of automated weather stations close to avalanche starting zones. To bridge the gap between these local predictions and fully data-and model-driven regional avalanche danger maps, we 5 developed and evaluated a three-stage model pipeline (RAvaFcast v1.0.0), involving the steps classification, interpolation, and aggregation. More specifically, we evaluated the impact of various terrain features on the performance of a Gaussian process-based model for interpolation of local predictions to unobserved locations on a dense grid. Aggregating these predictions using an elevation-based strategy, we estimated the regional danger level and the corresponding elevation range for predefined warning regions, resulting in a forecast similar to the human-made avalanche forecast in Switzerland. The best-performing model 10 matched the human-made forecasts with a mean day accuracy of approximately 66% for the entire forecast domain, and 70% specifically for the Alps. However, the performance depended strongly on the classifier's accuracy (i.e., a mean day accuracy of 68%) and the density of local predictions available for the interpolation task. Despite these limitations, we believe that the proposed three-stage model pipeline has the potential to improve the interpretability of machine-made danger level predictions and has, thus, the potential to assist avalanche forecasters during forecast preparation, for instance, by being integrated in the 15 forecast process in the form of an independent virtual forecaster.
... Therefore, we used two different approaches to explore "accuracy". In the first study, Techel et al. (2020) compared the forecast danger level (DLfx) and the unpublished SL with the danger level assessments made after a day in the field by observers (DLobs). The results showed that, in case of differences between DLfx and DLobs, most of the time the SL closest to DLobs was forecast, indicating that the "error" in the forecast was often less than a "full" DL. ...
In public avalanche forecasts, avalanche danger is commonly summarized using one of five ordinal danger levels. This strong simplification of a complex phenomenon inevitably leads to a loss of information. To compensate for this information loss and to ease interpretation of the forecast, the Swiss avalanche warning service introduced sublevels, that indicate whether the forecast danger is in the lower (-), middle (=) or higher (+) range of the danger level. After a 6-year test period, during which it showed that forecasters can reliably assign sublevels to a danger level, the sublevels were published for the first time in the public avalanche forecast in the winter season of 2022/2023. To evaluate the general understanding and the usefulness of the sublevels, and to understand how they influence people's backcountry behavior, we performed a user survey towards the end of the season. The survey with 3403 participants showed that participants perceived the sublevels as generally useful (85%) and that they understood them in the intended manner (94%). Sublevels also had a positive influence on the planning of backcountry tours. Faced with a forecast sublevel (+), almost half of the participants claimed they would change their plans to a less exposed tour than they had before when the same danger level was given but without sublevel information. This self-stated influence of the sublevels on the choice of a backcountry tour was especially important for participants with little or no avalanche training, showing that sublevels seem most useful for less experienced forecast users, who may find it difficult to extract similar information from the forecast text description.
... Furthermore, a strong tendency towards over-forecasting (by one level) has been noted, with fore-casts rarely being lower compared to nowcast assessments of avalanche danger (e.g. Techel et al., 2020b). ...
... In contrast, decreasing avalanche danger is often linked to comparably minor and/or slow changes in snowpack stability (e.g. Techel et al., 2020b). While these changes are gradual in nature, these can only be expressed in a step-like fashion using the five-level danger scale. ...
... Previous studies estimated the accuracy of the Swiss avalanche forecasts to be in the range between 75 % and 81 % (this study -see Sect. 5.2; Techel and Schweizer, 2017;Techel et al., 2020b). Our classifiers reached these values: the overall prediction accuracies of RF 1 and RF 2 were 74 % and 72 % (compared to D forecast ) and 76 % and 78 % (compared to D tidy ), respectively (Table 1). ...
Even today, the assessment of avalanche danger is by and large a subjective yet data-based decision-making process. Human experts analyse heterogeneous data volumes , diverse in scale, and conclude on the avalanche scenario based on their experience. Nowadays, modern machine learning methods and the rise in computing power in combination with physical snow cover modelling open up new possibilities for developing decision support tools for operational avalanche forecasting. Therefore, we developed a fully data-driven approach to assess the regional avalanche danger level, the key component in public avalanche forecasts, for dry-snow conditions in the Swiss Alps. Using a large data set of more than 20 years of meteorological data measured by a network of automated weather stations, which are located at the elevation of potential avalanche starting zones, and snow cover simulations driven with these input weather data, we trained two random forest (RF) classifiers. The first classi-fier (RF 1) was trained relying on the forecast danger levels published in the official Swiss avalanche bulletin. To reduce the uncertainty resulting from using the forecast danger level as target variable, we trained a second classifier (RF 2) that relies on a quality-controlled subset of danger level labels. We optimized the RF classifiers by selecting the best set of input features combining meteorological variables and features extracted from the simulated profiles. The accuracy of the models, i.e. the percentage of correct danger level predictions , ranged between 74 % and 76 % for RF 1 and between 72 % and 78 % for RF 2. We assessed the accuracy of forecasts with nowcast assessments of avalanche danger by well-trained observers. The performance of both models was similar to the agreement rate between forecast and nowcast assessments of the current experience-based Swiss avalanche forecasts (which is estimated to be 76 %). The models performed consistently well throughout the Swiss Alps, thus in different climatic regions, albeit with some regional differences. Our results suggest that the models may well have potential to become a valuable supplementary decision support tool for avalanche forecasters when assessing avalanche hazard .
... It is therefore important that avalanche forecasters assess avalanche danger as detailed as possible when preparing a public forecast, given the available data and resources. This level of detail may be greater than what is communicated in the forecast product (e.g., Walcher et al., 2018;Techel et al., 2020b). 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 (MacGregor, 2001;Statham et al., 2018a). ...
... Compared to absolute judgments, this approach requires more effort and is time-consuming but allows a finer discrimination within previously assigned categories (Kahneman et al., 2021). Such an approach has been used during the past 5 years in Switzerland, where forecasters assigned a danger level and a sub-level qualifier refining where within this danger level the avalanche conditions are expected (Techel et al., 2020b). This leads to our overarching research question: using such an approach to assign a sub-level qualifier to a danger level, can human avalanche forecasters forecast avalanche hazard at finer granularity than the five danger levels? ...
... To do so, an approach combining absolute and comparative judgments, as described in the previous section, is used. Forecasters first assign a danger level according to the definitions in the EADS and then make a comparative refinement using one of three qualifier terms (Techel et al., 2020b): plus or +: the danger is assessed as high within the level; e.g., a 3 + is high within 3 (considerable); ...
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.
... Compared to absolute judgments, this approach requires more effort and is time-consuming, but allows a finer discrimination within previously assigned categories (Kahneman et al., 2021). Such an approach has been used during the past five years in Switzerland, where forecasters assigned a danger level and a sub-level qualifier refining where within this danger level the avalanche conditions are expected (Techel et al., 2020b). This leads to our over-arching research question: using such an approach to assign a sub-level qualifier to a danger level, can human avalanche forecasters forecast avalanche hazard at finer granularity than the five danger levels? ...
... region C1 in a). Source: example is taken from Techel et al. (2020b). ...
... To do so, an approach combining absolute and comparative judgements, as described in the previous section, is used. Forecasters first assign a danger level according to the definitions in the EADS, and then make a comparative refinement using one of three qualifier terms (Techel et al., 2020b): ...
Forecasting avalanche danger at a regional scale is a largely data-driven, yet also experience-based decision-making process. In the case of public avalanche forecasts, this 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. 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 five years. Using a sequential analysis, we first establish which data is suitable and in which part of the danger scale by comparing their distributions at consecutive danger levels. In a second step, integrating these findings, we compare the frequency of locations with poor snow 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 snow 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 is sufficiently available in time and space, and (2) if a similar approach combining absolute and relative judgment is used. The sub-levels increase the predictive value of the forecast, opening the discussion on how this information could be provided to forecast users.
... Furthermore, a strong tendency towards over-forecasting (one level) has been noted, with forecasts rarely being lower compared to nowcast assess-150 ments of avalanche danger (e.g. Techel et al., 2020b). ...
... In contrast, decreasing avalanche danger is often linked to comparably minor and/or slow changes in snowpack stability (e.g. Techel et al., 2020b). While these changes are gradual in nature, these can only be expressed in a step-like fashion using the five-level danger scale. ...
... Previous studies estimated the accuracy of the Swiss avalanche forecasts in the range between 75 % and 81 % (this study, see Sect. 5.2; Techel and 505 Techel et al., 2020b). Our classifiers reached these values: the overall prediction accuracies of RF #1 and RF #2 were 74 % and 72 % (compared to D forecast ) and, 76 % and 78 % (compared to D tidy ), respectively (Table 1). ...
Even today, the assessment of avalanche danger is by large a subjective, yet data-based decision-making process. Human experts analyze heterogeneous data volumes, diverse in scale, and conclude on the avalanche scenario based on their experience. Nowadays, modern machine learning methods and the rise in computing power in combination with physical snow cover modelling open up new possibilities for developing decision support tools for operational avalanche forecasting. Therefore, we developed a fully data-driven approach to predict the regional avalanche danger level, the key component in public avalanche forecasts, for dry-snow conditions in the Swiss Alps. Using a large data set of more than 20 years of meteorological data measured by a network of automated weather stations, which are located at the elevation of potential avalanche starting zones, and snow cover simulations driven with these input weather data, we trained two random forest (RF) classifiers. The first classifier (RF #1) was trained relying on the forecast danger levels published in the avalanche bulletin. Given the uncertainty related to a forecast danger level as a target variable, we trained a second classifier (RF #2), relying on a quality-controlled subset of danger level labels. We optimized the RF classifiers by selecting the best set of input features combining meteorological variables and features extracted from the simulated profiles. The accuracy of the danger level predictions ranged between 74 % and 76 % for RF #1, and between 72 % and 78 % for RF #2, with both models achieving better performance than previously developed methods. To assess the accuracy of the forecast, and thus the quality of our labels, we relied on nowcast assessments of avalanche danger by well-trained observers. The performance of both models was similar to the accuracy of the current experience-based Swiss avalanche forecasts (which is estimated to 76 %). The models performed consistently well throughout the Swiss Alps, thus in different climatic regions, albeit with some regional differences. A prototype model with the RF classifiers was already tested in a semi-operational setting by the Swiss avalanche warning service during the winter 2020-2021. The promising results suggest that the model may well have potential to become a valuable, supplementary decision support tool for avalanche forecasters when assessing avalanche hazard.
... The avalanche danger scale, with its five levels -of which only three are used during 98% of the time, and the discrete nature of both the danger scale and the core zone, inevitably lead to a loss of information as transitions cannot be communicated easily. For a manual application, however, the advantage of a simple number may outweigh a certain loss of information as whole danger levels can likely be understood by a user more easily than a probabilistic forecast (Murphy, 1993;Techel et al., 2020b). Therefore, if more detailed information regarding the avalanche conditions would be provided in the forecast, the comprehensibility and the usefulness of this information should be well studied. ...
Rule-based decision frameworks are widely recommended to estimate the avalanche risk while planning a ski tour. However, these frameworks were developed relying primarily on accident data and usually did not consider backcountry travel data. Hence, they are not risk-based. Here, we address this gap and calculate the risk taken during backcountry touring in avalanche terrain and correlate it to the expected avalanche conditions as described in a public avalanche forecast. For this, we rely on 784 reported avalanche accidents and more than 2.1 million movement points in potential avalanche terrain, based on GPS tracks recorded in Switzerland for 14 winter seasons. Combining this data with the respective avalanche forecast, we show that risk increases fourfold from danger level 1-Low to 2-Moderate, and from 2-Moderate to 3-Considerable. Furthermore, at 2-Moderate and 3-Considerable, in the critical elevations and aspects specified in the avalanche forecast, the risk is nearly six times higher compared to locations outside this, so-called, core zone. For danger level 1-Low, where the Swiss avalanche forecast does not provide any information about the critical elevations and aspects, we derived a risk-based core zone. Within this core zone too, the risk is about six times higher than outside. These findings suggest an adaption of the rule-based decision frameworks to reflect the observed risk better. The proposed framework considers the strong influence of the elevation and reduces the effect of the aspect, compared to former decision frameworks. We emphasize that this new decision framework cannot replace on-site risk assessment. However, it allows backcountry users to come closer to the goal of achieving a minimum of avalanche risk while allowing a maximum of freedom of movement.
