Examining the operational use of avalanche problems with decision trees and model-generated weather and snowpack variables

Avalanche problems are used in avalanche forecasting to describe snowpack, weather, and terrain factors that require distinct risk management techniques. Although they have become an effective tool for assessing and communicating avalanche hazard, their definitions leave room for interpretation and inconsistencies. This study uses conditional inference trees to explore the application of avalanche problems over eight winters in Glacier National Park, Canada. The influence of weather and snowpack variables on each avalanche problem type were explored by analyzing a continuous set of weather and snowpack 5 variables produced with a numerical weather prediction model and a physical snow cover model. The decision trees suggest forecasters’ assessments are not only based on a physical analysis of weather and snowpack conditions, but also contextual information about the time of season, location, and interactions with other avalanche problems. The decision trees show clearer patterns when new avalanche problems were added to hazard assessments compared to when problems were removed. Despite discrepancies between modelled variables and field observations, the model-generated variables produced intuitive 10 explanations for conditions influencing most avalanche problem types. For example, 72 h snowfall was the most significant variable for storm slab avalanche problems, skier penetration depth was the most significant variable for dry loose avalanche problems, and slab density was the most significant variable for persistent slab avalanche problems. The explanations for wind slab and cornice avalanche problems were less intuitive, suggesting potential inconsistencies in their application as well as shortcomings of the model-generated data. The decision trees illustrate how forecasters apply avalanche problems and can 15 inform discussions about improved operational practices and the development of data-driven decision aids.

. Location of Glacier National Park in western Canada and locations of grid points for each elevation band where weather and snowpack data were generated with numerical weather prediction and physical snow cover models. Grid points used before the 2017 model update are shown with circles and grid points used after the update are shown with squares.
based on the type of information needed for their assessment (Table 1). Over the study period the three most common avalanche problem types were persistent slab (present on 61% of the days), wind slab (50%), and storm slab avalanche problems (46%).
Wet slab and glide slab avalanche problems were omitted from the analysis because they were never present during the study period.
To provide a meaningful analysis of persistent problem types, information about specific weak layers was also extracted 5 from the hazard assessments. A list of relevant weak layers was produced for each season by reading the problem description and snowpack summary sections of the public avalanche forecast. This method identified 54 distinct weak layers that were associated with persistent problem types over the study period. Each weak layer was identified by a burial date and assigned a status that changed over time based on its role in persistent problem types. Once buried by new snow, layers were initially assigned a status of surface in which case they may have been associated with a surface avalanche problem type (e.g., storm 10 slab avalanche problem). The status changed to active when layers became attributed to a persistent avalanche problem type and then changed to dormant after they were no longer attributed to a persistent problem (either the problem was removed from the forecast or the problem became attributed to a different weak layer). Fifteen of the weak layers were attributed to a persistent problem at least one more time after a period of dormancy. However, since the reawakening of persistent weak layers is likely driven by a different set of factors than the initial activation, we only considered the initial period of a layer contributing to a 15 persistent slab problem in our analysis. Another 41 potential weak layers were described in the hazard assessments but never  became attributed to persistent problems. These layers were also documented to compare with the layers that were attributed to persistent problems.

Weather and snowpack data
Weather and snowpack variables were produced for each day of the study period using data from a numerical weather prediction model and a physical snow cover model. Variables were selected to be similar to the types of information forecasters routinely consider in their hazard assessments. Model-generated data have a decisive advantage for our study as their continuity and consistency facilitate statistical analyses that would not be possible with the irregular field observations commonly used by 5 avalanche forecasters.

Weather variables
Weather variables were compiled for each day of the study period using hourly data from a numerical weather prediction model.

Data was compiled from the High Resolution Deterministic Prediction System (HRDPS) produced by the Meteorological
Service of Canada (Milbrandt et al., 2016). The HRDPS has 2.5 km grid spacing with initial and boundary conditions produced 10 by a coarser regional scale deterministic model.
A single representative grid point was chosen for each elevation band in Glacier National Park (Fig. 1). Preliminary analyses revealed that the gridded weather data was relatively homogenous across the park, which is mainly due to the 2.5 km horizontal resolution. This means the selection of grid points should not have a strong influence on our analysis. The main source of variability amongst the 236 grid points in the park was their elevation. Accordingly, grid points were chosen primarily based 15 on their elevation with a secondary consideration being locations where the forecasted precipitation amounts were average relative to all grid points. It is important to note that the HRDPS underwent two major changes over the study period; an upgrade prior to the 2015 winter that improved the overall quality of the forecasts and a switch to a new grid prior to the 2018 winter that required a new set of grid points for subsequent winters (Faucher, 2016).
To account for the evolutionary nature of avalanche conditions and provide a meaningful summary of the weather conditions 20 contributing to avalanche problems, hourly weather data were aggregated into summary statistics covering the 24 and 72 h periods prior to the hazard assessment (Table 2). Previous studies have consistently found weather conditions over these time periods as strong predictors of avalanche hazard (e.g. Schirmer et al., 2009;Yokley et al., 2014). Hourly surface weather data was extracted from the most up-to-date HRDPS forecasts, which was reinitialized every six hours. Precipitation was accumulated over the previous 24 and 72 h periods and partitioned into solid, liquid, and total components (using an air 25 temperature threshold of +0.5°C). Minimum, maximum, and average air temperature (2 m above surface) and hourly wind speed (10 m above surface) were calculated over the same time periods. Accumulated incoming shortwave and longwave radiation were calculated over the 24 and 72 h periods, as well as the maximum intensity of incoming shortwave radiation.

Snowpack variables
Snowpack conditions were simulated with the SNOWPACK snow cover model (Lehning et al., 1999). The snowpack structure 30 at each elevation band in Glacier National Park was simulated using the hourly weather data extracted from the representative HRDPS grid points (Fig. 1). Default SNOWPACK settings were used with wind transport disabled to produce flat field snow profiles representative of conditions in sheltered terrain. Daily snow profiles were produced at 08:00 to align with the hazard assessments.
To mimic existing forecaster practices, three interfaces were identified in the upper snowpack to produce variables relevant to surface avalanche problem types: a 24 h interface, a 72 h interface, and a storm interface (Fig. 2). The 24 and 72 h interfaces were identified by searching for all snow deposited in the past 24 and 72 h. The storm interface was identified by selecting all 5 snow deposited since the last full day without snow, similar to how accumulated snowfall is measured in the field on storm boards that are cleared in between stormy periods (Canadian Avalanche Association, 2016a).
Interfaces corresponding to the weak layers described in the hazard assessments were also identified to produce variables relevant to persistent avalanche problem types (Fig. 3). Since discrepancies arise between the burial date of layers used by forecasters and the deposition dates used by SNOWPACK, a method to search for the most likely weak layer was developed. First, a range of possible deposition dates were identified for each weak layer by searching for clear weather days prior to the burial date. Once buried, weak layers were identified by searching for the weakest layer that fell within the range of potential deposition dates using SNOWPACK's structural stability index to identify the weakest layer (Schweizer et al., 2006).
Snowpack variables were extracted from the simulated profiles to summarize slab and weak layer properties for each of the identified interfaces (Table 3). Variables characterizing slabs included slab thickness, average and maximum density, 5 average and maximum temperature, average and maximum hardness, and total liquid water content. Slab averaged values were calculated using a product sum with layer thickness. Variables characterizing weak layers included grain type, grain size, density, temperature, liquid water content, hardness, age (days since deposition), and whether there was a melt-freeze crust less than 5 cm below the weak layer. Four generic snowpack variables that did not depend on any interfaces were also calculated: height of snowpack, skier penetration depth, snow surface temperature, and the size of any surface hoar on the surface.

Verification of modelled data
Modelled snow heights and daily snowfall amounts were compared to assess the overall agreement between modelled variables and field observations. While an in-depth model validation was outside the scope of this study, sensitivity studies have shown physical snowpack models are most sensitive to precipitation inputs (Raleigh et al., 2015;Richter et al., 2020). Thus, the accuracy of modelled snow heights provides insights into the accuracy and reliability of the modelled data used in this 15 study. Parks Canada manually observes snow height and daily snowfall amounts every day at their below treeline study plot compared to the corresponding modelled snow heights at the below treeline and treeline grid points using Spearman correlation coefficients, percent differences, hit rates, and false alarm rates.

Decision tree analysis
Decision trees are a class of machine learning methods that provide simple and interpretable visualizations of complex non-5 linear relationships. Unlike other machine learning methods that focus on predictive performance (e.g., neutral networks, random forests), decision trees present relationships in ways that more closely resemble the human decision-making processes and are thus a helpful tool to understand the avalanche problem identification and assessment process. While the classification and regression tree (CART) method by Breiman et al. (1984) has been widely used, the method has issues with overfitting data and producing biased splits when the dependent variable is unevenly distributed. An alternate method that overcomes 10 these weaknesses is conditional inference trees (Hothorn et al., 2006), a method that uses statistical hypothesis testing to recursively split data sets until no more statistically significant splits exist in the data. When visualized in a tree diagram, the most significant splits in independent variables are presented higher on the tree, and the terminal nodes at the bottom of the tree show the resulting distribution of the dependent variable. In the present study, we employed conditional inference trees to explore relationships between weather variables, snowpack variables, and avalanche problem types using the partykit implementation in R (Hothorn and Zeileis, 2015). Since the focus of the analysis was to explore relationships rather than construct predictive models, the trees were made to be relatively simple by only including variable splits that had p-values smaller than 0.0001, except in analyses with smaller dataset where we kept the threshold p-value at 0.05. To compare various decision tree models, contingency tables were produced by comparing the 5 original observations with the classifications produced by each decision tree model (using the most common value in each terminal node). The accuracy, hit rate, and false alarm rate were calculated from the contingency table to describe the ability of each decision tree model to represent the original data (see Appendix A for definitions of these statistics).

Surface problem types
Decision trees were computed to determine the presence or absence of each surface avalanche problem type. The presence 10 or absence of a surface problem at each elevation band for each day of the study period was fit to the relevant weather and snowpack variables for that elevation band and day (960 days which are equivalent to 2880 elevation band days). Weather variables included those listed in Table 2 for the past 24 and 72 h periods. Snowpack variables included the four generic snowpack variables in Table 3 and the slab and weak layer variables for the three types of surface interfaces (24 h, 72 h, and since the last day without snow).

15
To offer a better understanding of how avalanche problems were applied, two decision trees were computed for each surface problem type: one using only weather and snowpack variables and a second using an additional set of contextual variables (Table 4). The contextual variables provided information about the hazard assessment and included temporal information (e.g. specific seasons, number of days since 1 October), elevation band, and the presence or absence of other surface problems the same day and the previous day. This information further strengthened the representation of the evolutionary character of 20 avalanche hazard in our analysis.

Persistent problems
Instead of analyzing whether persistent slab problems existed on a specific day, decision trees were computed to explore how persistent problem types related to specific weak layers. This approach reflects the more complex process of assessing persistent problems, where forecasters track the evolution of multiple weak layers over time rather than simply considering the upper snowpack. For simplicity the analysis was limited to the treeline elevation band where persistent slab problems were 5 most prevalent (Table 1). The analysis included the same weather variables as the surface problem analysis, but the snowpack variables were restricted to the slab and weak layer variables pertinent to the specific weak layer being tracked, plus the four generic snowpack variables shown in Table 3.
Several decision trees models were fit to compare conditions when weak layers were associated with persistent problems (active status) with conditions before (surface status) and after (dormant status). First, a large decision tree with a three level 10 dependent variable was used to examine the statuses of all 54 weak layers in a dataset that included 714 cases (i.e., weak layer days) of surface conditions, 866 cases of active conditions, and 280 cases of dormant conditions. The dormant cases were restricted to the first seven days after layers became dormant to focus on times when the problem was most relevant and avoid skewing the dataset with large numbers of cases with deeply buried dormant weak layers. Additional decision trees with binary dependent variables were fit to focus on the onset and the end of persistent problem types in more detail. Onset was analyzed 15 by narrowing the focus to the final three days of layers having a surface status and the first three days of having an active status.
A similar decision tree was fit to focus on the transition from active to dormant, however the number of days before and after the transition was expanded to 14 since no significant variables were found for shorter time periods.

Variable selection
The weather and snowpack variables included in our dataset exhibit natural correlations, which can negatively affect statistical 20 analyses and make the interpretation of decision trees more challenging. To avoid the problem, highly correlated weather and snowpack variables were removed from the analysis. Variables with pair-wise Spearman rank correlations above 0.9 were identified and the variable with the largest mean absolute correlation in the dataset was removed. A total of 23 highly correlated variables were removed from the surface problem dataset and 15 highly correlated variables were removed from the persistent problem dataset (see Appendix B for correlation plots). 25 3 Results

Snow height verification
Modelled snow heights and daily snowfall amounts were strongly correlated with field observations over the study period in Glacier National Park (Table 5). Whereas snow heights were consistently underestimated at the treeline grid points (Fig. 4), they were either overestimated, underestimated, or closely agreed at the below treeline grid points depending on the season. 30 Over the entire study period the modelled snow heights were an average of 36% and 5% lower than the observed snow heights at treeline and below treeline, respectively. The overall accuracy of predicting days with more than a trace of new snow was 86% (i.e. HN24 > 0.1 cm). Days with more than a trace amount of snowfall were predicted with a hit rate of 91% and a false alarm rate of 10%, while days with no snowfall or a trace amount were predicted with a hit rate of 77% and a false alarm rate of 21%. Based on the quantity and timing of modelled precipitation, the model-generated weather and snowpack variables are expected to capture the majority of weak layer and slab formation events over the study period, but also likely contain 5 inaccuracies about the detailed properties of the slabs and weak layers.

Storm slab avalanche problems
When only considering weather and snowpack variables, the most significant variable to explain the presence of storm slab avalanche problems was the height of new snow over the past 72 h (Fig. 5). Storm slab problems were present 62% of the time when the height was more than 7 cm and 77% of the time when the height was more than 16 cm. The next most significant variable was the maximum hourly wind speed over the past 24 h, with storm slabs becoming more common when small amounts of new snow coincided with higher wind speeds. Other variables appearing lower in the decision tree included the grain type beneath the last 72 h of new snow (storm slab problems were more common when this was rounded grains), the average slab density of snow from the past 72 h (storm slab problems were more common when the density was greater than 5 113 kg m -3 ), and the maximum intensity of incoming shortwave radiation in the past 72 h (storm slab problems more common when shortwave radiation was weaker).
When adding contextual variables, the first and most significant variable was whether a storm slab problem was present the previous day (Fig. 6), with the problem persisting 81% of the time. The subsequent splits describe when storm slab problems were newly added to a hazard assessment (i.e. left branch of the tree) or when storm slab problems were removed (i.e. right 10 branch of the tree). The most significant rule for adding a new storm slab problem was height of new snow over the past 24 h, with a problem being added 35% of the time when there was more than 5 cm of new snow. The most significant rule for removing a storm slab problem was the amount of incoming longwave radiation over the past 24 h, with the problem more frequently removed after a period of clear skies (i.e. lower values of incoming longwave radiation). When adding a new storm slab problem, the second most important variable after a threshold amount of new snow was whether there was already a wind 15 slab problem in the hazard assessment, with very few cases of both storm and wind slab problems present on the same day.
Wind slab problem variables on the right branch of the decision tree show a similar theme of wind slab problems replacing storm slab problems after a period of clear skies.
When using the decision trees for classification, the addition of contextual variables increased the percentage of correct classifications from 78% to 88% (Table 6), increased the hit rate from 76% to 84%, and decreased the false alarm rate from 20 35% to 13%.

Wind slab avalanche problems
When only considering weather and snowpack variables, the most significant variable for wind slab avalanche problems was the maximum air temperature over the past 24 h (Fig. C1). Wind slab problems were more common when the maximum air temperature was less than -9°C. Snowpack variables in the decision tree included the grain type beneath the last 72 h of 25 new snow (wind slab problems were more common when the grain type was precipitation particles, rounded grains, or faceted 13 https://doi.org/10.5194/nhess-2020-274 Preprint. Discussion started: 7 September 2020 c Author(s) 2020. CC BY 4.0 License. 14 https://doi.org/10.5194/nhess-2020-274 Preprint. Discussion started: 7 September 2020 c Author(s) 2020. CC BY 4.0 License.

Problems Weather Snowpack Elevation
Variable type 15 https://doi.org/10.5194/nhess-2020-274 Preprint. Discussion started: 7 September 2020 c Author(s) 2020. CC BY 4.0 License. crystals), the size of any surface hoar on the surface (wind slab problems were less common when there was large surface hoar on the surface), the height of new snow since the last snow free day, the grain type that formed before the last snow free day, and the total height of the snowpack.
When adding contextual variables, the first and most significant split was whether a wind slab problem was present the previous day (Fig. C1), with the problem persisting 82% of the time. The variables that contributed to adding a new wind slab 5 problem included elevation band (wind slab problems were uncommon at the below treeline band), the presence or absence of a storm slab problem the same or previous day, the particular season, and the time of year (wind slab problems were more common before 7 January). The variables that contributed to removing a wind slab problem were the presence or absence of a storm slab problem the same or previous day and the size of any surface hoar on the surface (wind slabs were removed more frequently when surface hoar had grown larger than 1.9 mm).

10
When using the decision trees for classification, the addition of contextual variables increased the percentage of correct classifications from 74% to 92% (Table 6), increased the hit rate from 59% to 85%, and decreased the false alarm rate from 50% to 12%.

Dry loose avalanche problems
The most significant variable impacting dry loose avalanche problems in the decision trees with only weather and snowpack 15 variables was the skier penetration depth (Fig. C2). Dry loose problems were present 19% of the time when the skier penetration depth was greater than 18 cm. The only other significant variable was the average density of snow deposited since the last snow free day, with dry loose problems more common when the average slab density was less than 139 kg m -3 . When contextual variables were included in the analysis, dry loose problems persisted 64% of time when they were present the previous day.

Cornice avalanche problems
The analysis of only weather and snowpack variables found significant variables for cornice avalanche problems were the total height of the snowpack (Fig. C4), maximum of shortwave radiation over the past 24 h, maximum hourly wind speed over the past 24 h, the height of new snow since the last snow free day, and the snow surface temperature. When including contextual 25 variables, persistence played a dominant role in the presence of cornice problems, with cornice problems persisting 74% of the time and no significant variables for removing a cornice problem. The most significant variable influencing when to add a new cornice problem was elevation band (cornices were primarily present in alpine elevation band) and then a secondary split based on the maximum shortwave radiation over the past 24 h (cornice problems were added with large amounts of incoming shortwave radiation). 30 The decision tree with weather and snowpack variables never resulted in situations when cornice problems would be classified, but the decision tree with contextual variables would predict a cornice problem if it was present the previous day.
Since cornice problems were very rare, these decision trees had high accuracies (97% and 98%).

Persistent avalanche problem types
The 54 weak layers attributed to persistent avalanche problem types were found to have significantly larger grain sizes than another 41 potential weak layers that were also described in the hazard assessments but did not result in persistent problems.
A decision tree model comparing these two sets of layers found layers with grain sizes larger than 1.1 mm were more likely to become attributed to persistent problems (not shown).

5
The most significant variable influencing a known weak layer's association with persistent avalanche problems was the density of the slab above the weak layer. When analyzing all the cases for the 54 weak layers from the time they were buried until one week after they were no longer associated with a persistent problem found these layers were associated with persistent problems 66% of the time when the maximum slab density was greater than 214 kg m -3 , compared to only 15% of the time when the slab density was below that threshold (Fig. 7). Weak layer age was the second most influential variable, appearing 10 several places in the decision tree. The situations when these weak layers were most often associated with persistent problems were when the slab density was relatively high and when the weak layer was between 6 and 26 d old. The most common situations when these weak layers had not yet developed into persistent problems (i.e. surface status) were when the slab density was low and when weak layers were young (with threshold ages of 6, 8, and 18 d appearing in the tree). Two additional situations delayed the onset of persistent problems; periods of low incoming solar radiation and when there was a low-density 15 weak layer with more than 21 cm of snow over the past 72 h. These situations both describe conditions that would be expected during storms, suggesting the weak layer would more likely be associated with a surface problem type. The most common situations when these weak layers were no longer associated with persistent problems (i.e. dormant status) were when they were older (with threshold ages of 18 and 26 d appearing in the tree) and after periods of sustained wind (minimum hourly wind speed greater than 0.4 m s -1 over the past 72 h).

20
The dominant conditions related to forecasters decisions to add or remove persistent problems become clearer when narrowing the focus to the actual start and end of persistent problems. Due to the smaller size of the data sample, these analyses revealed fewer significant variables at the p = 0.0001 level, so the maximum p-value for variables shown in the decision trees was increased to 0.05. Maximum slab density was again the most significant difference between the three days before layers became associated with persistent problems and their first three days of being associated with persistent problems (Fig. 8). Exceeding 25 a maximum slab density of 193 kg m -3 increased the frequency of the layer being a problem from 32% to 57%. Snow surface temperatures colder than -18°C further increased the frequency of weak layers becoming associated with persistent problems.
Comparing conditions before and after weak layers became dormant revealed no significant difference for time periods of 3 or 7 d before and after the transition. Expanding the window to compare conditions two weeks before and after the transition revealed several significant differences (Fig. 9). First, weak layers were more commonly associated with a persistent problem 30 when the maximum slab density was less than 292 kg m -3 . Additional factors that contributed to weak layers persisting were an age less than 16 d, and air temperatures above -8°C. The slab density above weak layers was the most significant variable for both the onset and the end of persistent problems, with weak layers most often associated with persistent problems when the maximum slab density was between 193 and 292 kg m -3 .

Weather Snowpack Time
Variable type show how forecasters integrated weather and snowpack information into their hazard assessment process. The decision trees with contextual variables always started by asking whether the problem was present the previous day, illustrating an iterative Bayesian process where each assessment starts with the prior assessment and then considers new information (LaChapelle, 1980).
The addition of contextual variables resulted in decision tree models that better represented the original data, largely due 10 to the prominent role of persistence. The improvement was indicated by a higher accuracy, higher hit rates, and lower false alarm rates when using the trees for classification (Table 6). For example, when adding contextual variables, the percentage of correct classifications by the storm slab decision tree increased from 78% to 88%, the hit rate increased from 76% to 84%, and false alarm rate dropped from 35% to 13%. The decision tree for wind slab problems had similar improvements when adding contextual variables. The improvements were less clear for dry loose, wet loose, and cornice avalanche problems because these 15 problems were less prevalent in the dataset. Even with statistically significant variables in their decision trees, the terminal nodes still suggest these problems were absent most of the time, unless it was known that the problem was present the previous day. This also reinforces the important role of persistence in avalanche forecasting.
Another dominant theme was adding new avalanche problems often had more significant and intuitive explanations than removing old avalanche problems, suggesting problems were added more consistently by forecasters than they were removed. 20 This is evident by the greater number of significant variables in the left branch of trees with contextual variables (i.e. situations when a problem would be added) than in the left branch of these trees (i.e. situations when a problem would be removed).
For example, there were no significant weather or snowpack variables to explain the removal of dry loose, wet loose, or cornice avalanche problems. Similarly, when analyzing the onset of persistent avalanche problems there were several significant differences in conditions three days before and after the onset of the problem (Fig. 8), but there were no significant differences 25 three or even seven before and after the end of the problem. These patterns suggest forecasters likely have greater confidence, precision, and consistency when adding new problems. Removing problems could be more inconsistent due to uncertainty about when the likelihood or consequence of avalanches has reduced to the point where problems no longer need to be listed (Lazar et al., 2012). However, it is important to remember that our analysis did not include the full range of observations that avalanche forecasters consult when removing an avalanche problem (e.g. avalanche observation data). 30 Significant interactions between storm slab and wind slab avalanche problems were apparent in the decision trees ( Fig.   5 and Fig. C1b), as at least one of these two problems were listed in the hazard assessment on 83% of the days over the study period. Some operational policies suggest using only one of these two problems at a time to effectively communicate the most relevant risk management approach (Klassen, 2014). The decision trees with contextual variables show this practice slab and cornice problems could be achieved with model configurations that better capture snow transport (e.g. Vionnet et al.,

2018).
Models also offer opportunities to capture conditions that are difficult to measure in the field. For example, slab density above a weak layer was found to have a significant influence on persistent avalanche problem types. Slab densification is difficult to continually quantify from field observations but is likely a process that forecasters hypothesize about when assessing 5 persistent problem types. Models could help forecasters explicitly test hypotheses about slab density trends and similar physical properties. Although this analysis used weak layers identified by forecasters, weak layer detection could also be automated by models (Monti et al., 2014). For example, the 54 weak layers in this analysis were found to have significantly larger grain sizes than other potential weak layers over the same period, suggesting layer properties could be analyzed to detect potential weak layers. However, our approach was more appropriate for understanding how forecasters applied persistent problem types.

Applications and limitations for decision support tools
Data-driven decision support tools like the decision trees presented in this study have the potential to improve the accuracy and consistency of avalanche hazard assessments (Clark, 2019; Lazar et al., 2016;Statham et al., 2018b;Techel et al., 2018). Providing guidance on avalanche problems could provide more tangible information to forecasters about necessary risk mitigation measures than previous decision aids that focused on danger level or avalanche occurrence. Expanding the decision 15 tools to provide insight into additional avalanche problem attributes such as the location, size, and likelihood of avalanches could offer additional value.
While this study focused on model-generated weather and snowpack variables due to their consistency and temporal coverage, decision support tools could also include a broader range of inputs that represent the data available to forecasters more comprehensively. First, including field observations of weather and snowpack conditions would be a meaningful complement 20 to overcome issues with model errors. Second, including avalanche activity data would likely improve aspects of the decision trees such as providing better explanations of when avalanche problems are removed. A challenge would be formatting field observations in a way that allows them to be integrated into a statistical model in a consistent way. Future decision aids should consider leveraging the respective strengths of field observations and model-generated variables.
Using operational datasets to build predictive models will continue to face issues with messy and inconsistent patterns that 25 could perpetuate errors in human assessments (Guikema, 2020). The decision trees presented in this study only showed the most significant variables by restricting the p-value to 0.0001 to keep the diagrams tidy and easy to interpret. However, adjusting the maximum p-value to 0.05 revealed decision trees with many additional significant variables that become more challenging to interpret. While such models may offer considerable predictive power, the underlying rules should be grounded in physical conditions and operational priorities to provide optimal and unbiased decision support.  Terminal nodes show the proportion of cases with cornice problems present (green) and absent (yellow).