Enhancing the operational value of snowpack models with visualization design principles

Forecasting snow avalanches requires a reliable stream of field observations, which are often difficult and expensive to collect. Despite the increasing capability of simulating snowpack conditions with physical models, models have seen limited adoption by avalanche forecasters. Feedback from forecasters suggest model data is presented in ways that are difficult to interpret and irrelevant to operational needs. We apply a visualization design framework to enhance the value of snowpack models to avalanche forecasters. An established risk-based workflow for avalanche forecasting is used to define the ways 5 forecasters solve problems with snowpack data. We address common forecasting tasks such as identifying snowpack features related to avalanche problems, summarizing snowpack features within a forecast area, and locating problems in terrain. Examples of visualizations that support these tasks are presented and follow established perceptual and cognitive principles from the field of information visualization. Interactive designs play a critical role in understanding these complex datasets and are well suited for forecasting workflows. Preliminary feedback suggests these design principles produce visualizations 10 that are more relevant and interpretable for avalanche forecasters, but additional operational testing is needed to evaluate their effectiveness. By addressing issues with interpretability and relevance, this work sets the stage for implementing snowpack models into workstations where forecasters can test their operational value and learn their capabilities and deficiencies.

4. Algorithm level. This is the level where idioms are produced from raw data with a computer. Issues arise when algorithms are too slow. At the algorithm level, most snowpack model visualizations are time consuming for forecasters because they are poorly integrated into their workstations (accessibility). Munzner (2014) also describes that visualization problems can be attacked from two possible directions within the nested model: top-down approaches that first understand the domain and tasks and then design visual idioms accordingly, and bottom-5 up approaches that start with developing new algorithms and idioms. Most existing snowpack model visualizations were developed with bottom-up approaches that began with model development followed by the creation of visualizations of the model output. Bottom-up approaches allow novel visualizations that reveal nuances and anomalies in new types of data, but also have the potential to not solve the intended problem (Munzner, 2014). While it is worth considering bottom-up designs that take advantage of the unique capabilities of snowpack models, it is also important to carefully examine the domain and 10 tasks of avalanche forecasting to establish top-down design principles that support forecasting needs.

Domain of avalanche forecasting
Avalanche forecasting is a common task for all operations that manage short-term avalanche risk (e.g. ski areas, transportation corridors, backcountry warnings, resource extraction). The forecasting process consists of iterative data analysis and is dominated by human judgement and inductive logic (LaChapelle, 1980;McClung, 2002). Statham et al. (2018) surveyed existing operational 15 practices within North American avalanche forecasting operations to develop a standard framework for this process. The resulting conceptual model of avalanche hazard (CMAH) identifies the key components of avalanche hazard and provides standard workflow and terminology to guide the forecasting process. The CMAH is a risk-based framework that is consistent with other natural hazard disciplines and can be applied to any scale in space or time. A central part of the CMAH is the concept of avalanche problems that represent individual, identifiable operational concerns that can be described in terms of 20 their potential avalanche type, location, likelihood, and size (Statham et al., 2018). Under the CMAH, avalanche forecasting is viewed as sequentially answering four questions: 1. What type of avalanche problems exist? 2. Where are these problems located in the terrain?
3. How likely is it that an avalanche will occur? 25 4. How destructive will the avalanche be?
Over the past decade, the CMAH has been widely adopted by all industry sectors in North America (Statham et al., 2018), which clearly indicates that it is a useful model to describe the domain situation of avalanche forecasting.

Task and data abstractions for snowpack analysis
Given the importance of avalanche problems in avalanche forecasting practices, any operational visualization of data should 30 consider this abstraction to help forecasters identify and characterize avalanche problems. Assessing avalanche problems consists of integrating a complex array of data that includes observations of avalanches, snowpack, weather, and terrain (Statham et al., 2018). There is no structured or standardized way this data is used to answer the CMAH questions, as the analysis relies on subjective judgement and heuristics (LaChapelle, 1980), however there are common practices for interpreting field observations. Snowpack models produce data that is analogous to manual snow stratigraphy profiles, which is a key type of field data used by forecasters. Building off familiar visual representations is an effective way for people to understand new 5 types of information (Blackwell, 2001), and thus examining existing practices for visualizing and analyzing snow profiles provides insight into ways snowpack models could be visualized to support forecasting tasks.
Forecasters perform several analysis tasks with snow stratigraphy profiles to help them assess avalanche problems and develop a comprehensive mental model of hazard conditions. Individual snow profiles are either recorded in tables of unstructured text or illustrated as a hardness profiles (Canadian Avalanche Association, 2016a). Forecasters learn to identify relevant 10 snowpack features in these profiles, and then compare multiple snow profiles along with other observations to summarize the snowpack conditions within a forecast area. Forecasters summarize snowpack data by writing a concise overview of snowpack conditions in their forecast area. The goal of a written snowpack summary is to organize and reduce data, focusing on average conditions along with potential anomalies and outliers (Canadian Avalanche Association, 2016a). Some operations also illustrate their snowpack summary with generalized stratigraphy profiles for their forecast area (Fig. 1).

15
Tracking trends in snowpack conditions over time is another common forecasting task, which is most often done with tables of text. Temporal trends in the likelihood and size of avalanches are particularly relevant. For example, the InfoEx forecasting workflow allows forecasters to track weak layers in their forecast area with qualitative summaries of their status and depth each day of the season (Haegeli et al., 2014). Simple observed snowpack data is plotted as time series (e.g. daily snowfall at fixed observation sites), but complex data like snowpack structure is rarely visualized temporally.

20
To help forecasters answering the four key questions about avalanche problems posed by the CMAH, visualizations of snowpack model data should help forecasters identify, compare, and summarize snowpack features and highlight trends over time.

Information visualization principles
The field of information visualization studies how to leverage the human visual system to off-load cognitive work and visualize 25 information effectively. Information visualization principles should be considered when designing the visual appearance and interactive components of tools for snowpack model data (i.e. the visual encoding and interaction idiom level of the nested model). These principles consider effective ways of representing data visually and are explained in greater detail in textbooks by Ware (2012) and Munzner (2014). The following list summarizes information visualization principles that are relevant when visualizing snowpack model data:

30
-When representing information visually, designers encode data to visual features such as: spatial position, size, color, or shape among others. Color can be further divided into hue (the actual color), luminance (the brightness or darkness of a color), and saturation (the intensity of the color). Through years of perception studies, standard guidelines for mapping these visual features to data types have been established (Cleveland and McGill, 1984).
-Visual encodings should present data in ways that match the capabilities of our visual system. Hence, categorical and ordered data should be encoded with visual features that match human visual aptitudes. For example, when using colours, hues should be used for categorical attributes such as avalanche problem types and luminance (lightness or brightness) 5 should be used for ordered attributes such as avalanche likelihood.
-Designs should prioritize the importance of information and encode data to visual features that are perceived more quickly, accurately, and draw our attention to make this information more salient (i.e. noticeable) and discriminable (Cleveland and McGill, 1984). Spatial position is perceived the fastest and most accurately, and thus the most important attributes should be encoded by their position in a visualization. After spatial position, designs should consider the 10 hierarchy of salience for non-spatial visual features. For example, size features such as length and area are more salient than colour features such as hue, luminance, and saturation. For a comprehensive breakdown of this hierarchy see Munzner (2014).
-Choose designs that are accessible and effective for common types of colour blindness. For example, red-green colour blindness (deuteranopes) affects roughly 8 % of males of European descent (Birch, 2012).

15
-Interaction reduces cognitive load and helps users understand data by asking questions and performing queries. A practical guideline for designing interaction idioms is the visual information seeking mantra of Shneiderman (1996): "overview first, zoom and filter, then details on demand". The initial visualization should provide an overview of the entire dataset, and then the interactions should allow the users to change the view to see subsets of the data, and then visualize details about features of interest. This design approach offers users a flexible way to explore data, while being able to maintain a sense of context and orientation.
-Comparison tasks are often more effective when seeing multiple frames in a single side-by-side view rather than changing views over time. The human perceptual system is effective at reading spatial information in parallel, whereas changing 5 views with animations or multiple tabs relies on human memory and results in substantial cognitive load (Ware, 2012).
-Visualization idioms should present data with the smallest number of spatial dimensions, avoiding three-dimensional visualizations and using one-dimensional lists where possible. Displaying three-dimensional data on planar surfaces has numerous issues with depth-perception and over plotting (Ware, 2012).

Applications of visualization design principles 10
This section presents applications of the visualization design principles using simulated snowpack data for 8 January 2018 in Glacier National Park, Canada. On this day the avalanche danger rating was considerable at all elevation bands with two avalanche problems (Parks Canada, 2018): a storm slab problem at all elevations (size 1 to 2 avalanches were possible to likely) and a persistent slab problem at treeline and below treeline elevations (size 1 to 3 avalanche were possible to likely). Simulated profiles were produced by forcing the physical snowpack model SNOWPACK (Lehning et al., 1999) with gridded 15 meteorological data from the Canadian HRDPS numerical weather prediction model (Milbrandt et al., 2016). Meteorological data was extracted at 236 grid points in the park and at each grid point a single flat field profile and four virtual slope profiles were simulated (38 • slopes in four cardinal directions). A total of 1180 profiles covering an area of 1354 km 2 provide a sample data set to present visualizations of regional snowpack conditions.

Identify snowpack structure patterns with colour 20
Snowpack features related to avalanche problems should be easy to identify in visualizations of snowpack structure. The standard colour palette for snow grains (i.e. Fierz et al., 2009) creates undesired emphasis on certain types of snow. Important features such as thin weak layers have relatively low perceptual salience while less important features such as melt forms and ice formations have relatively high salience. The colours also make it difficult for individuals with colour blindness to distinguish important features. 25 We propose a perception-informed colour palette for snow grain types that emphasize features related to avalanche problems (Table 1). Similar perception-informed colour palettes have been proposed to improve the interpretation of visualizations in meteorology and oceanography (Stauffer et al., 2015;Thyng et al., 2016). The proposed colour palette groups grain types into four categories based on their role in avalanche problems: persistent weak layers (surface hoar and depth hoar), new snow layers (precipitation particles and decomposing and fragmented particles), bulk layers (rounded grains and faceted crystals), and melt 30 and ice form layers. These groups were visually related using analogous color schemes (e.g. the hues are perceptually close to each other) that remained visually discriminable. The visual salience of these groups was adjusted using properties of color such as how dark they appear (i.e. luminance) and how vivid the colors are (i.e. saturation). In this way a visual hierarchy of importance was created. Weak layers that tend to take up the smallest area were made the most salient by using strong contrast against other grain types, next new snow was made salient. Finally, the other layers formed the lowest level of perceptual salience and serve as a neutral background. All colors were made to be perceptually distinct. Accessibility for common types 5 of colour blindness was also considered (see Table 1). Unique colours were also assigned to melt-freeze crust and rounding faceted particles, as distinguishing these sub-classes was deemed meaningful for avalanche forecasters. A simplified colour palette was also designed using only the four main categories of grain types non-model experts ( Table 2). The simplified palette uses analogous colours to the full palette and maintains a similar visual hierarchy. Table 1. A perception-informed colour palette for snow grain types that emphasizes features related to avalanche problems 10 and is effective in grayscale and for common types of colour blindness. Table 2. Simplified colour palette for groups of grain types related to avalanche problems.
The colour palettes were tested with common visualization idioms such as hardness and timeseries profiles (Fig. 2). Comparing 15 the standard and redesigned colour palettes shows how the new palettes simplify the interpretation of the profiles by drawing attention to the most important snowpack features on 8 January 2018. The increased salience of the thin depth hoar layers highlights a potential persistent slab avalanche problem and the new snow highlights a potential storm slab avalanche problem.

Identify avalanche problem types from multiple profiles
Visualizing information from an ensemble of snow profiles is an effective way to identify snowpack patterns in a forecast 20 area. Identification and summarization tasks can be done fast and effectively by deriving visual summary statistics from distributed visual information. For example, humans can visually calculate correlation coefficients, clusters, and averages with their visual perception systems (Szafir et al., 2016). The volume and continuity of data produced by snowpack models offers new opportunities for summarizing snowpack structure that are not possible with human observed snow profiles. When  Figure 2. Comparison of timeline and stratigraphy profiles with standard colours for grain types, perception-informed colours for grain types (Table 1), and perception-informed colours for grain type groups (Table 2).
used in combination with a colour palette that emphasizes snowpack features related to avalanche problems, profile ensemble visualizations can help forecasters identify prominent avalanche problem types.
A simple and powerful summary is obtained by plotting multiple grain type profiles side-by-side (Fig. 3). In this example, 1180 profiles are sorted from thinnest to thickest and over 46 000 individual snow layers are shown in a single view. Despite the large volume of data, and a few prominent snowpack features pop-out and attention is drawn to the main snowpack patterns in 5 the forecast area. Since this visualization is specifically designed for the task of identifying potential avalanche problem types, other idioms are required for visualizing geospatial patterns in a meaningful way (see Sect. 3.3).
Another summary visualization that draws attention to potential avalanche problem types is produced by aggregating layers by their age or deposition date (Fig. 4). Simulated profiles can be aligned and aggregated by the deposition date of each layer to summarize the main features amongst a set of profiles. The prevalence of different grain types is determined by counting the 10 percentage of profiles containing grain types for each day the season. Grain types associated with persistent weak layers are emphasized with a diverging horizontal scale to distinguish them from other grain types. The persistent weak layers are also easier to notice in this visualization because they occupy a greater area than in Fig. 2 and 3 where their size is proportional to layer thickness. While it is also possible to produce an aggregated stratigraphy profile from aligned layers (e.g. Hagenmuller   Each layer is given a random horizontal position within the bin to allow visual summary statistics. Grain types are coloured using the perception-informed palette from Table 1. grains on the surface and a potential persistent slab avalanche problem is apparent from the salient surface hoar and depth hoar layers that are buried 30 to 50 cm below the surface (Fig. 3) and formed in early December 2017 (Fig. 4).

Locate avalanche problems in terrain
When locating avalanche problems in terrain, the description of the terrain depends on the context and scale of the forecast (Statham et al., 2018). For example, regional forecasters describe terrain by elevation bands and aspects while highway 5 forecasters reference named avalanche paths. Partitioning snowpack data into distinct terrain classes and comparing sideby-side views of the data for each terrain class is an effective way visualize complex geospatial patterns. High-dimension (3D) visualizations are tempting to characterize mountainous terrain, particularly with high density model datasets, but there is large potential for misinterpretation on two-dimensional displays due to depth perception issues and over-plotting (Ware, 2012). Instead, using eyes to simultaneously compare visualizations for different types of terrain has low cognitive load and 10 less potential for misinterpretation.
To provide insight into the spatial distribution of avalanche problem characteristics, the simulated profiles from Glacier National Park were partitioned into bins for elevation band and aspect classes to support regional-scale forecasting (Fig. 5). A randomized horizontal position (i.e. jitter) was applied to each layer to reduce over-plotting and randomize the order within a bin. The jitter plot allows the user to derive visual summary statistics about the snowpack structure in each terrain class and make comparisons between different terrain bins such as: snow depth generally increases with elevation, except for south and west facing slopes in the alpine, there is more new snow on north and east aspects, buried surface hoar layers are more prevalent on north and east aspects, and 5 the early December 2017 weak layer is more prevalent at treeline and below treeline elevations.
These types of visual patterns could help forecasters localize avalanche problems in their terrain. Different types of terrain bins could be applied for other forecasting contexts to highlight differences between relevant types of terrain. Examples include sub-regions, avalanche paths, or classes of ski terrain (e.g. Sterchi et al., 2019).

Compare distributions of avalanche size and likelihood
10 Avalanche size is easily visualized by aligning layers by depth rather than height. Layer depth is more relevant to forecasting avalanches than layer height, as weak layer depths correlate to the destructive potential of slab avalanches (McClung, 2009).
From an information visualization perspective, comparisons are more effective on aligned scales, and thus aligning layers by depth allows users to browse the distribution of depths for specific weak layers. From the distribution of layer depths in Fig. 3 and Fig. 5, forecasters could estimate the potential sizes of storm slab and persistent slab avalanches.

15
The CMAH defines the likelihood of avalanches as a combination of sensitivity to triggers and spatial distribution (Statham et al., 2018), making it a relatively difficult attribute to visualize. Options for visualizing avalanche likelihood include encoding related attributes with visual features such as shape, size, or motion in any of the previous idioms or by designing new idioms that focus specifically on likelihood. Information about the spatial distribution of a problem can be derived by counting relevant features amongst a set of profiles (e.g. Fig. 4). Sensitivity to triggers is an assessment of snowpack instability, which snowpack 20 models estimate using stability indexes based on the mechanical and structural properties of the layers (Schweizer et al., 2006).
We derive a relative measure of sensitivity to triggers (S) from SNOWPACK's structural stability index (SSI). The SSI combines a stress-strength ratio with differences in hardness and grain size to calculate a value between 0 and 6, where lower values correspond to less stable layers. To visually emphasize unstable layers, SSI was transformed into a relative measure of sensitivity to triggers: where the SSI for each layer is scaled inverse exponentially to produce an ordered variable that correlates with the sensitivity categories from the CMAH (i.e. unreactive, stubborn, reactive, touchy). This transformation produces values between 0 and 1 and exaggerates differences for weak layers with low SSI. The numeric value of the sensitivity measure does not have an interpretable meaning but illustrates relative patterns when applied in visualizations.   Figure 6. Providing information about the likelihood of persistent slabs avalanches by scaling the size of each layer's dot with its sensitivity to triggers (derived from the structural stability index). Snowpack layers from 1180 simulated profiles are partitioned into terrain class bins for elevation band and aspect. Elevation bins include alpine (ALP), treeline (TL), and below treeline (BTL) and aspect bins include four cardinal directions (north, east, south, west). Each layer is given a random horizontal position within the bin to allow visual summary statistics. Grain types are coloured using the perception-informed palette from Table 1. We present two examples of visualizing likelihood information with this relative measure for sensitivity to triggers. The terrain class visualization in Fig. 5 was modified to scale the dot size of each layer to its sensitivity to triggers (Fig. 6). This creates greater emphasis on sensitive weak layers, where the number and size of weak layer dots in a terrain bin relate to the likelihood of persistent slab avalanches in that type of terrain. Another visualization specifically designed for likelihood is given in Fig. 7, where the left panel provides information about the spatial distribution of each layer and the right panel provides 5 information about their sensitivity to triggers. Spatial distribution is shown by the prevalence of each layer by age (i.e. Fig.   4), while sensitivity to triggers is shown with the distribution of the relative sensitivity of each layer by age. The side-by-side comparison of spatial distribution and sensitivity to triggers provides information about the potential likelihood of persistent slab avalanche problems. For example, the weak layers that formed in early December 2017 are more widely distributed and sensitive to triggers than the weak layers that formed in late October (i.e. avalanches are more likely).

10
It is important to note that we are presenting these likelihood visualizations more to illustrate the concept than as a practical decision aid. It is known that the modelled stability index does not provide meaningful information about layers near the surface where storm slab avalanches occur (Schweizer et al., 2006), andMonti et al. (2014) has highlighted issues between the modelled stability indices and field observations of snowpack instability. At the domain situation level, creating links between snowpack models and the CMAH addresses operational challenges faced by avalanche forecasters (Statham et al., 2018). Reflecting the broad adoption of the CMAH, the proposition of using snowpack models to characterize avalanche problems has gained more interest from the Canadian forecasting community than 10 snowpack model tools produced over the past decade. The CMAH may not characterize the domain situation for all possible snowpack model users, as problems such as terrain selection or civil protection likely require distinct design considerations.
At the task and data abstraction level, the visualization of snowpack summaries has received consistent positive feedback from forecasters. The side-by-side profile summary (Fig. 3) visualizes snowpack patterns in a way that is not possible with traditional snow profile data and can help forecasters build a mental model of the snowpack structure in their forecast area.

15
Other operational tasks could benefit from bottom-up designs that leverage the spatial and temporal coverage of snowpack models, such as using stratigraphy timelines to visualize temporal trends.
At the visualization and interaction idiom level, some forecasters suggested reducing the number of colours in snow profile visualizations to make them easier to interpret. The perception-informed colour palettes (Table 1 and 2) achieve this while following established perceptual and cognitive principles to draw attention to the most important features. The user testing 20 exercise evaluated the users' ability to interpret the visualizations by performing simple tasks with the interactive dashboard.
Four out of five participants correctly performed task such as comparing snow height over different elevations, identifying the depth of prominent weak layers, and summarizing new snow amounts. The remaining participant made mistakes with filtering and selection, highlighting the importance of designing interactions that are simple and intuitive.
At the algorithm level, the operational prototypes provided daily updated visualizations in a timely manner with fast response 25 time for interactions. The main concern at the algorithm level was the prototypes were accessed externally from existing workstations, which created a major barrier to access. Integrating snowpack model visualizations into forecasting workstations is a critical next step. Testing in an operational setting would allow further validation at the domain and abstraction levels by measuring user adoption and observing how designs are used to perform operational tasks. Although the designs presented in this paper follow established visualization principles, testing in real forecasting scenarios is needed to validate their actual 30 operational value.
We present visualization design principles that increase the interpretability and relevance of snowpack model outputs. These are two of the four major perceived issues with operational snowpack model tools identified by Morin et al. (in press). The nested model for visualization design (Munzner, 2009) provides a framework for defining the domain of avalanche forecasting and the necessary tasks that are needed to analyze data. Tasks required to assess avalanche hazard are described by applying the widely 5 adopted conceptual model of avalanche hazard (Statham et al., 2018). From these tasks, we apply information visualization principles to design visual representations of snowpack model data in ways that leverage the human visual system to understand the complex nature of the data. Preliminary feedback from avalanche forecasters suggests these designs are easier to interpret and provide more relevant information than previous visualizations of snowpack model data.
A key idea in these designs is shifting from bottom-up scientific visualizations towards information visualizations that 10 address user needs. As highlighted by Grainger et al. (2016), other types of environmental models would likely see improved adoption by shifting towards information visualization. When using numeric models as a tool for assessing natural hazards, visualizations will be more effective when the designers make links to established risk frameworks and carefully consider the tasks performed by operational decision makers.
A critical next step is implementing these designs into operational forecasting workflows. By addressing issues with the 15 interpretability and relevance of snowpack model data, these designs will allow forecasters to learn the capabilities and deficiencies of snowpack models in a meaningful way. The same design principles should be considered when visualizing other types of avalanche and snowpack data, as the same domain situation and task abstractions apply when analyzing field observations. Interaction idioms should play an important role in understanding of complex model data, as they allow users to perform custom queries, test and validate hypotheses, and discover inconsistencies and anomalies. Interactions that compare 20 model data with observation data would be particularly powerful in building trust in the models and addressing issues with their integrity. This process was critical in the adoption and trust in numeric weather predictions models by meteorologists (Benjamin et al., 2019), and just like meteorologists, avalanche forecasters could become active participants in model validation and improvement.
. The code and data used to produce the visualizations are published as a data file containing the simulated profiles and an R script that 25 produces each of the visualizations. The interactive dashboard is available at https://avalancheresearch.ca/pubs/2019_horton_snowpackvis.
. All authors worked on the conceptualization of this paper. SH prepared the data and software, SN contributed to visualization ideas and designs, and PH provided supervision. SH prepared the manuscript with review and editing from the other authors.
. The authors declare no competing interests.