To explore the reliability of avalanche size estimates provided by avalanche practitioners, we developed a survey consisting of 10 photographs of avalanches (see the Supplement). The photographs used were originally captured for the purpose of documenting avalanches. They show clearly identifiable avalanches in various terrain and were chosen to provide a diverse set of perspectives, sizes and illumination, as well as the presence of various reference objects. In the survey, each participant was asked to estimate the size of each of the 10 avalanches using the five-class integer scale, where we refer to estimates as “full” sizes (for instance, size 3; see Table ). After estimating the full size of an avalanche, participants had the opportunity to provide an intermediate size (“half” size, nine levels; for instance, size 2.5). As a second task, we asked participants to rate the importance of the factors characterizing avalanche size for their size estimations on a four-point Likert scale as either very important, important, less important or not at all important (factors: destructive potential, dimensions, runout and volume; Table ). We designed the survey with a European audience in mind and only later decided to extend it to North America. For this reason, runout and volume were included as factors even though they are not part of the North American avalanche size definition. Similarly, in the European definition typical length and volume are presented under the headline typical dimensions , a term which is not present in the North American definition.

The survey was sent to avalanche practitioners, primarily regional avalanche forecasters in Europe and North America, through personal contacts or using forecast center mailing lists. The survey was available in English, French, German and Italian. We asked participants at the beginning of the survey whether they were avalanche forecasters and in which country they work. In total, 170 responses were received: 105 from Europe and 65 from North America. The proportion of professional avalanche forecasters in our dataset was 86 % (146). The other 24 participants either had additional roles besides forecasting or worked closely with the avalanche warning service, for example as avalanche educators, mountain guides, ski patrollers or field observers for a warning service. The professional forecasters in our sample were from the United States (39 participants), Italy (33), Canada (17), Norway (15), Spain (10), Austria (10) and Switzerland (7), while all other countries had 2 or fewer participants.

To investigate the reliability in avalanche outlines, we asked nine people, who map avalanches as part of their professional duties, to map six avalanches in the area around Davos (Switzerland) from winter 2020/21.

For each avalanche, we provided three to six photographs and indicated the approximate location by giving the name of a ridge or summit in the proximity of the avalanche (distance 50 to 300 m). Mapping was conducted using operational mapping tools, which provide the user with a topographic map at best with a scale of 1:10000;, orthophotos resolution of 10×10 cm; and slope incline classes for areas steeper than 30∘ (resolution 10 × 10 m). Each participant was asked to map the six avalanches with the same accuracy as they usually would when mapping avalanches. In addition to the nine participants, we used the avalanche outlines that were initially mapped for documentation purposes in winter 2020/21.

To create a reference for the map analysis, we georeferenced one image per avalanche with the monoplotting tool developed by , then drew and exported the avalanche outlines. Since this approach allows for a much more accurate localization of avalanche outlines, we used these as a reference in this study. For one avalanche (Fig. a), the deposit was obscured by a tree in the only photograph where the whole avalanche was visible. This part of the avalanche was therefore disregarded in the analyses including this reference.

The third experiment concerned the mapping of avalanche outlines from remotely sensed imagery. In addition to comparing mapped avalanche outlines between individuals, this experiment allowed for exploring some of the potential factors influencing the quality of mapped avalanche outlines illumination, snow conditions, avalanche type, image resolution;.

We selected two georeferenced images acquired under different snow conditions (see Table ) without artifacts, without saturation and as 16-bit radiometric information. The images were processed in Agisoft Metashape. To obtain different resolutions, we bi-linearly resampled the data in the red–green–blue channel (RGB) to 25 cm, 50 cm, 1 m and 2 m spatial resolution (for native resolutions, see Table ). For separating illuminated from shaded areas, we used a support vector machine classifier to calculate a shadow mask seefor details.

We provided a standardized introduction to the five participants, who were all familiar with avalanches and remotely sensed imagery. All visible avalanches were to be digitized in the software ArcGIS Pro, starting from the coarsest (2 m) and ending with the finest resolution (25 cm). Images with higher resolution were only made available after the mapping of the (one-step) coarser resolution had been completed. Participants could not re-examine their earlier mappings. They had access to the topographic map at best a scale of 1:10000; and slope incline classes for areas steeper than 30∘ (resolution of 10 × 10 m). We instructed participants to outline all areas with signs of avalanche activity rather than drawing individual events; thus, they were asked to delimit all visible avalanche regions but not to separate them into individual avalanche polygons. The participants did not see the mapped outlines from other participants before they had finished with the highest resolution.

A total of 170 people participated in the survey and estimated the size of 10 avalanches, shown in Fig. . The agreement rate between any two size estimates Pagree was 0.53, ranging from 0.22 to 0.68 for individual raters. Nine of these raters had an agreement rate lower than the 95th percentile of the 170 participants (Pagree≤0.39), indicating particularly low correspondence with avalanche size as perceived by others. Each of these nine raters suggested at least for one avalanche a rather “extreme” avalanche size, a size which less than 10 % of the participants had chosen. Without these nine raters, the agreement rate was 0.54, which is only marginally higher than the overall agreement. Considering all responses, the mean reliability rel of individual estimates was 0.72, ranging from 0.47 to 0.82, and, if excluding the nine raters with the lowest agreement with others, rel was 0.73.

On average, the agreement with the size considered correct (s‾j) was Pcorrect= 0.74, or, if treating a simple most frequent vote smode as the reference size, Pmode was 0.66; 16 participants were in full agreement (Pcorrect=1) with the avalanche sizes considered the most likely size s‾j, while the nine raters with the lowest agreement with others also had low values of agreement with s‾j. Excluding these resulted in Pcorrect= 0.76.

In addition to the 66 % of the respondents who provided the same size estimate as the most frequently chosen size smode, another 29 % chose the second-most popular neighboring size. Thus, in total 92 % of all estimates fell into two adjacent size classes highlighting that there was a reasonable consensus on the most likely size(s). Relaxing the definition for agreement even more as in, 97 % of the responses were size smode±1, ranging from 46 % for Fig. b to 98 % for Fig. f. The average number of different full size classes chosen was 3.7, ranging between two for Fig. f and five for Fig. j. The latter example means that each of the five size classes were indicated at least once. This shows that even though most votes were in correspondence with one of the two most frequent size classes, at least some estimates regularly deviated strongly from this opinion.

An intermediate size class was given in 26 % of all cases. The agreement of the intermediate size estimated by a respondent with the most frequently indicated intermediate size smode.intermediate was 0.49, and for smode.intermediate±0.5 the agreement was 0.74, while Pcorrect was 0.53. The most frequent intermediate size was always between the two most frequent full sizes, underlining that a share of participants differed in their estimates less than a full size (Fig. ). The mean agreement rate Pagree, when allowing for full and intermediate sizes, was 0.37; the reliability rel was consequently 0.61.

To explore if the size of an avalanche relative within an image and in relation to the surroundings influences size estimation, we included one avalanche twice though the image was cropped and flipped (avalanches in Fig. a and c); 168 out of 170 participants rated both avalanches. Of those 168, 78 % indicated the same size, and 15 % rated the avalanche one size larger in the close-up view in Fig. c than in the overview in Fig. a, whereas 7 % rated the avalanche one size smaller in the close-up view compared to the overview. The shift in the proportions is statistically significant (proportion test: p=0.036).

When comparing the results from Europe and North America, we found the agreement of individual raters in the most frequently estimated size smode to be identical (Pmode=0.66). This approach slightly favors European respondents, as these contributed a larger share of responses (Europe: N=105, North America: N=65). Considering s‾j instead, the agreement Pcorrect was 0.74 overall and 0.66 for both Europe and North America individually. The higher overall agreement results from the definition of the reference size s‾j, where two sizes were considered correct if the reference size was located in between two values. North Americans had a tendency to assign smaller sizes than their European counterparts. This is most notable for the three largest avalanches (avalanches in Fig. b, g and i), with a median size of 4 by Europeans and a median size of 3 by North Americans. With the exception of Fig. e and f, differences in size estimates were statistically significant (Wilcoxon rank-sum test: p ranging from 0.045 to <0.001). Within their continents, respondents had a similar agreement with each other (proportion test: p=1): on average, Pagree was 0.53 within Europe and 0.56 within North America, resulting in rel of 0.73 and 0.75, respectively. Intermediate sizes, which are more commonly used in North America, were chosen in 31 % of the cases by North Americans compared to 23 % by Europeans (proportion test: p=0.368). When using intermediate sizes, the agreement with the mode intermediate size smode.intermediate was 53 % for Europe (±0.5: 75 %) and 49% for North America (±0.5: 81 %).

Among the factors used to determine avalanche size (Table ), runout was considered the most important with 56 % of respondents considering this factor as very important, followed by volume (very important: 39 %), dimensions (29 %) and destructive potential (20 %) (Fig. ). Comparing responses from Europe and North America, we found the most frequent response to be identical for all four factors (very important for runout, important for the other three). However, runout was considered significantly less often as very important in Europe (46 %) compared to North America (72 %, proportion test: p=0.001). The factor volume showed a similar pattern with significantly more votes from North America (46 %) than Europe (18 %) for being very important (proportion test: p<0.001) and the opposite pattern for volume being either less important or not at all important (Europe: 39 %, North America: 7 %, proportion test: p<0.001). The differences between the continents for rating the importance of destructive potential and dimensions were not significant (proportion test with rating of very important: p=0.260 and p=0.718).

The 10 participants, all very familiar with the study area, centered the avalanches based on oblique photographs around the corresponding reference mapping (Fig. ), identifying the correct locations.

On average, the overlapping proportion of the mappings of any two participants of IoUpairwise was 0.52, varying from 0.32 for the worst pairwise agreement to 0.69 for the best one (Table ). Individual pairwise comparisons are shown in Fig.  in the Appendix. When comparing individual mappings to the area mapped by at least one person as AoU, the mean IoUall is 0.31, ranging from 0.21 to 0.41. Only a fraction of 9 % of the combined area of union (AoU for all 10 participants) was identified by all participants as an avalanche (area of overlap for all 10 participants, AoO) showing the considerable scatter of individual mappings. Comparing individual mappings to the reference resulted in a mean IoUref of 0.60, with a minimum of 0.40 and a maximum of 0.80. These areas of higher agreement between participants, visible in darker hues in Fig. , coincide with the outlines from the reference mapping, highlighting that variations happened around the reference. In other words, individual mappings had a higher correspondence with the reference mapping compared to mappings by other individuals (see also Fig.  in the Appendix). The large variation between individual mappings was also shown when analyzing the absolute values of the mapped areas (Table ): the largest mapped area was between 2 and 4 times larger than the smallest mapped area (avalanches in Fig. b and f). Additionally, the comparison with the reference showed a systematic tendency towards underestimation of the area, as in all cases the median mapped area was between 10 % and 36 % smaller than the reference area.

Two of the raters had statistically lower pairwise overlap in their area of avalanche activity (mean IoUpairwise≤ 0.49) compared to the other eight raters (mean IoUpairwise≥ 0.69, Wilcoxon rank-sum test: p=0.011 and p=0.002; see Fig.  in the Appendix). These two raters also had the lowest agreement with the reference mapping (IoUref≤ 0.44), lower than the other eight (IoUref≥ 0.53).

When visually comparing the mappings, differences can be observed not only between image resolutions (Fig. ) but also between participants (Fig. ). The mean of the pairwise overlapping proportion of avalanches of IoUpairwise increased with increasing image resolution from 0.46 at 2 m resolution to 0.68 at 25 cm resolution (Table ). Considering the area classified as an avalanche by three or more raters as the reference showed an increase in IoUref from 0.64 at 2 m resolution to 0.80 at 25 cm resolution (Table ). Regarding the influence of illumination conditions, all IoUpairwise scores were higher in illuminated areas compared to shaded areas of the image (for instance, at 25 cm resolution – illuminated: 0.77, shaded: 0.54; Table ). Snow conditions also influenced the agreement of the mappings (Figs.  and ): the mean IoUpairwise was higher in wet-snow conditions (25 cm resolution: 0.90) compared to dry-snow conditions (25 cm resolution: 0.66; Table ). Moreover, individual mappings were also much more similar in wet-snow conditions compared to dry-snow conditions with the variations in IoUpairwise ranging for dry-snow conditions at 25 cm resolution between 0.56 and 0.77 (mean: 0.66; standard deviation: 0.07) and for wet-snow conditions between 0.88 and 0.91 (mean: 0.9; standard deviation: 0.01; Fig. d). Overall, variations in mean IoUpairwise were smaller across resolutions (0.02 to 0.22) than the differences between the minimum and maximum IoUpairwise within one resolution (0.20 for 25 cm resolution to 0.43 for 2 m resolution). This is especially pronounced for dry-snow conditions (Fig. ). The large variations between different experts are also reflected in the avalanche area that was consistently identified by one person over all four spatial resolutions (dark red in Fig. ).

One of the five participants had a lower pairwise agreement compared to the other four (for instance at 2 m resolution: IoUpairwise≤ 0.41 vs. IoUpairwise of 0.37–0.76), although this was not significant (Wilcoxon rank-sum test: 2 m resolution of p=0.088, 25 cm resolution of p=0.055). Considering the area classified as avalanche by three or more raters as the best approximation of a ground truth, the mean agreement with this mapping ranged between IoUref=0.64 and IoUref=0.80 (Table ). Again the same participant had the lowest mean agreement.

Our results show that it is difficult to achieve consistent size estimates of avalanches: in only 53 % of the cases did any two size estimates agree, and in 66 % of the cases an individual estimate agreed with the most frequent size among the respondents (mode). It showed, however, that in most cases disagreements were comparably small with 92 % of individual estimates being either equal to the mode smode or equal to the second-most frequent neighboring size or, if considering intermediate sizes, that 74 % of the estimates were within one intermediate size class. Comparing our results to previous studies investigating agreement for avalanche size estimates (Table ), the agreement rate with the most frequent size ranged from 62 % to 84 % . The high agreement rates in and are probably related to the fact that these studies relied on a small number of experienced practitioners with a comparably similar background and training. In contrast, both the studies by and our study included participants from numerous countries and thus respondents with different avalanche backgrounds, leading to a more diverse group of avalanche practitioners. Moreover, speculated that the changes introduced in the avalanche size definitions shortly before their survey may have lowered the agreement. The reliability of an individual size estimate in this study was 0.73, highlighting the uncertainty associated with these data. Thus, using size estimates by an individual as ground truth when developing or evaluating models, a perfect model can achieve more than 73 % accuracy only by chance if the errors a model makes are independent from the errors contained in the avalanche size labels.

In our survey, we found the lowest agreement with smode for the three largest avalanches (s‾j≥3.5) (avalanches in Fig. b, g and i). For these three avalanches, smode differed between Europe and North America. argued that practitioners have more experience with smaller avalanches (sizes 1, 2 and 3), which are much more frequent than larger avalanches, which may cause size estimates of large avalanches to be more variable and less accurate.

We noted systematic differences between size estimates provided by North Americans and by Europeans, with North Americans tending towards smaller sizes (Fig. ). This might stem from differences in the European and the North American definitions (see Fig. ): for the typical length the European definition provides a range, whereas in the North American definition a typical value is given. The European definition encompasses the North American values, describing typical length for the smaller avalanches, while it coincides with the upper bound for size 4 and provides only a minimum value for size 5. Another difference is that the North American definition includes typical mass, in line with the definition introduced by , while in Europe typical volume is defined. Combining deposit volume with density measurements or density estimates of the deposit, mass may be determined (mass = volume × density). For instance, calculating the mass of avalanches assuming a mean density of 390 kg m-3, measured from 95 avalanches at Rogers Pass, British Columbia , avalanches are almost 4 times larger in the European compared to the North American definition; 4 times larger corresponds approximately to a half size (e.g., for size 2 the mass according to the definition is 102=100 t; for size 2.5, it is 102.5=316 t). Consequently, the significant intercontinental differences may be, at least partially, attributed to the differing size class definitions.

We also observed differences in the importance ranking assigned to the factors determining avalanche size, with both the criteria runout and volume being considered more relevant for size estimation by North Americans compared to Europeans. This is particularly noteworthy as neither a description of runout nor an indication of volume are part of the North American size definitions (Table ). We found the size-determining factor of destructive potential to be considered the least important by North Americans and the second-least important by Europeans. This is surprising as both the definitions in Europe and North America state that avalanche size is classified according to destructive potential e.g.,. Furthermore, this finding is also contradictory to that in the study by , where destructive potential was the highest-rated factor. One possible reason for the low importance of destructive potential in our study might be related to the study design, where supplementary information about damage to property or people, beyond the photographs, was absent. Thus, destructive potential had to be inferred from avalanche properties like width, length and volume, which is arguably the normal situation when estimating avalanche size.

To find out if the way an avalanche is shown in an image influences size estimations, we included one avalanche twice, changing the perspective and zoom. Even though 78 % rated the avalanche in Fig. a and c the same, we observed a significant proportion of larger estimates in the close-up view. We suspect that this might be caused by the perception of the avalanche being larger when covering more area in the photograph. But our sample is small, and understanding the effect of perspective and area covered by avalanche would require further investigations with a more meaningful sample.

Study 2 required participants to first find the location of the avalanche on the map, matching the topography visible in the images with the topography as shown on the map, before mapping was possible. The 10 experts, all very familiar with the study area, located the avalanches in the same place. This first step of the assignment would have been more difficult for someone that did not know the area well, possibly resulting in entirely different locations and hence mappings. Thus, the mean overlapping proportion of any two mappings (IoUpairwise), which was 0.52 in our study, may potentially be 0 if an avalanche is located in the wrong place. We therefore assume that an IoUpairwise of 0.52 may well describe the upper limit of agreement in mappings from oblique photographs.

Study 2 allowed for a comparison with a reference mapping using a methodology superior to the approach the 10 experts used. Thus, the agreement between experts' mapping and the reference mapping can be interpreted as the proportion correct and, hence, allows for assessing the experts' competence in mapping avalanche outlines. The overall proportion considered correct (IoUref) was 0.6, with a clear bias towards smaller mapped areas compared to the reference. The results showed that experts were not equally competent, with the proportion correct ranging from 0.4 to 0.8 (Table ). The agreement between individual mappings and the reference mapping is larger than the agreement in the mappings between participants (IoUpairwise=0.52). This means that the reliability of individual mappings would be underestimated when relying on a measure like the agreement rate between domain experts. If competence is known, it would be possible to weigh individual mappings if two or more mappings were available, likely resulting in more reliable results. Overall, we consider the mapping of avalanches using oblique photographs a challenging task to perform consistently and accurately.

In study 3, five participants had to identify the avalanches in the remotely sensed imagery, for each of four image resolutions, before mapping them. In other words, whether a point is identified as an avalanche is a combination of existential and extensional uncertainty , addressing the following questions: is there an avalanche, and where are the boundaries? This uncertainty was lower with higher image resolution, and for illuminated compared to shaded parts of the image, it allowed participants to identify avalanche area more consistently and confirmed the findings of earlier work . Moreover, snow conditions influenced agreement too, with higher agreement under wet-snow compared to dry-snow conditions (Fig. ). We suspect that this difference is caused by the presence of liquid water in the case of a wet snowpack, which leads to more pronounced avalanche boundaries compared to dry-snow conditions.

Another important finding from these studies are the large differences in the areas mapped as an avalanche by the experts. For the six avalanches in study 2, the largest mapped area was between 2 and 4 times larger than the smallest mapped area (Table ). For study 3, variations across resolutions were found to be smaller than the variations in IoUpairwise per resolution, suggesting individual experience and competence have a larger impact than the underlying spatial resolution. If avalanches would be classified automatically using area or by extracting width and length from the mapping e.g.,, completely different size classes may result due to these variations. For instance, comparing the mappings of 4000 avalanches with the reported size estimate, showed that the median area of size-2 avalanches was about 3 to 5 times larger than size-1 avalanches or that the mapped area of size-4 avalanches was about 7 times larger compared to size-3 avalanches pp. 49 and 51. Comparing these values to the variation observed in the mappings by different experts in our study suggests that estimating avalanche size based on mapped area would, quite frequently, result in different size estimates.

Several sources of error may impact the reliability of tasks involving human judgment and estimation. These are related to the data being suitable; the skill or competence of the rater interpreting the data; and, finally, the level of generalization (e.g., forcing avalanches into five size classes). In the context of size estimation and outline determination, low image resolution, lack of image context/reference objects and unfavorable illumination conditions such as shade or diffuse light may impact the conclusiveness of the data for the task to be performed and, hence, reliability. Similarly, differences in the raters' backgrounds and experiences may lead to different levels of competence in interpreting the data. Furthermore, variations in the perception and definition may result in systematic variations in size estimates or outlines. Moreover, the binary choice between an avalanche and no avalanche, when mapping avalanche outlines, does not allow for expressing uncertainty. This may lead to more pronounced differences in the avalanche area identified by participants. Addressing and being aware of these issues will improve the quality of avalanche size estimation and outline determination.

The results of this study indicate that size inventories from North America and Europe, different warning services within Europe, or simply different domain experts may systematically differ in their assessments. Consequently, transferability of size inventories between continents or different warning services may be limited. To achieve a common understanding and comparable size estimation, in particular for expert forecasters and for observers, we suggest a joint effort of the continental and national avalanche associations, together with avalanche forecasters and other avalanche practitioners, to develop training tools that help standardize the size estimations. An in-depth analysis of current protocols or training programs could be fruitful and serve as a first step in tackling this issue. One option might be training people in the “all observables approach” advocated by , which requires imagining the objects that might be destroyed in the track or start of the runout zone of an avalanche.

In the meantime, the uncertainty related to size estimates may be reduced taking into account second estimates and/or jointly deciding on the size in the case of disagreements e.g.,. Additionally, we recommend the use of intermediate sizes in the following way: first the (full) size class should be estimated. In a second step, the assessor may judge whether avalanche size is low or high or in the middle of the class like suggested for evaluative social judgments by. If it was low or high, this would result in the intermediate sizes between the chosen and the upper or lower adjacent full size. Practically, this could mean that from a full size 2 the assessor may, in a second step, assign size 1.5 if the avalanche is at the lower end of size 2, assign 2.5 if it is at the upper end or keep a full size 2 if it is a “typical” avalanche for that size.

While the observed variations in avalanche outlines may partly be attributed to a different background and level of experience, we argue that it is partly caused by the lack of a common, precise definition of where to delimit an avalanche. We are not aware of any unambiguous, actionable guideline of where exactly to place the visible outline. Arguably, there is no “natural”, self-evident definition, especially for dry-snow avalanches. Consistency, in the sense of repeatability across expert annotators, can perhaps only be achieved through a generally agreed consensus that includes shared but, to some degree, arbitrary conventions. It appears that a standardization effort may be beneficial and that standardized training could go a long way towards reducing the spread between different experts and organizations, even if some causes of variability, e.g., lighting conditions after a large snowfall cannot easily be controlled and will remain.

If reliable mappings from photographs are required, we recommend second mappings, jointly deciding on the extent of the outline, using a monoplotting tool e.g., or the overlay image capabilities of Google Earth. For remotely sensed imagery we advocate using a spatial resolution of 50 cm or finer for the detailed segmentation of specific avalanches, whereas a resolution of approximately 2 m seems to be sufficient to capture the overall avalanche activity over a larger region. Intermediate resolutions may provide a reasonable compromise between the precision of individual outlines and large-area coverage at a reasonable cost and immediacy. Recording the perceived uncertainty while mapping might help as well as using the area of agreement from several mappings or jointly discussing areas of disagreement. We generally recommend aiming for good illumination for mapping avalanches, especially under dry-snow conditions.

In all three studies, we relied on comparably small sets of images: in study 1, the size survey, we aimed at a high response rate, which came at the cost of a smaller selection of different avalanche examples (10) since we did not want to introduce a bias by showing different images to different participants. The two mapping tasks (studies 2 and 3) were rather time-consuming and were, therefore, limited to few participants and to few examples. They may therefore be regarded as pilot studies, whose results should be interpreted with caution, keeping in mind the comparably small datasets and the potential particularities of the data. In particular, biases may be present due to the homogeneous sample from participants with a similar background and training, as well as due to the chosen oblique photos and remotely sensed data.

For study 1, we provided photographs but no maps and no additional information, for instance on damage which may have occurred. This certainly made the size estimation task somewhat more difficult, as we often suspect either that a map is available or that the person is familiar with the avalanche path, which may both help in estimating avalanche dimensions. We did not provide maps together with the photographs, as we wanted to avoid introducing a bias related to the (un)familiarity with a specific map design. We have, however, tried to compensate the lack of an accompanying map through the presence of reference objects (people, trees, ski lifts, etc.) in our example photographs to help participants determine the potential damage and volume/mass. Nevertheless, photographs in Fig. a and e lack recognizable reference objects, and the other photographs (Fig. ) must be analyzed carefully to identify the ski tracks, people and houses. Agreement for avalanches next to clearly visible reference objects could therefore be higher than in our analysis.

For the avalanches mapped from oblique photographs (study 2), we speculate that study participants, being aware that their mappings will be analyzed in detail, may have paid more attention to finding the exact boundaries than during routine documentation work. We acknowledge that our sample size of six avalanche examples covers only a fraction of possible viewing angles, snow and avalanche conditions, and terrain characteristics. Furthermore, all participants were well acquainted with the study area and had extensive experience with mapping avalanches using oblique photos. Thus, we regard our results as a best-case scenario. Still, we believe that within the range of (fairly typical) conditions captured by our set of pictures, the evaluation is representative of avalanche outlines currently used in Switzerland. We encourage further research to ascertain the worldwide validity of the results.

Finally, we would like to point out that three of the authors were also involved as participants in the studies (one in study 1 and two in studies 2 and 3). Particularly in studies 2 and 3, with few participants, this may impact results favorably and suggests the presented findings should be treated as an upper bound.