Avalanche risk prediction: to ski or not to ski?
Image credit: Dreamstime
Machine-learning tools and data analytics could be key in assessing, and one day successfully predicting, snow avalanche hazards in mountainous regions. Ideal in preventing loss of life. Perfect for skiing enthusiasts.
Rapidly falling masses of snow, ice and rocks may be beautiful to witness from afar, but inevitably, avalanches pose a serious threat to anyone on snowy mountainsides. More than anything, they can be fatal due to their sheer force and seeming unpredictability.
In fact, snow avalanches claim an average of 100 lives in Europe every year, according to the European Avalanche Warning Services (EAWS) – March 2019 saw the last avalanche incident on UK soil with the tragic death of three climbers on Scotland's Ben Nevis.
According to Simon Trautman, an avalanche specialist at the US Forest Service’s National Avalanche Center and Northwest Avalanche Center in Washington, disastrous avalanches occur when massive slabs of snow break loose from the mountainside and shatter like broken glass as they race downhill. These hazards can travel faster than cars on a motorway, up to 100mph (160km/h)... it’s safe to say you don’t want to be at the bottom of a mountain or skiing down one when this happens.
Although the risks are proven to be deadly, this has not stopped many climbers, snowmobilers and skiers alike from participating in snow activities. Indeed, winter tourism is an important industry in the European Alps, which is the largest ski destination in the world, capturing 44 per cent of skier visits in 2017/18, according to the 2019 International Report on Snow & Mountain Tourism.
To help skiers and mountain-lovers to soak up the winter sun without the worry of getting caught in an avalanche, former software developer Günter Schmudlach developed a website for ski tourers which uses machine-learning algorithms to create an automated avalanche risk assessment for approximately 1,200 backcountry ski tours in Switzerland.
The tool, known as Skitourenguru, is used by backcountry (unmarked or unpatrolled areas outside a ski resort's boundaries) skiers and snowboarders to support rational decision-making in the planning phase before a ski tour, and whether to choose a less risky summit or even to not ski at all.
In many modern avalanche assessments, experts in the field often use the 3x3-Rule (Assessment and Decision Framework), designed by Swiss mountain guide and author Werner Munter. This method divides a ski trip into three phases: trip planning, local evaluation and individual slope. Throughout each phase, three factors – conditions, terrain and human – must be considered.
“During the trip, usually there is a gain in information,” says Schmudlach. “And in each stage, the skier decides upon all available information.
“We believe in the systematic usage of data and machine learning throughout the first stage, the trip planning,” he adds. “That particularly applies to the trip selection process.”
Skitourenguru targets backcountry skiers and snowboarders who know avalanche theory and actively practice avalanche awareness; winter sports enthusiasts still need to be able to cope with the challenges of the alpine terrain during the ski season.
Avalanche courses usually start with planning a specific route, but who provides the initial route suggestion? The instructor? “At home, you won’t have an instructor,” Schmudlach says. “Even those who know ‘their’ mountains well will have difficulties memorising 1,000 routes and make a rational route selection.”
This is where the tool could come in handy for these adrenaline junkies. Skitourenguru provides a customised list of ski tours that show a low avalanche risk based on data collected from multiple sources.
The algorithm used in the tool is based on a quantitative reduction method (QRM), which calculates the avalanche risk for each point of a route by combining a digital elevation model with the latest avalanche bulletin (updated twice a day).
To gather the data provided for the system, Schmudlach receives an avalanche bulletin at 8am and 5pm daily via an interface from the Institute for Snow and Avalanche Research (SLF). Then the QRM combines the information from the avalanche forecast with the characteristics of the terrain, and the tool evaluates different skiing routes in around ten minutes.
Regarding the data obtained from SLF, Skitourenguru compiled documents of daily avalanche forecasts over the past 19 years, with a total of 5,262 avalanche forecasts. The QRM was derived from data of about 1,800 severe avalanche accidents over this period as well as 50,000km of GPS tracks from real-life backcountry ski tours. Aside from Skitourenguru, this large set of data was also collected by mountain sports information portals Camp2Camp and Gipfelbuch.
Schmudlach explains that it is a challenge for avalanche forecasters to collect data directly from temperature and precipitation, which therefore means they use alternative methods: indirect meteorological data.
“It’s an old dream of avalanche forecasters to create avalanche forecasts directly from meteorological data like temperature, precipitation, sunshine etc. Regrettably, that’s not possible,” Schmudlach says. “Of course, avalanche forecasters use weather meteorological data, but only as one data source among many others.”
The data from SLF is collected from around 180 IMIS stations, as well as other MeteoSwiss automated stations, 200 observers, snowpack models, meteorological forecasting models and weather reports. As there is a huge volume of data, which can only be handled appropriately if the values are processed and visualised properly, Global Information System (GIS) technology is used as a basis for the interactive spatial data visualisation. In this context, both measurements and observations or assessments can be presented and statistical values can be calculated.
Skitourenguru uses this ready-made avalanche forecast of the SLF, and therefore this indirect meteorological data is processed within the algorithm of the tool.
“Collecting sensor data by crowdsourcing doesn’t make sense. It would just bring plenty of data without meaningful patterns,” Schmudlach argues. “The snow cover is by far more complex to be captured by some temperature or precipitation parameters.”
Subsequently, the results from this method are aggregated to a single risk-indicator. Like a traffic light, the risk-indicator identifies whether the route is of low risk (green), an elevated risk (orange) or high risk (red).
“Machine learning is definitively the answer to create avalanche risk prediction, if used during the first stage, the trip planning,” says Schmudlach. However, he adds, there is still scepticism about pushing the models into the second and third stage of Munter’s decision framework (local evaluation and individual slope).
“The information gained during the ski tour is of rather qualitative and subjective nature,” he explains. “Moreover, there is no process (and in most cases, no communications infrastructure) to enable the user to collect real-time data easily while in the outdoor context. That makes it difficult or impossible to collect training data and develop models.”
Skitourenguru calculates a risk-indicator for each of the 1,200 or so routes in Switzerland twice a day. From there, backcountry skiers and snowboarders can select routes from the list and plan these tours in detail.
When the backcountry ski tour is progressing, knowledge regarding snow, weather, terrain and people increases. With the additional knowledge, the result of the planning becomes less important and is replaced by a differentiated risk evaluation of every single slope on the routes picked out.
Regarding the success rate of the tool, Schmudlach says that around 500 to 1,000 people visit the platform every day, with around 10,000 of Switzerland's skiing community using the site regularly. “It’s hard to find a backcountry skier that doesn’t know the website,” he says.
Switzerland is currently the only area fully covered by the tool, but Schmudlach told Gipfelbuch in an interview that there are currently demo versions – which means the results are not reliable yet – of the tool that cover Austria, Bavaria, South Tyrol and north-east Italy.
“For next winter, the test stage is planned, the next winter is the full version,” Schmudlach tells Gipfelbuch. “Until then, it is important to build the route records.”
Furthermore, Schmudlach aims to allow users to digitise routes on the map themselves. The tool would rate these routes, and “if the user then does the ski tour, picks up a GPS track and uploads it, a second review could take place. This would give the ski tour operators feedback on the magnitude of the risks taken,” he says.
Although Skitourenguru uses machine learning to create avalanche risk assessments to help plan a ski trip, it does not predict the chances of an avalanche. However, a recent study by a group of researchers from a number of universities across the world – the University of Tehran (Iran), Oxford Brookes University (UK) and Texas A&M University (US) to name a few – deployed two machine-learning models to determine the chances of an avalanche.
According to the researchers, the results of both models, support vector machine (SVM) and multivariate discriminant analysis (MDA), showed that they had an “excellent performance in snow avalanche modelling”.
“Sensitivity analysis indicated that the topographic position index, slope, precipitation, and topographic wetness index were the most effective variables for modelling,” the team reports. “A snow avalanche map indicated that the high snow avalanche hazard zone was mostly near the streams and was matched with hillsides around the water pathways.
“Findings of study can be helpful for land use planning, to control snow avalanche paths, and to prevent the probable hazards induced by it, and it can be a good reference for future studies on modelling snow avalanche hazards,” the researchers add.
That being said, there is a still a long way to go for these models to become a key component in avalanche predictions. If anything, many experts believe that in-depth knowledge of the mountainous regions and a better understanding of what causes them still plays a vital role in avalanche predictions.
When combined with data analysis and machine learning, however, this prior knowledge and monitoring of conditions could one day pave a way to more accurate and quicker predictions, to prevent people from being caught in such dangerous conditions.
For now though, a risk assessment tool, such as Skitourenguru, may be considered the strongest option for skiing and snowboarding enthusiasts – allowing them to pursue their passion for the winter sports whilst helping them to avoid potential danger.
What triggers an avalanche?
Avalanches can be caused by many things. Some of these are natural. For example, new snow or rain can cause built-up snow to loosen and fall down the side of a mountain. Earthquakes and the movement of animals have also been known to cause avalanches.
On the other end of the spectrum, artificial triggers can also cause the natural phenomenon – from skiers and snowmobiles to gunshots and explosives. Indeed, human-triggered avalanches can start when a person walks or rides over a snow slab with an underlying weak layer. This weaker layer collapses, causing the overlayer mass of snow to crack and start to slide.
However, in recent years, the impacts of global warming have been felt in the mountainous regions, where the rise in temperatures is above average, affecting both glacier landscapes and water resources. Such repercussions of these changes have increased in the frequency and intensity of snow avalanches.
Snow scientists say extreme avalanches are among the accelerating impacts of climate change in mountain regions, and global warming can affect avalanches in several ways. For example, more moisture in a warmer atmosphere can fuel more extreme snowstorms, which means bigger avalanches, and warmer temperatures and make snow layers collapse and slide.
Chris Wilbur, an avalanche consultant based in southwestern Colorado in the US, surveyed 240 avalanche experts and snow scientists in 2018 to learn how global warming is changing the risks associated with dangerous snow slides.
“The findings were consistent, with an increase in observed avalanches,” he concludes, “and even more predicted, especially wet snow and wet slab avalanches because that’s related to warming temperatures.”
Much like in the US, Switzerland has also experienced similar impacts in recent years and is already adapting avalanche mitigation plans based on global warming.
Scientists and engineers in the mountainous country expect more extreme snowstorms, so are therefore building higher avalanche barriers to prevent masses of snow from sliding off the peaks.
Furthermore, Swiss experts are updating avalanche hazard maps as they expect that global warming will put new areas at risk. According to Perry Barthelt of the Swiss Federal Institute for Forest, Snow and Landscape Research, the organisation have detailed satellite measurements of extreme avalanches on 2017-18, which they hope will help assess future risks.
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.