We chat with Dr. Pascal Haegeli about artificial intelligence and snow science. Haegeli is an expert in avalanche safety research working on projects in Canada and worldwide. You can find him teaching and mentoring graduate students at Simon Fraser University.

 

***

 

 

This interview has been lightly edited for clarity and length.

 

The High Route: What exactly does artificial intelligence mean in your snow science community?

Pascal Haegeli: That’s a big question. Over the last two years, at least amidst popular culture, artificial intelligence has really been on the minds of almost everybody in society considering the launch of Chat GTP, etc.

The news around Chat GTP certainly highlighted the power of that technology, or at least large language models, that could change how we work, interact with computers, and affect our daily lives. From a scientific point of view, artificial intelligence didn’t just show up two years ago; it’s been a long development, and there are several different tools that these types of tools can provide us with.

It gives us better tools to look at patterns within large datasets and find complicated relationships between things we observe and things we want to predict.

In the past, relating it directly to snow science, for example, our statistical tools required us to have a very good understanding of the physical processes in the snow to define our prediction models in a way that can represent those prophecies in the best possible way. 

These newer machine learning methods and algorithms can find these relationships more independently. This allows those tools to represent more complicated relationships without us having to apply these models or at least specify these models in a very detailed way in the first place. And that allows us to look at larger datasets more quickly and find patterns more easily.

 

THR: When referring to data sets, can you give us an idea of what that data set looks like? Are the data from several snow pits, for example, on the same slope?

Haegeli: We could think of that data as simple files, for example, where you have detailed records of all the layers, and you’re trying to predict whether there is an unstable layer within this snow profile and where it is located. That’s a fairlycomplicated data set because you have hardness, depth, grain forms, size, and so on. And you’re trying to predict if there is instability within this profile.

So, the predictor we’re looking for is instability—this is a fairly complicated task. We’ve done it routinely with field tests, but if you only have a photo, that would be a much more difficult task. If you want AI predictions based on the profile alone, that is fairly complicated. We can use some of these machine learning algorithms to make that connection. These machine learning algorithms can detect these relationships. That’s an example of a data set like that. 

To develop these models, we, in turn, need to have data sets that allow us to train the model to check what that relationship looks like; we need a data set where we know all the profile characteristics as well as the outcome. The outcome, in this case, is the test results. Once the algorithm has learned that relationship, we can apply it to other data sets where we only have profiles, and we can then use a machine learning algorithm to predict whether this profile is stable or not.

 

THR: I’ve read some of the literature, and there’s a study from the SLF where they’re talking about something like 75% predictive accuracy when using AI and cross-referencing those predictions to what they’re observing in the field. From your experience, I know these aren’t commercially available products, or at least I’m aware of them. What does the accuracy look like? Snow stability is dynamic and depends on many variables. The snowpack or my test results could be different 10 meters from where I’ve just assessed the snowpack. 

Haegeli: In our pre-interview, you said you’re critical about these tools. And that’s a good mindset to have. These tools can only be as accurate as the data set; it all depends on the quality of the datasets that we give it to train and how representative that data set is. An accuracy of 75% doesn’t seem that high if you’re trying to make decisions that your life can depend on. And as you said, there’s so much variability in the snowpack out in the field; it makes a difference whether you ski 10 meters on one side or 10 meters on the other side. 

We need to be very careful about what we use these technologies for. What are the assessments that we let the machine make? And how do humans interact with these assessments? 

So, for example, the real strength of these algorithms is that they can consistently process large amounts of data very quickly. They are less good about knowing the context or whether the current situation is outside of the norm or what we’ve seen in the past. So, for example, a good application of AI is simulating the snowpack’s evolution using a numerical and numerical model across 1000s of locations in Western Canada. Now, forecasters only have time to look at some of these simulated profiles; we can get an AI algorithm that looks at how many are potentially unstable. Where are they? How deep the potential weak layer is, etc. This information gives a forecaster a general sense of the conditions and gives them a meaningful starting point. 

The forecasters then must combine that information with other available observations to validate it and put it through the human filter to provide meaningful information to the forecast users. It’s all about what exactly you want to do with this tool—and understanding its strengths and weaknesses in specific applications. And then, how do you integrate the information into your working process?

 

THR: How do you see avalanche forecast centers using this tool as it evolves and improves? Do you see it as a data supplement when writing a forecast? Or, maybe there’s some gray area here, and forecasts become more automated for regions without the resources or it’s too difficult to access the physical terrain. 

Haegeli: In general, the technology will help avalanche forecast centers become more efficient and, therefore, use the human resources they have more effectively in areas where it makes the biggest difference.

So, it might help to streamline observations and interpretation.

As I mentioned before, lots of forecasting centers are now starting to use snowpack simulation. This can quickly give the forecaster an overview of what’s going on, where they can then start their forecasting processing in a more targeted way. That might free up some time for them to do forecasts in the areas where, in the past, they have not forecast for. The homework that they have to do upfront is more efficient. 

I spent quite a bit of time in Switzerland last winter. The forecast key in Switzerland uses some of the models developed there and includes a model that looks at stability and snowpack simulation; they also have a model that can predict the danger rating based on their assessment.

That would be another application of AI: it offers an additional perspective, and a model forecasters can compare to their assessment. It creates an additional forecast based on the knowledge captured in historical assessments.

 

THR: What did you learn about that gray zone between the model’s predictive capacity and on-the-ground truthing in Switzerland? Did you find the AI models most aligned with what was occurring in the field?

Haegeli: I would say, in general, that the particular model predicting the danger rating is quite good. Forecasters are now learning this model’s strengths and weaknesses, like when it works well and, generally, what situations where it’s off. One of our challenges in avalanche forecasting is that, most of the time, it is relatively straightforward. And the model does an equally good job compared to the forecaster. 

The challenging situations are where a good model can be most useful. Under those circumstances, the model also has its challenges because there are few edge cases in the original data set that was used for training. 

In those instances where the forecaster has to decide what exactly is the condition right now? And how do I best communicate that? It’s critical to understand where the model is coming from, and what kind of biases that model will have. And we’re currently in the process of learning.

 

THR: As far as recreationists, and using an app that promises guidance regarding what terrain to ski, how much faith should we put in such tools—tools that harness AI in some capacity to provide guidance?  

Haegeli: There are two different audiences we’re talking about. So far, we have talked about a professional audience, avalanche forecasters, who can use some of these tools to make their work more efficient. In general, there is a healthy dose of skepticism. But, I think this type of user understands the limitations of these tools, and, and is willing to put in the time to learn about their strengths and weaknesses.

When we’re talking about recreational users, several things come together that make it very challenging. First, there’s the hype about AI. People may not be aware of how much work it takes to develop some of these tools, so they provide meaningful insights. The tools that developers can use are easily accessible these days, and you can easily throw a bunch of data at it and get an output. 

However, whether that output is meaningful or not, is a completely different story. 

One of my PhD students, over the last year, he’s tried to create a model that predicts the viable run lists for a heli-ski operation. For those new to the run list concept, a heli-ski operation has a run list, or the run they will approve (or shut down) to ski on any given day. Let’s say they have 300 runs. Every morning, they do their hazard assessment. Then, they go through their list of runs to decide whether that particular run is open or closed for guidance.

Putting that list together takes a lot of time in the morning. And so having a tool and model that would at least give a guide a first guess that under these conditions in the past, these runs were typically open versus these ones were closed. 

It required a tremendous amount of work for this student to describe the hazard in the best possible way and describe the terrain that is/was actually skied in a meaningful way to give the modeling a chance to come up with a meaningful pattern. The model he came up with is fairly good at determining if a particular run should open if it were open the day before or if the run is closed when it was closed the day before.  

But the transitions, which are what we’re interested in, are actually really hard to predict. This is all to illustrate that it takes a tremendous amount of work to produce models so that they produce meaningful output.

Let’s go back to the basic user. Much of our research has shown that it’s actually challenging for people to take the forecast information and then apply that to terrain and make meaningful decisions about where it’s appropriate to ski and where it’s not. That’s a challenging decision-making and assessment process. But it’s where the rubber hits the road. Thisis a challenging process, and people are looking for shortcuts.

And so, in my opinion, this creates somewhat of a perfect storm; people developing these apps, because they or a developer can write them and they create flashy outputs. And I’m not saying that these outputs are necessarily wrong, but there are serious limitations of what these tools can do. Then, you have users on the other side who are looking for these types of tools and are not fully aware of the limitations of these tools. I think having a very healthy dose of skepticism is critical. Be very cautious about how you use these tools and fully understand that they’re just an additional perspective or provide additional input for your existing decision-making process. Don’t just give a thumbs up and justify that you want to do a trip or a run that otherwise looks a little bit questionable.

 

THR: So as I noted informally in an email or as we picked up the call, I tend to have a healthy dose of skepticism. But, oftentimes, I’m behind the curve and should adopt sooner when it comes to certain technologies. 

Haegeli: I think everybody wants to move toward tools that help users interpret the avalanche forecast and make informed decisions—we’re all trying to work towards that. And I think we have to do it in a very informed and transparent way. I have a master’s student with me starting this fall who will study some of the existing decision support tools in Europe along these lines. There’s a fairly popular website that basically provides a risk assessment for an existing by taking an avalanche forecast and layering it over terrain. It is called skitourenguru.ch. 

In simple ways, it figures in the danger rating and then determines the aspects and elevation bands where that danger rating applies and it basically calculates a risk metric for routes, depending on how that route interacts with that terrain,and how steep the terrain is and so on.

It’s a fairly sophisticated algorithm that does that. And then it spits out a weighting between zero and one. Then, it colors the routes greens, oranges, and reds according to that value. And it is super popular.

Our research aims to better understand how people use the tool and how much they trust those assessments. Does a group, when a week is colored in green, does it just mean go and they don’t use any additional information or don’t do any local assessments? Does this user just follow that track, and that’s it? Or, are they using the site and the data output as a starting point to decide if a route is probably more suitable right now than another route. We’ll also look into the additional information they may use to make those choices. We want to understand more about that dynamic. We want to use that as the foundation for developing our own tools so that we can develop them in a more informed and targeted way. 

 

THR: Where does AI come into play for more and less experienced recreationalists? 

Haegeli: For those who are more experienced, I think it will offer an additional tool that will provide you with additionalperspective. For sophisticated backcountry users, I don’t think that much will change.

I also think we’ll be able to provide tools for less experienced people that will help our team to make better-informed decisions.The way I see it is that some of these tools, because they make forecasters more efficient, they will therefore be able to create forecasts for more regions, or they will be able to update the forecast more frequently. This may create efficiencies that allow us to potentially create more targeted forecast products for different types of users. That will allowus to support a wider range of backcountry users than we have in the past. I think that’s where the biggest gain will be, not necessarily for the sophisticated end user.