World of Science | Frank Techel (SLF) on people and models

In addition to the physical snowpack model SNOWPACK, which uses weather data to calculate the layering of the snowpack at a specific point, three machine learning ("AI") models are used at the SLF. These models have "learnt" statistical correlations between snow cover simulations, weather data and avalanche observations or hazard levels using training data sets. The learnt correlations are used to predict relevant avalanche parameters (probability of triggering, danger level) without the intervention of human forecasters.

PG: In your publications, you write that the trend in Swiss avalanche warning is moving away from a purely "expert-based approach" and towards a "data and model-driven approach". Is that a fundamental goal for you?

FT: I don't know whether a purely data- and model-driven approach is really the goal, but I see a stronger integration of ever-improving models into the forecasting process as a logical development. Until about five years ago, all we really had in avalanche warnings were today's observations, today's measurements and a weather forecast for the following day. We also had the SNOWPACK snowpack model, but it was rarely used in our area. So weather models were the only models that really played a role. All the rest was done by the avalanche wardens using their experience, their knowledge and their gut feeling. That's what I mean by Expert Based Approach. 

We have had a lot more model data available for a few years now. On the one hand, SNOWPACK, which we are increasingly using for forecasting, and on the other, the machine learning models that come on top of SNOWPACK, so to speak. A lot has happened since the last PowderGuide article on this topic was published. At that time, we had already partially tested the model chains, but they are now very stable and in operational use. In addition to all the programming work that was necessary for this, the training of the avalanche wardens is also extremely important. We all have to learn what the models can do and, above all, what they can't do.

Are you also noticing this trend in neighbouring countries, towards models and away from purely human expertise? 

Yes, when I talk to colleagues in Tyrol, or in Norway or Canada, everyone realises that models offer a great opportunity. I'm not talking about AI, but about all models. There is certainly still a lot of implementation work to be done, as well as research, but models have great potential to improve avalanche warning. I am trying to promote this transfer - that the models really do come from research to us in avalanche forecasting. 

What does the "model-driven approach" look like in day-to-day operations?

I don't know if I would call it a model-driven approach yet. But the models already offer a valuable second opinion on how the existing data can be interpreted. So when I make a forecast, I continue to make exactly the same considerations as I did a few years ago, but I also have a model that tells me something about the probability of spontaneous avalanches, for example.

Let's talk in more detail about the models you use. Can you briefly tell us what SNOWPACK does?

SNOWPACK takes weather data and uses it to calculate the snow cover, so to speak. Every weather event has an influence on the snowpack, be it an increase with a new layer of snow, or when meltwater penetrates, or anything else. SNOWPACK then simulates these processes for a specific point;

SNOWPACK is not an AI model, but a physical model that calculates what happens in the snowpack when it rains or snows on it, for example, based on our understanding of the process.

Genau, I. 

You now have several AI models -  including the Danger Level Model, the Instability Model and the Natural Avalanche Model. What do they do?

As I said, we have the very complex simulations from SNOWPACK at hundreds of data points and for different exposures. All in all, that's a lot of data that first has to be made usable and interpretable for the avalanche warning system. With the AI models, we try to filter out the most relevant information. 

All three AI models are trained to answer a very specific question. They can each do one specific thing. The hazard level model, for example, has used data from the last twenty years to learn which combination of weather data and SNOWPACK snow layers correlates approximately with which hazard level.

The instability model is different. It only takes the simulated layer profile and goes through every layer combination. What does the layer look like? Could it be a weak layer and is there a snow slab over it? The model has learnt the correlations on the basis of slide block tests. It then gives a probability that this is a typical combination of a weak layer and the "slab" above it. Based on past data, we then expect a more or less weak slip block result. The most unfavourable layer-board combination in a simulated profile then classifies the snow profile as weak or stable.

The spontaneous avalanche model is based on the instability model and also takes fresh snow as a parameter. It has learnt from a historical data set with observed avalanches. We used the model operationally for the first time last winter and then analysed how the model's predictions correlate with radar-detected avalanches. The radar data comes from certain avalanche tracks where permanently installed sensors recognise when an avalanche is about to occur. Such systems are used, for example, for automated road closures on exposed traffic routes. This is an interesting data set for evaluating the spontaneous avalanche model, because the radar registers the avalanches virtually in real time. Human avalanche observations, on the other hand, are usually somewhat delayed because we don't always notice the avalanches as soon as they start, or because it's snowing and we simply don't see anything. The comparison with the radar detections confirmed our feeling that the model reacts with a slight delay. This is probably due to the fact that the model was trained with human observations, which are also slightly delayed. It is therefore very important for all models that we know how the models are trained, as this also determines potential errors.

The AI-based, statistical models are trained with your forecasts and observations and the snowpack simulations from SNOWPACK. This means that we still need to understand the physical processes and can't just rely on the magic of AI somehow, right?

Exactly. I would very much like to have models that represent the physical processes as well as possible. I never just look at the output of the instability model, I always go back to the SNOWPACK simulation and look at some stations to see whether the layers look plausible. It is extremely important that we at least begin to understand what is happening in these long model chains. The AI models come at the very, very end and are actually just superimposed on everything that happens beforehand.

The AI models help to extract the relevant information. SNOWPACK spits out lots and lots of information and the AI then filters it for us. Can you summarise it like this?

Yes, you can summarise it like that. Ultimately, these are simply models that end up on top of SNOWPACK. But these "small" models have led to quite a big leap in the use of SNOWPACK in our forecasting service. This is because they suddenly make the complex output of the SNOWPACK model digestible by extracting relevant information from it that is relatively easy to understand. I therefore think it would be important if we could integrate the models even more easily when creating the forecast.

How does that work in practice? I say three plus. You say four minus, the model says three plus. Then we form the median. What does the discussion look like then?

In this case, if we include the model, we would start with 3+ as a starting point. Let's assume that I am now very unhappy with giving the situation a three for tomorrow. Then my task would be to explain to the others why I don't think this is right. As data-based as possible, although this is of course difficult because our human interpretation is always involved in the end. In this discussion, decisions are likely to be made one way or the other. And that is also one of our weaknesses, because we humans are not always consistent.

So it's still human?

It's clearly human. We are still primarily human avalanche monitors. However, the models can support us and show us where we perhaps need to take a closer look today. 

Will it stay that way in the future?

I don't know what an avalanche warning will look like in the future. Maybe in ten years' time the forecast will have a high resolution, maybe models will do a large part of the forecasting. Who knows? Maybe that's why it's all the more important that we can still provide explanations and communicate to the user: "Watch out right there..." In other words, that human voice that you bring in. What the avalanche forecast of the future will look like is still very open to me.