Show HN: Open Avalanche Project – Using ML to Improve Avalanche Forecasting | Hacker News Reader

Show HN: Open Avalanche Project – Using ML to Improve Avalanche Forecasting

openavalancheproject.org

42 pointsscottcha10 comments

Hi, this is a side project I've been working on for the better part of the last year. Unfortunately we've also had 5 avalanche deaths in Washington (my home state) in the last week alone. Preventing these deaths is my primary goal.

I originally came up with this idea when I was working on another project to help the Ski Guides of Hokkaido Japan share snow and avalanche information as there is no avalanche forecasts there.

In the US (as opposed to Canada and Europe) our avalanche forecasting is generally only lightly funded. I wanted to find a way to increase the capability of our forecasting while also building a platform which allowed us to improve the forecast products themselves. The Open Avalanche Project is meant to both be a data and experimentation platform for this. Operationalizing the data and using ML to predict the aspects of an avalanche forecast seemed like a good way to begin to move towards that goal.

This goes beyond what existing avalanche forecasts do in that it has a gridded resolution as opposed to a regional estimate. It also has been built to operate at scale as I would like to cover every avalanche hazard area in the world with a forecast.

I'd love to answer any questions folks have about the approach taken or the goals of the project.

This is super cool! Note that even in California there was an avalanche in-bounds at a ski resort last week. The problem is real.

https://www.mercurynews.com/2018/03/03/rescuer-describes-har...

It might be worth reaching out to Joel at Open Snow (http://opensnow.com/about/contact) to see if they'd be willing to share any of their historical weather data (assuming it's what you need, I don't know the ins and outs of weather data other than being an Open Snow user and knowing they have historical records for snowfall for much of the US).

Open Snow pulls most of their SWE data from SNOTEL sites and NRCS reports.

As someone who has backcountry skied in Washington for the past 15 years color me interested!

What are the weather inputs into the models and do you think that will capture enough? Wind transportation seems easy enough to think about but stuff like hoar frost seems harder to predict using available weather data, but maybe that's just clear cold days?

I do wonder as well whether the models should actually be regionally specific. Sure avy patterns are different but that likely falls out if you are basing it all on weather. IE, WA can act like CO when we have many weeks in a row of cold and vice versa during warm spells there.

Anyways, will be keeping an eye on this. Enabling places with no current forecast to have something at all is a noble goal, though I am skeptical that the recent WA deaths would have been prevented, that's just education and group dynamics sadly.

The current model captures data from three separate inputs: The NAM weather forecast model--primarily wind, air temperature and precipitation values SNOTEL data--on the ground snow values include snowpack depth and SWE

SNODAS-which has things like snowpack temperature and sublimation rates

The input combines these and has a lookback of ~21 days to allow for things like buried hoar frost. I do think there is work to be done to validate that buried weak layers are well represented in the model but currently the model is a pretty good representation of the human forecasts which do account for those.

I do agree that regional variation is a factor. One example is that I feature engineered a Long Term Cold feature which we know is a primary cause of snowpack faceting leading to buried deep layers. This is more of a continental feature of avalanches as opposed to a coastal. My model doesn't consider this feature important. I currently attribute this to being only trained on costal forecasts. I have reached out to both to the Colorado Avalanche and Utah avalanche centers to work on how to best incorporate their historical forecast data in to the modeling to assess this further.

This is cool! I have been looking into building something similar along these lines for a while (I'm a data scientist and avid backcountry skier in the Wasatch). Have you seen: https://utahavalanchecenter.org/ ?

I would potentially be interested in contributing to this. One of the approaches I have thought would be beneficial is to generate probabilities of a slide occuring on a given slope using snow profiles as features in the model. Here we have the Utah Avalanche Center who go visit slides and generate snowpack reports like this: https://utahavalanchecenter.org/avalanches/37960

There are also a bunch of other snow pits dug with reports on slopes where a slide didn't occur. I would be neat to be able to plug in a slope (aspect, pitch, elevation, wind loading, etc) and have a red/yellow/green generated. The problem I ran into was obtaining the data - it all looks to be stored in random places.

Great stuff! Super interested for a European version. Its been close to my heart for a long time. I'd like to see a change in the risk scale. 1-5 just doesn't work. If its 4-5 you don't ski. 1 is basically bomber. 2 is where all the safe skiing happens 3 is where all the deaths occur.

Seems like 3 could be a whole system of grading or simply 3- or 3+. I know that there is huge risk with adding that indicator as people then take it for 'gospel' but informed is better than ignorance.

Browsed the repo a bit. From where do you get your surface-level forecast data?

The surface-level forecast data is the NAM model which I pull from NOAA. Eventually I want to try the GFS (with a resolution tradeoff) or the EU models (with a cost tradeoff).

Reply on news.ycombinator.com