3
$\begingroup$

I'm doing my thesis on avalanche prediction using machine learning.

For my input features I'm using avalanche accidents with features like slope, altitude, facing direction of the slope, combined with according weather data from the day the avalanche occurred.

I want to predict an avalanche when certain variables combine and create a deadly avalanche situation. So 1: the avalanche occurs. 0: The avalanche does not occur.

The only data in my database are occurred avalanches, I got around 200 samples. So I don't have any data of a non deadly avalanche situation, which is mostly the case.

My question is if a One Class SVM is a good approach to take on this classification?

Improve this question
$\endgroup$

3 Answers 3

Reset to default
4
$\begingroup$

Your problem seems to belong to novelty detection in the general area of OCC problems.

So, the short answer is: yes. You can apply SVDD (Support Vector Data Description) method to get the smallest hypersphere containing samples in the dataset and then assess whether a new observation is an outlier or not.

Of course, the less representative your dataset is, the less accurate your classifier will be.

Improve this answer
$\endgroup$
2
$\begingroup$

You can use a method of data mining to predict avalanches, however, there are some pit falls which I can provide you based on my basic avalanche knowledge from mountaineering.

  1. What do you want to predict? Spontanous avalanches (mainly threating villages and roads) or human triggered avalanches (mainly affecting skiers). The factors for these are completely different
  2. Getting data has already been mentioned. There are some data sets related to avalanche incidents, for example at the swiss avlanche research institute: https://www.slf.ch/de/lawinen/unfaelle-und-schadenlawinen/alle-gemeldeten-lawinenunfaelle-aktuell.html However, there is naturally little data about cases where no avalanche was triggered and where an avalanche was triggered but nobody harmed and therefore not reported. There have been some tries to estimate the number of people on tour based on touring reports in the internet.
  3. Getting precise data is even more of a problem. Consider figure 2 in this weeks report: https://www.slf.ch/de/lawinenbulletin-und-schneesituation/wochen-und-winterberichte/201819/wob-18-25-april.html It compares the same slope at a time difference of 45 minutes and it looks completely different.
  4. Feature selection is another big issue. You mention that you want to use weather data from the day of the incident. I think this is drawing the wrong conclusions as most skiing avalanches happen during the weekends and probably in slightly better weather. Also most people will be sensible and not go ski touring on risky tours on risky days. This has a big potential to skew your data and your model
Improve this answer
$\endgroup$
2
$\begingroup$

Could you look for any possible way to get non-avalanche data?

1) Avalanches happened in a mountains chain. Could you add to your data neighboring peaks data from the same day avalanche happened?

2) You may have good insights from data exploration. For instance, what is the minimal slope that mountain should have to be able to produce avalanches? Range of temperatures?

3) Could you look for other data sets (with non-avalanches entries) that you could combine with your data?

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ it is incorrect that there is no way to learn a classification model if the training data does not contain at least two classes. The field which addresses this is known as "one-class classification" (or anomaly detection, or outlier detection, it probably also has other names). In this field the broad approach is to build a model of the data and define a distance measure and threshold. New data points which are closer than the threshold are labelled as the training data and those which are further are labelled as the other class. Hope this helps. $\endgroup$
    – cfogelberg
    Apr 26, 2019 at 19:36
  • $\begingroup$ Thanks for the comment and explanation. I didn't know that the outlier detection can also be treated as classification. It is a very interesting point of view. I have edited reply to not mislead others. $\endgroup$
    – Tatyana
    Apr 26, 2019 at 21:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.