Computers Hunting for Avalanches

Hello all! Today we’re going to be talking about avalanches. Avalanches are, as you probably already know, dangerous–they are potentially fatal, and can also cause road closures, force evacuations, and cause economic damage. As such, having a good risk assessment for avalanches is important–and as it turns out, difficult.
For a good risk assessment, one needs a full set of data of avalanches in both time and space. If one is using the traditional methods, which estimate some variables of avalanches from safe distances–this depends on the weather, if it’s daylight, and is time and labor intensive and…well, this isn’t going to yield the data set needed.
Let me introduce today’s article. “A method for automated snow avalanche detection through use of synthetic aperture radar (SAR) imaging” was written by H. Vickers, M. Eckerstorfer, E. Malnes, Y. Larsen, and H. Hindberg. It was published in the American Geophysical Union’s Earth and Space Science in 2016. 
An image of an avalanche in the Mount Cook area of New Zealand in 2013. Image credit NASA Earth observatory.
Here, the authors are trying to create a fully automated (no human input) algorithm so that computers can find avalanches all by themselves, using radar. Radar has some advantages over traditional methods; it doesn’t much care about weather, or lighting conditions, and has been shown to be able to point to avalanches. Radar works by sending some form of signal and having it ‘backscattered’, or reflected back the same direction it came from; depending on items such as the surface it hits and how far away it is, this backscatter varies, and features are discernible. Avalanches appear as long, downhill, tongue-shaped things that weren’t there on radar images taken at a previous time.
The algorithm the authors aimed for has been tried before in a few different way, but always ran into trouble; they were perhaps confused by artificial snow, wind-sculpted snow features, snow-covered vegetation, if the avalanche had been covered by new snow, or were not fully automated. So, the authors decided to try what they hadn’t seen before: a fully automated avalanche-hunting algorithm using SAR (synthetic aperture radar) data. The data is from a spacecraft called Sentinel-1A. The data is available worldwide, but the authors focused on Norway, as that is where they were based. The orbit of Sentinel-1A goes over the same area at the same angle every 12 days, but passes near the poles (i.e. Norway) every 4 or 5 days, with different angles, which can be useful for monitoring.
An image showing two unusual avalanches in Tibet in 2016. The first two are natural colors, and the third is from SAR data, the same kind of data our authors are using. Image credit NASA Earth Observatory.
To test the algorithm, the authors made use of a storm system that had passed through Troms County, Norway just after New Year’s 2015. The storm brought heavy winds and rain and many avalanches resulted. The basic steps of the algorithm:
  1. Compare an image to an image with identical viewing angle (12 days prior). Each image is split into smaller ‘windows’ for efficiency.
  2. Filter out any parts of the image that are in shadow or otherwise cannot have an avalanche detected there due to the angles of the radar data. Also filter out any water that could look like avalanche debris, and slopes higher than 35°, as most avalanches are on slopes less than this.
  3. Is there is enough difference between the backscatter in the same section of the window between the two images,  and at least 1% of the pixels have such backscatter difference, mark it as an avalanche. If not, the computer categorizes it as ‘not avalanche’.
  4. Repeat the above until all sets of images are complete.
To test the algorithm, an avalanche expert looked at the radar images and marked the avalanches. The authors used two areas as case studies, the first being Tamokdalen, Norway. The original, ‘reference’ image was taken 13 December 2014, the next was 6 January 2015, after the storms. Here, the algorithm marked 40 of the 70 avalanches found manually, which is 57%, and 62% of pixels overlapped with both methods. The algorithm generally agrees in size and location. The most disagreement came from smaller avalanches, possibly due to the 35° cutoff used. The algorithm also marked areas as avalanches that were not marked manually. For a test of images from December 16, 2014 (reference) and the 9th of January 2015, the algorithm wasn’t as good, only finding 7 of 53 manually marked avalanches. It did snow between the 6th and the 9th of January, and this may be a factor in this. They also combined images from the 6th, 8th and 9th of January; the algorithm found only 6% more than it did the first time (68%), so this was not an impressive improvement.
The authors then did a case study in the area of Lavangsdalen, Norway. The reference image here was taken the 13th of December 2014, once again compared to the 6th of January 2015. The algorithm performed much the same; detecting 28 of 39 (72%) of avalanches, not quite in the same size or shape as manually marked. It again missed smaller ones an marked avalanches that weren’t seen manually. For one of these, the authors went to the site and marched around the edge of an avalanche with a GPS, comparing this ‘ground truth’ to both the manually marked avalanche and what the computers saw. The manually marked avalanche was nearly identical, and the algorithm mostly agreed–it found 379 pixels, the manual method 365, and 321 pixels were agreed upon as an avalanche.
Obviously, improvements can be made. The authors went through a lot of math to show that the threshold of backscatter difference required by the algorithm to mark something as ‘avalanche’ is decent, though in other places different numbers may be better. Changing this could also increase misidentification, and where the best fit is will depend on things like whether or not the algorithm is being used to decide on evacuations or not. The authors also note that how snow backscatters depends on how wet it is, and other factors; wet snow reference images will be more difficult. The 35° slope cutoff could be a factor, as well as the possibility of avalanche debris on frozen lakes (which would not have been seen here due to the filter set for bodies of water). There is definitely room for improvements, and the authors promised to keep working on it.
Carlowicz, Mike. Avalanche in Aoraki/Mount Cook National Park. NASA Earth Observatory, 15 Feb. 2013. Web.
The source of the first image in today’s post. The original page also includes a second, previous photo for reference.
Vickers, H., M. EckerstorferE. MalnesY.  Larsenand H. Hindberg (2016), A method for automated snow avalanche debris detection through use of synthetic aperture radar (SAR) imagingEarth and Space Science, 3, 446–462,  doi:10.1002/2016EA000168.
Today’s main article.
Voiland, Adam. A Second Massive Ice Avalance in Tibet. NASA Earth Observatory, 21 Oct. 2016. Web.
The source of the second image in today’s post.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s