Judges’ Queries and Presenter’s Replies

  • Icon for: Diane Cook

    Diane Cook

    Judge
    May 20, 2013 | 12:48 p.m.

    How early in the avalanche can it be detected? Early detection would be important for timely reaction.

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 20, 2013 | 01:17 p.m.

    At most, the avalanche is detected within five minutes of its occurrence. The detection workflow processes five minute chunks of data at a time and only pulls out the significantly powerful events using spectral flux based thresholding. This event selection acts as a filter, removing the small, insignificant non-events before classification.

    To speed things up, the event selection could be performed in an online fashion (as the data is being received)- at which point the avalanche could be detected within the first five seconds!!

    I hope that answers your question!!

  • May 21, 2013 | 03:47 p.m.

    Please give a few more details about how your compressive sampling method works. It sounds as if it might be very useful in all kinds of remote-sensing settings.

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 21, 2013 | 07:40 p.m.

    Compressive sampling (or sensing) has been growing ever since its formulation in the mid 2000’s. Instead of trying to explain how compressive sampling (or sensing) works in this small area (and without mathematical symbols), I will instead direct you to three publications of increasing complexity. The first is from Wired magazine…

    1) “Fill in the Blanks: Using Math to Turn Low-Res Data Sets Into Hi-Res Samples”
    http://www.wired.com/magazine/2010/02/ff_algori...

    2) “Compressed Sensing: Making Every Pixel Count”
    http://www.ams.org/samplings/math-history/hap7-...

    3) “An Introduction to Compressive Sampling”
    http://dsp.rice.edu/sites/dsp.rice.edu/files/cs...

    I’m sorry for my lack luster answer, but it’s simply too hard for me to convey how compressive sampling works without mathematical symbols, a white board, or even just hand gestures..

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 21, 2013 | 07:50 p.m.

    Compressive sampling has been used in all sorts of domains, including MRI data acquisition and reconstruction, geophysical inversions, and (so-called) analog to information conversion. A team at Rice University even created a single pixel camera!!!

    Rice University maintains a terrific list of resources and publications related to compressive sampling. Here’s a link: http://dsp.rice.edu/cs

  • May 21, 2013 | 05:01 p.m.

    how does your performance at avalanche detection compare to previous work?

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 21, 2013 | 07:48 p.m.

    There are two sets of previous research that my work improves upon. The first comes from France and the second from Iceland. In the French version, researchers were able to reach >90% accuracy in classifying avalanches, but the classification scheme was NOT TRAINED AUTOMATICALLY! Instead, the French researchers had human experts create fuzzy logic rules to discern, for example, an avalanche from thunder. My research improves upon this work by making the training completely automatic; the workflow does not require any human judgement other than identifying several avalanches in the initial training set. In the Iceland version, researchers automatically trained machine learning models to identify avalanches, but only reached 65% accuracy. My work, on the other hand, reaches >90% accuracy with fully automated training.

  • May 21, 2013 | 07:59 p.m.

    Typically sensor nodes forward their measurements to other sensor nodes until the measurements reach the sink. Alternatively, sensors may send the data directly to the sink, creating a star topology? Please describe the wireless sensor topology that you are using, the restrictions on distances that you need to observe during deployment, and the expected frequency of replacing the sensor nodes when they run out of power.

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 21, 2013 | 08:12 p.m.

    We use a simple star topology with long-range radios (802.15.4 XBee Pro with one mile LOS range). Additionally, we sleep the radio and employ a double buffering scheme to reduce the radio’s uptime.

    In terms of power consumption, the sensor nodes must be designed to last all winter without battery replacement. This is because the sensor nodes have and will be deployed in potentially dangerous (avalanche-prone) areas that are difficult to reach. Thus, in our Swiss installment we used a 10W solar panel and 30 Ah battery. (There is the added challenge of poor battery performance at cold temperatures to boot).

    Because of the added cost per sensor node (i.e., solar panel + charge controller + larger battery), decreasing the sensor node’s power consumption (using compression, for example) is of paramount importance. Additionally, this is why compressive sampling is so attractive- we can place hard guarantees on the amount of compression (and thus power consumption) that will occur.

  • Further posting is closed as the competition has ended.

Presentation Discussion

  • Small default profile

    Becky

    Guest
    May 21, 2013 | 05:47 p.m.

    Skiers thank you. Keep up the good work.

  • Small default profile

    irv kauffman

    Guest
    May 22, 2013 | 12:14 a.m.

    Very interesting information

  • Icon for: Joni Falk

    Joni Falk

    Faculty
    May 22, 2013 | 08:59 a.m.

    Thought this video was great at showing the practical implications of your research. Really enjoyed it, thanks!
    Is 93% accuracy an advance on what other methods have been able to detect in the past?

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 22, 2013 | 09:51 a.m.

    Thank you!! I had a great time putting this video together..

    To answer your question, there are two sets of previous research that my work improves upon. The first comes from France and the second from Iceland. In the French version, researchers were able to reach >90% accuracy in classifying avalanches, but the classification scheme was NOT TRAINED AUTOMATICALLY! Instead, the French researchers had human experts create fuzzy logic rules to discern, for example, an avalanche from thunder. My research improves upon this work by making the training completely automatic; the workflow does not require any human judgement other than identifying several avalanches in the initial training set. In the Iceland version, researchers automatically trained machine learning models to identify avalanches, but only reached 65% accuracy. My work, on the other hand, reaches >90% accuracy with fully automated training.

  • Small default profile

    Rebecca Buccini

    Guest
    May 22, 2013 | 11:50 a.m.

    Skiers will love this info. Keep up the good work.

  • Small default profile

    Vlad

    Guest
    May 23, 2013 | 12:10 a.m.

    How does the topology of the wireless network and geography of the area affect collected data and classification precision of avalanches? :P

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 23, 2013 | 12:28 a.m.

    Geography affects the effective distance and line of sight direction of the wireless transmissions. The geology alters the seismic waveforms and velocities as well; e.g., the seismic signal from the avalanche will travel faster in granite than in sedimentary rock..

  • Small default profile

    Ann Brazzel

    Guest
    May 23, 2013 | 01:33 a.m.

    Very Interesting. Thank You for all your hard work.

  • Small default profile

    Ricardo Barrera

    Guest
    May 25, 2013 | 09:44 a.m.

    Are there many avalanches that go undetected? If so, are they too remote to matter to any human interest or are they still important to your research and the success of your technology?

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 26, 2013 | 12:12 p.m.

    Few avalanches go undetected.. Even if the avalanche is “remote”, i.e., far away from a major highway or recreational area, then it’s still useful information. In avalanche forecasting, knowing when and where avalanches are occurring is really useful information. Often times an avalanche will run on an “indicator path” before sliding on something with higher risk..

  • Small default profile

    Ryan Leard

    Guest
    May 29, 2013 | 09:21 a.m.

    Awesome work. This certainly seems to be valuable work for transportation authorities in all the mountain passes prone to slides. This could even be applied to ski resorts that have known slide paths as a way to instantly alert the patrol teams.
    – I wonder if your system could also be used to located, or at least guide, patrollers to buried skiers and riders….

    Cheers.

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 29, 2013 | 10:33 a.m.

    That’s an interesting proposition.. i.e., trying to identify riders in the seismic data. Another more practical way would be to use infrared beam sensors at the top and bottom of a known slide path, and track how many people visit a particular ski run. (We’ve built two different prototypes of a system called SkinTrack that was designed to count the number of people and beacons going through backcountry access gates..). Perhaps we could combine our technologies..?

    Thanks for the question!

  • Icon for: Marc Rubin

    Marc Rubin

    Presenter
    May 29, 2013 | 10:35 a.m.

    Here’s an article about “SkinTrack”: the backcountry access gate monitoring system that we have been working on: http://arc.lib.montana.edu/snow-science/objects...

  • Further posting is closed as the competition has ended.

Icon for: Marc Rubin

MARC RUBIN

Colorado School of Mines
Years in Grad School: 3
Judges’
Choice

Wireless Sensor Network Technology for Avalanche Monitoring: From Hardware to Detection

In this poster, I talk about three research challenges of building a wireless sensor network that can automatically monitor and detect avalanches. My research focuses on three aspects of this wireless system: 1) high-precision and easy to use wireless hardware, 2) efficient wireless communication, and 3) automated avalanche detection. First, I describe efforts to create a “shield” that easily connects existing wireless hardware with geophysical sensors like seismic geophones, self-potential electrodes, and infrasound microphones. Second, I talk about wireless data compression and how a new technique, known as compressive sampling, can significantly reduce radio transmissions and save power. Third, I discuss how data mining and machine learning algorithms have been used to automatically detect snow avalanche events in real-world seismic data.