back to article We trained an AI to predict how bad a forest fire will be. It's just as good as a coin flip!

Forest fires have apparently ravaged over four million acres of land across the United States so far this year, and the problem is only getting worse with global warming. Enter technology's hottest solution: Machine learning. Scientists and engineers from the University of California, Irvine (UCI), have built a decision tree …

  1. Archtech Silver badge

    Fair to middlin'

    Right half the time - that's not so bad. It's similar to the Met Office's weather forecasts for a few days ahead - although they are much more accurate when looking decades ahead.

    Any high-tech AI solution that does as well as tossing a coin shows real signs of promise.

    Although, when you think of it, being right half the time is exactly the worst you can do.

    1. Baldrickk Silver badge

      Re: Fair to middlin'

      When tossing a coin, sure. They mentioned that the standard to beat was 33% in this case.

    2. Cuddles Silver badge

      Re: Fair to middlin'

      "Right half the time - that's not so bad."

      It depends what it's compared with. Pure chance isn't a sensible comparison because no-one has ever tried to predict how bad a fire will be by rolling a die. The much more relevant question is how this compares to a firefighter, disaster management person, or whoever else might generally be tasked with making such predictions.

    3. VikiAi Silver badge

      Re: Fair to middlin'

      Several years back the British Met office was boasting an 80% accuracy on daily forecasts.

      NewScientist editorial pointed out that, since local weather tends to run in 6-8 day cycles, saying "Tomorrow's weather will be much the same as today's" should get a better accuracy-rate than 80%!

  2. David Bond

    Well at 50% it is probably (I guess around 50% chance this is right), as close to guessing with the 3 options just basing it upon what the weather is like. Raining likely to be small, so predict small, but could end up falling into the medium, unlikely to be large. It's hot and dry, likely to be large, predict large but but could fall into medium, unlikely to fall into small. Based on that you have a little under 50% change of being correct, all depending upon the types of fires and weather they used to train and test against. But they did say it was simple.

  3. fajensen Silver badge

    Retarded or Clever use of machine learning, depending on wether the success criteria were 'good predictions' or 'solid money'!

    There are not enough forest fires (yet) to build a dataset that is large enough for the current crop of machine learning to work. In former times, people would have used a mathematical model based on simplified physical processes and statistics. That would have taken longer to develop, been a few hundred lines of math-heavy code, maybe a paper in 'IEEE' or 'Nature' and it would probably be more accurate too. But, not 'Machine Learning' so not wanted!

  4. Dan 55 Silver badge

    More training data needed then

    Aren't we lucky Brasil has some for us?

    1. This post has been deleted by its author

    2. This post has been deleted by a moderator

      1. JetSetJim Silver badge

        Re: More training data needed then

        ML/AI is like curve fitting. If you have 2 data points, it's pointless fitting an order-100 polynomial to it. In contrast, linear fitting *can* work for 1000 data points, but you are possibly over-simplifying your solution too much.

        The data has to be sufficiently modellable with the features offered to the AI (even if that model is unknown). This team are attempting to predict the size of the fire from only two variables (which is slightly at odds with the article statement of "..built a decision tree algorithm to predict how big a forest fire will grow given the time of day, weather conditions, and local vegetation"), which are quite possibly the only two that can be reliably extracted from historic data, but given their accuracy of only 50%, there is clearly more at play in the underlying model (and even then, the chaotic nature of the process may prove limiting)

        Overall their model boils down to:

        (a): if VPD < 0.66 kPa, goto (b) else goto (c)

        (b): if VPD < 0.56 kPa, it's "small", else "medium"

        (c): if spruce fraction < 0.04 it's "small", else "large"

        Interestingly, the abstract of the paper says that they tried more complicated models and they were less accurate, and rather interestingly just using the Vapour Pressure Deficit was 49.2% accurate, and just using Relative Humidity was 47.2% accurate. They probably know a boat-load more about forest fires than I do.

        1. This post has been deleted by a moderator

          1. JetSetJim Silver badge

            Re: More training data needed then

            It's not "my" model/system/data - I'm merely commenting on their methodology. The objective seems to be to map some simple input features that are easily measurable into a more accurate determination of the possible spread of the fire - perhaps something that could be quickly evaluated on detection of a fire to give extra information to emergency services. What they've made is something more accurate than rolling a dice, but my position is that the complexity of the system they're trying to model may be too much to get much more accurate.

            If you think you can do better with your iffy language processing AI patents, go ahead

            On a separate note, it might be interesting to see if finite element analysis techniques offer some benefit to this, but I doubt I'm the first person to think of this! Equally, some AI image processing techniques would be interesting to throw into the mix

            1. This post has been deleted by a moderator

  5. This post has been deleted by a moderator

  6. This post has been deleted by its author

  7. Alister Silver badge

    built a decision tree algorithm

    Let's hope it's fireproof

  8. rnturn

    That spokesperson's hands must ache from all the hand waving that he was doing to justify the barely so-so results their AI achieved.

  9. NantucketClipper

    Good thing scientists don't use computer models to predict climate change

    Oops. I just Googled "climate change computer models", and found out that that's exactly what the scientists do. No worries, though. We IT people have 100% confidence that those computer programs predicting the Earth's doom in 10 years (or was it 20? 50? -10? -20?) are 100% proven and reliable, not to mention the rock-solid data they use. Woooh... (sigh of relief)

  10. md56

    That's highly significant

    Reminds me of a model (not AI, pure math) I built many years ago to predict binary choice in random walking parameter values around (what would now be called a sort of evolutionary fit) to find the best fit. N around 100,000. After about a month of trundling around (and a power outage...) it delivered an amazing answer -- on then-common frequentist statistics, the F values were >>1E6, wow, incredibly significant. But then I looked at the proportion of correct predictions -- around 50.2%... My current answer: Don't use frequentist statistics!

  11. s.coffield

    Hi all, lead author here. Thanks for the discussion. I'm more of an ecologist than a machine learning specialist (though we do have a couple computer scientists and statisticians as co-authors), so it's interesting to see your perceptions

    Yes, basically we turned a 1-in-3-odds problem into the odds of a coin flip, on average. What failed to be included in the article from our email interview was that we still find this encouraging and worthy of sharing for a couple reasons. The performance is best for the largest fire group, which cause about 90% of the destruction. The model can "catch" 65% of the fires that do become large (recall), with a precision of 53%. This allows us to do good enough to identify a small subset of fires that will account for a majority of the burned area (for example, 40% of fires accounting for almost 80% of the burned area, etc by Figure 7). That could absolutely be useful information for triaging efforts. In some of these cases, dozens of fires are breaking out in a day or two, and managers need to make decisions faster than is possible by running full fire spread models on every ignition.

    We certainly would encourage future research to use more complex input information to increase this accuracy (our learning curve analysis suggests that the number of fires in the dataset isn't what's limiting accuracy), in order to increase its usability for managers on the ground. In essence, we're setting up a new framework that is especially relevant for areas where fire frequency is increasing due to climate change, and hopefully others can build off it.

    Fire prediction is inherently a really difficult and chaotic problem, so it's interesting that just a couple simple variables can explain even this much of the variance from the time of ignition, especially when many of these big fires burn for weeks or months.

    TL;DR: Far from perfect, but useful (and scientifically interesting) nonetheless

  12. Douglas Wardle

    You could have just asked the donald

    Your forests simply need a good raking, and harvest for sale the trees that died from rot. And, as he has said, if the Canadians didn't have so many trees there would be more demand for real 'Murican produce. I'm surprised scientists need to investigate this when they could be kept occupied denying climate change. As volunteers, of course.

    1. Carpet Deal 'em Bronze badge

      Re: You could have just asked the donald

      Your forests simply need a good raking

      Funny you should say that: some of the worst wild fires can be traced to refusing to let smaller fires burn long enough to clear out the undergrowth, letting fuel build up until the inevitable happened. So yes, a good raking would actually help.

      (Not to miss your sarcasm, of course)

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019