Thresholds for test fires?

For the prize (click “show” below to see Proposed Prize Design), we are envisioning that fires should not be suppressed by teams until they reach a certain threshold—this is to ensure that the autonomous systems being developed don’t inadvertently drop water on people BBQing or having campfires.

Proposed Prize:
[spoiler]We are interested in focusing a prize specifically on wildfire suppression. In light of the increasing risk to lives and assets, the focus is on suppression of Wildland-Urban Interface fires, before these escalate into large fire events that put communities at risk. Here is an initial description of a proposed prize. In the other topics in the community, you’ll find different categories of feedback we would love to get on this design.

We understand that a prize designed in this way may result in a different approach or paradigm for firefighting, and we invite you to explore—for example, with faster detection and response, can fires be extinguished without traditional containment and control strategies?

The Proposed Prize: We are proposing a prize that works something like the following: Teams are invited to test a fully integrated autonomous system that rapidly detects, responds and suppresses wildfires. There will be a competition testing area of X by X acres (NOTE: we are proposing testing in an outdoor environment). At the beginning of the test, competition officials will ignite a fire somewhere within the testing grid. Once Y threshold of spread, temperature, flame height, or another variable is crossed, the competing team will have Z minutes to detect and completely extinguish the fire. The overall cost of the teams’ system must be no more than C dollars. Solutions must not pose an risk to the environment and/or lives.[/spoiler]

What would an appropriate threshold be? We’ve had it suggested we should be looking at infrared heat signatures—would that be a good way to think of a threshold? Or would fire behavior (flame length, rate of spread), be better measures for a threshold worth attacking? If so, what is the right cut-off number and why?

Is there another way altogether to ensure that people making small campfires are not targeted by the systems developed in this competition?

Share any thoughts, examples, ideas, or comments you may have!

When thinking of the threshold, please have in mind minimizing the risk to lives and assets. What indicates best that a fire had transitioned into a large fire event?

Also we mentioned BBQing and campfires above, but another category of “good” fire we might want to threshold out would be prescribed burns or agricultural fires. How can we make sure our autonomous system is sensitive (and fast) enough to put out potential megafires before they spread without attacking these fires that don’t warrant a response?

@DanSelz Leaving a fire to burn is often a complex management decision - typically fires near more urban areas, where resources may be based, would be extinguished immediately due to their proximity unless intentionally set. At this case they usually look at the growth of the fire in terms of acerage as a measure. It isn’t the best measure, but probably the one most commonly used today. For instance, once a fire reaches more than 5 to 10 acres it is now much harder to manage. It is not all burning that area at once, but how much has burned, which would mean the Fireline, which is burning, is the circumference around that still burning.

A more factual measure is the fire intensity, or the heat-release rate per unit length of Fireline, however this is very difficult to measure. A high fire intensity in one space doesn’t indicate things are all lost if it’s over a small area as well, so there is a complication to how these limits are defined.

I feel like something missing from this conversation is input from more people who actually work as fire managers. They may have some specific situations that are problematic where a more specific solution could be looked at.

Hi @mgollner , thanks so much for this comment and all your input in this community. A couple questions on your input:

Re: acreage …another way of thinking about this threshold questions, is we’re trying to figure out 2 thresholds, the first being “what signifies that a fire is significant enough to target” (and not a camp fire say) and the second being “what is the maximum threshold we can tolerate” before the fire gets out of control? "
I’m assuming the 5-10 acres you refer is basically the second threshold (what is the biggest possible fire we can tolerate before it’s out of control). Any thoughts on what the first type of threshold might be (i.e. what is the size at which a supression system should bother targeting and trying to suppress in the first place?) We’ve heard some experts talk about fire perimeter of maybe a couple meters square as being a good initial threshold?

Re: more people who work as fire managers, completely agree and actively working to find these people and get them into the community! If you (or anyone else reading this) know of any working fire managers or incident commanders, please let us know (feel free to reach out to me at my xprize email if you prefer), and we’d love to invite them in!

@DanSelz I think it may be more about what the surrounding conditions are of a fire to “target”. In other words, if there is a danger of the fire spreading then any fire you would hope you could extinguish immediately upon detection. What is cut out is “false alarms”. For satellite detection, this includes things like agricultural burning, reflections, etc. Probably you need at least a few square meters to get a detection. Time from initiation of the fire may be something you want to think about.

Hi @mgollner , thanks so much for the prompt response! Agreed with you that cutting out false alarms seems really important. That’s the rationale behind our thinking with the “Threshold to Respond” metric we were talking about; how do we make sure that we’re not targeting “false positives”–events that might seem like incipient wildfires but are not. Our thinking was that hopefully there was some relevant size metric we could use as a threshold (i…e a fire with a perimeter bigger than X sq feet is NOT a false alarm and therefore should be suppressed), but agricultural fires are a good example of a false alarm that might not be weeded out with a threshold size.

Maybe it would be helpful as a starting point to think through all the false positives we would want to rule out, specifically in the WUI. So far a list might look something like:

  • “Recreational” fires (campfires, BBQs etc.)
  • Agricultural fires
  • Reflections
  • Oil or gas fires

Anything else I’m missing that might make such a list of false positives to avoid? Are there any common thresholds or metrics we could use as a way of making sure teams’ solutions aren’t responding

Also, to your point about “time from initiation”…I’m assuming you mean the time teams have to suppress the fire once it’s ignited? We just started a discussion topic here to discuss that exact thing.

Hi @mgollner , thanks so much for the prompt response. If I’m understanding you correctly, I think that what you’re getting is essentially how do ensure that teams’ systems are sensitive enough to not respond to false alarms or false positives (which we gather is a significant and ongoing issue in efforts to detect wildfires).

Above I talked about some of the examples of false positives that should NOT be attacked by the teams’ solutions–specifically campfires and barbecues. I think you’re totally right that things like agricultural or prescribed burns as well as reflections also fall into this category.

What’s challenging here is that in our prize design, we had been hoping to address false positives like campfires or barbecues with a size threshold (i.e. teams should only respond to fires once they get to a certain size), but the examples you point out, indicate that maybe having a threshold size will not solve this problem of false positives.

Maybe a helpful step would be listing all the different sorts of false positives we might want to exclude and then seeing if there are any shared characteristics amount them (physical or fire behavior) we could use as a threshold for teams to not attack. So far I think we’ve identified the following false positives, but please let us know if there are more:

-Agricultural burns
-Other prescribed fires (e.g. for forest thinning to prevent wildfires)
-Oil and gas fires

  • Recreational fires (e.g. campfires or bbqs)
  • -Reflections

Anything else?

Is it possible to exclude certain ‘false positives’ using a multi-variate formula? For example (concerning target indicators), it is not simply the size (a minimum of sq. meters, or, maximum of acres) but also the intensity (heat output PLUS size/area) of the fire…PLUS additional factors such as the rate of fire spread/growth, and, weather/environmental variables (humidity, drought/dryness, wind, wind direction, etc.)…Obviously, each of these is yet to be determined (TBD).

A decision-making algorithm might include a formula like:

Fa (fire area) + Fi (fire intensity) + Frg (fire rate of growth) + Wd, h, w (weather; a composite variable of several key conditions; here; dryness, humidity, wind) = Sy/n/c (suppression effort: yes/no/continue monitoring).

Note: lower case letters are subscripts for the various (capitalized) fire and weather parameters.

@marz62 indeed, increasingly we hear that the key distinction is stationary versus not. Many false positives, i.e, a factory, solar panels, water glint can be mapped out, but then we have the issue of the unpredictable hot spots, i.e., a campfire. The movement seems to be key and as such, fire behavior; regarding fire behavior - do you happen to be familiar with any advancements in real-time measuring of these?

@Eti - Ah yes, the ‘X-factor’ (unpredictable hotspot) is always present in some form, it seems. Somewhere in my forensic science files I have some data on fire spread pattern analysis that I think might be applicable (with modification, of course) to wildfire scenarios…I will need some time to tack this down. Will try to follow up in a reasonable amount of time.

Follow up:

1] I must add another variable to the hypothetical formula: ‘fuel type’ (type of timber, dead/living wood, biomass on ground, mixed fuel [i.e., shrubs, trees {previously burned?}, fallen trees, dead wood, etc.])…which will have a direct impact/influence on a fire’s ‘rate of growth’.

2] Am still reading over my fire forensics files/papers (will follow up again).

@Eti @marz62 - there are a series of fire detection algorithms already used for the active fire mapping program. Active Fire Mapping Program. The newest addition (not necessarily online yet) is using GOES 16 to detect fires about every 15 minutes. the resolution isn’t as good as VIIRS but they do a lot of work to remove false positives. There is still an element of human thought to segregating these fires but most of the initial algorithms are published in many articles they use.

@marz62 Thank you for looking into these. Indeed these are relevant variables, and I believe we’ve discussed them before in this forum - I rather meant if you know of players out there who managed to measure these in realtime (a very challenging task, but could help inform resource mobilization); Nonetheless, any insights from your fire forensics are interesting - may I ask, is your background in fire investigations?

@mgollner Thank you for your helpful comment. We had the pleasure to talk about some of these tools and programs with professionals who use them and/or contribute to their development. Very helpful to get your additional insights; I will make sure to look into the link you’ve shared.

@Eti - Howdy…to answer your question: No. My background is varied (owing to my broad interests) but is most recently (since 2010) in LAW (public defense: mitigation research, forensic sciences) and science journalism. During my most recent 4 year contract as a law researcher, I compiled an extensive database of research papers, organized by topic/discipline (including fire forensics). The topic of wildfires also overlaps with some environmental/climate change issues which I sometimes reported on through the science blog (from 2010 - 2014) as a core writer/reporter.

@mgollner @Eti - thanks for noting this (re: use of algorithms; it seemed like a good/logical way to approach fire prediction, but I was not personally aware of these existing algorithms, and so, I just came up with a basic “Drake-type” equation as an illustration). I will check out that link and review the content – perhaps I can glean some key (but over-looked) information.

@mgollner - I understand that fire prediction can be as much ‘art’ as ‘science’… Interestingly, so too is much of Forensic Science (the topic of much critical discussion/debate in recent years among the nation’s scientific community)… Human interpretation (often based upon experience rather than data) plays an out-sized role in forensic determinations (e.g., whether one pattern from source A matches the pattern on source B, etc.)…this fact of subjectivity in Forensics (as applied to the law) has been criticized by the National Academy of Sciences (2009), PCAST (2016) and NIST (2017).

As it turns out, ‘fire forensics’ is predominantly focused on determining cause (which makes sense) and not predicting its spread (i.e., its pattern of spread, post fire extinguishing)…In point of fact, most house or building fires, once extinguished, readily reveal their spreading patterns through straightforward analysis (often an aerial photo clearly reveals where the fire originated).

So, while I am not giving up on this research path entirely, I feel it (fire forensics) has only limited applicability for our challenge design.

I would only note at this point that any wildfire prediction method must also avoid false negatives (type II errors), which would be worse that acting on a false positive. But, I imagine that false negatives are rare (?) given that there are eyes on the ground (manned fire towers, etc.) to confirm satellite, drone, or airplane sightings.

One other note: regarding existing fire prediction algorithms…how accurate are these predictive algorithms; have they been evaluated against real world data? I am assuming so, but i would like to see the research myself.

Forensics in fire is MUCH more of an art than a science than fire behavior prediction (in my opinion). There is a lot of work to be done in that field and re-creating events by only the final outcome is much more difficult. In fire behavior prediction you potentially have some data over time which you can work with, I.e. changing weather which can be applied to known theorems, even if they are empirical and rough, and generally do well at predicting when a fire will spread faster, slower, in this direction or not, etc. Transition points, such as crowning or not, are harder to predict and generally throw off predictions more than other variations.

There are many “small” false positives from satellites. Single pixels won’t necessarily dispatch a fire department, rarely do people watch the satellite feed until they have a problem. This thought process is changing but I don’t know what will be implemented. There’s a variety of false positives that are real events that just don’t really mean they’re a wildfire. Many papers by Giglio et al. and Schroder et al. from Univ. Maryland and NOAA on this.

There is a good comment in the technical documentation for FARSITE talking about “model validation”. I’d agree with his assessment it’s not entirely possible for large scale fires quite yet. There’s work by Rod Linn using a full 3D solver, FIRETEC recreating fires run at Eglin AFB as well, these were so incredibly sensitive to the wind sensors that you see it’s very hard to do any true validation scheme at this time. But these algorithms are relatively good when “gaming”, I.e. looking at one action vs. another and seeing how the outcome will change, even if the absolute value may be off. in other words, the trends are generally correct. Using real-time data, like data assimilation, is an approach we’ve suggested would be very beneficial for the field, that’s in large part to how weather predictions have improved so much.

@mgollner This is very interesting, thank you for sharing, we had a chance to speak with professionals in the field (recreating fire events) and it’s fascinating. We also heard about major improvements in weather observations/predictions in the last years, and how much social media added to that. Speaking of designing fire events - we now began looking into designing our test-fire, a question was just posted: Design Fire? — XPRIZE Community