How AI is learning to predict wildfires

The Manila Times

How AI is learning to predict wildfires  A European climate research institute is using artificial intelligence (AI) to more accurately predict the probability of wildfires across the globe. The European Centre for Medium-Range Weather Forecasts (ECMWF) has created a new Probability of Fire model, which uses machine-learning methods to analyze a wider set of data for more precise predictions of the wildfire locations. By incorporating fuel, such as vegetation levels that could help the blaze to spread, and sources of fire ignition into the analysis alongside predictions of weather conditions, ECMWF can go further than predicting just fire danger, and look at the likelihood of a fire igniting in a particular location.  ECMWF/REUTERS  Subscribe to The Manila Times Channel - https://tmt.ph/YTSubscribe    Visit our website at https://www.manilatimes.net    Follow us:  Facebook - https://tmt.ph/facebook  Instagram - https://tmt.ph/instagram  Twitter - https://tmt.ph/twitter  DailyMotion - https://tmt.ph/dailymotion    Subscribe to our Digital Edition - https://tmt.ph/digital    Check out our Podcasts:  Spotify - https://tmt.ph/spotify  Apple Podcasts - https://tmt.ph/applepodcasts  Amazon Music - https://tmt.ph/amazonmusic  Deezer: https://tmt.ph/deezer  Tune In: https://tmt.ph/tunein   #TheManilaTimes #tmtnews #wildfire

Transcript

00:00This video is made possible in the description of the video.

00:30So here we present a new method to forecast the fire, danger. It uses machine learning

00:46method to combine a lot of information, not only weather but also fuel in terms of availability

00:55and dryness and most importantly source of ignitions. So where people live, road access

01:04to places. This is a substantial advancement on what was done before because previous method

01:12only used weather. So you see, I mean it's quite, it looks quite, much better than the

01:28classical FWI. And it's interesting that we've got the high risk actually of the desert regions

01:35where we don't have any fuel at all. And actually the real risk that we see in the machine learning

01:40model is in a different location. It's where actually a lot of fuel exists but there's

01:44also those weather conditions required for the fire. This is clearly not an information

01:51that is clearly not available for the standard fire weather index. Yeah, because you can see

01:56on the circles and triangles where we actually observe the fires. It's very much more in line

02:02with this model. This is really good prediction.

02:18So historically for fire forecasting we use what's called the fire weather index. And this

02:23is a simple physics based model where we use four weather variables, temperature, wind, precipitation

02:29and humidity to forecast the chance that if a fire does occur, how intense it will be.

02:36So what we know from that is that it doesn't account for a lot of things. It doesn't account

02:40for fuel, it doesn't account for ignition sources and things like that. So what we try to do

02:44here is we try to incorporate more data into a machine learning framework. So although a lot

02:49of this data can't be explained physically, we can use machine learning to get a better forecast

02:55way of using that data without the knowledge of the physics underlying it. So we incorporate

03:01fuel, ignition sources and also the existing weather that we previously used in the machine

03:06learning model to predict not just fire danger but the probability that fire might occur in

03:11a given location.

03:12I think machine learning has been a very, it can be a very effective way to merge together

03:18different type of information. Traditionally I mean physical models are able very well to

03:28Yeah, for example recent case in Los Angeles where fire really broke out in the wildland urban interface.

03:38This was really very severe because the previous seasons were actually characterized by very

03:45wet conditions which created an abundance of fuel that then was burned during the event.

03:55And of course this new method, the probability of fire, having the memory of the fuel abundance

04:03in their formulation allowed to really identify those regions that could be much more affected

04:10compared to simpler method they only consider weather. And this is why our prediction in this

04:16case was much more precise and pinpoint the exact location when very close to Los Angeles where

04:23fire really occurred.

04:24Yeah, yeah, so the purpose here is to try and produce a cost-effective piece of machine.

04:46that people can take away and use. So it's very expensive maybe to train but once we've trained

04:59the model here using our high processing computing power other centres can then take it and run

05:04it very cheaply and they can run it on their local laptops for example with very little cost.

05:09so I think the machine learning really really works.

05:38I think the machine learning really helps with getting the more precision on the location

05:42of where a fire might occur and this is relevant for agencies involved in suppression activities

05:47because then they can allocate resources into the locations where the fire might actually

05:52occur as opposed to just a widespread warning system that currently exists.

06:08In some ways it's a step change in fire forecasting because we're really going from suggesting where

06:28fire danger might happen, where there's the potential for fires to exist, to saying this

06:34is where we think fires will actually exist and for that reason it's quite a large leap

06:38forward I think in terms of fire forecasting.

07:08I think in terms of fire understanding, which is a form of fire being khiated further

07:10day to death I think in terms of fireструк begrudging looks radio

Category

Transcript

Recommended