Wildfires versus big data: how South Africa’s 49-year-old fire warning system has stood the test of time

Share on facebook
Share on twitter
Share on linkedin
Share on email

When wildfires came close to destroying the lush, 317-year-old Lourensford Wine Estate in Somerset West at the start of the year, few expected that it was to be the start of one of the worst seasons for sporadic blazes in the region for many years.

A prolonged drought, strong winds and many conspiracy theories about arson have led to a tough 2017 for Cape Town. As early as 14th January, the City issued a statement saying that it had already spent half of its annual budget for helicopter resources.

Fortunately, there have been few reported casualties as a result of the unusual number of wildfires, which is testament to the skill and experience of those who are call to try to stop them. Fighting bush fires in the Cape or anywhere else is difficult, dangerous work.

It’s also increasingly high tech, as innovations in the world of disaster response improve communications and tools. Which is why it might surprise you that one of the key weapons South Africa uses is old fashioned, outdated and also extremely effective at what it does.

What is fire risk?

Fires are inevitable and, in some cases, essential. They’re part of nature’s way of clearing out dead and dry plant matter and opening up land to new growth, and there’s no way of stopping them altogether. They’re also very dangerous, and very expensive. Which is why over the decades climatologists and data scientists have developed risk models of where they think fires are mostly likely to break out.

These Fire Danger Indexes (FDIs) are common throughout the world, from Canada to Australia. Firefighters and local governments use them to allocate resources in particular areas on a particular day, and to decide whether or not a small blaze might develop into huge one. Based on the current FDI for a region, they might send a fire engine, a helicopter or nothing at all.

FDIs are also used by foresters and farmers. In South Africa people are legally obliged to check the FDI for their region before burning a firebreak, for example. Under the terms of the National Veld and Forest Fire Act (1998), the South African Weather Service (SAWS) is tasked to produce an FDI daily alert and circulate it in multiple languages to three television channels, two newspapers and three radio stations in every region.

Other organisations can produce their own FDIs using different calculations – and they do, but more on that later – but SAWS has the sole mandate for producing FDI alerts.

The FDI, then, is taken as seriously as it deserves. But what is it and how is it worked out?

A simple equation

In South Africa, the official FDI is calculated according to the Lowveld Fire Danger Index (LFDI), which produces a rating of between 1 and 100, which is in colour-coded bands for blue, green, yellow, orange and red as below.

The full list of warning states in the Lowveld Fire Danger Index.

The LFDI is something of a quirk in modern meteorological tools though. It was first developed in Zimbabwe/Rhodesia in 1968 by a scientist called Michael Laing, and slowly adopted by local weather stations in Mpumalanga and other Lowveld areas before becoming the national standard much later on.

“A lot of modern scientists look down it because it’s an empirical scale,” explains Kevin Rae, the chief forecaster at SAWS, “Which is calculated on three primary components.”

While predicting weather in general is notoriously difficult – SAWS has recently had a new Cray supercomputer installed in its basement – the FDI is calculated on three basic measures which are captured from weather stations. The formula is worked out using air temperature, relative humidity and wind speed. How hot is it? How dry is it? How fast is the wind blowing? The actual equation looks like this.

[su_highlight background=”#99c0ff” color=”#000000″]FDI = {(T-35)-((35-T)/30)+((100-RH)*0.37)+30}
Where T = Temperature (degrees C) RH = Relative humidity (%)[/su_highlight]

“The model has been refined over the years,” Rae says, “So there’s a final ‘antecedent rain correction’ based on rainfall in the last 21 days, and we now update it every six hours rather than every 12.”

There are some obvious problems with this model, says Rae. Wildfires spread for a number of reasons: the direction of the wind is probably the single most important one, but if a fire is moving uphill it can spread more quickly than if it is going down, as there’s more surface exposed to flames. Even cloud cover can make a difference.

The region is important too. As its name suggests, the LFDI was never intended to be used in the Cape regions: it might be hot and dry in the Karoo or the Kgaladei deserts, but the chances of sand spontaneously combusting is low no matter what the FDI says. There are plenty of other FDI models worldwide which take this kind of information into account too, modelling risk based on local plant types and “non-geographical data”.

Today, Rae says, SAWS uses incredibly sophisticated tools to create the FDI index. Using software purchased from Meteo France, predictive models from the British Met Office, satellite data and its own network of over 150 weather stations, SAWS can create highly accurate weather predictions and associated FDIs down to a 12km grid area across the whole country. That resolution will increase this year to 8km2, and Rae says it’s likely to go down to 4km2 in the near future. Just 20 years ago, the best weather models only worked to the level of 50 square kilometres.

It’s a sophisticated organisation with a lot of clever people working there. So why is LFDI so basic?

We use it because it works

Karen Steenkamp and Philip Frost work on weather and fire mapping and sensing at the Mareka Institute of the Council for Scientific and Industrial Research (CSIR). Earlier in his career, Frost pieced together a map of how the tragic 2001 blaze in the Kruger National Park spread, a process that he points out took days using physical photographs from satellites, but would now take minutes using online tools developed and distributed for free by Nasa.

Those Kruger fires, in which 23 people died, were a turning point for local climatologists. The key lesson learned was that at the time there was no system in place for combining weather data with fire risk data to produce detailed alerts, and a sudden change in the wind conditions was responsible for the blaze going out of control and causing loss of life.

CSIR got to work, publishing almost a paper a year on evaluations of risk indices and their appropriateness for South African conditions.

Despite this, LFDI remains, and there are two reasons that it has held on so long. Firstly, it works: in a 2013 paper Steenkamp and Frost showed that although other models were more sophisticated, that didn’t make them best suited to the local climate and biomes. Along with colleagues, the pair tested LFDI against the Canadian Fire Weather Index (FWI) and the Australian McArthur Forest Fire Danger Index (FFDI) and McAthur Grassland Fire Danger Index (GFDI). A fifth model, from the US, has also been tested in the past.

Using historical weather data and fire maps, what Frost and Steenkamp found was that while the FWI was the most accurate model for predicting fires the LFDI came second most of the time. Curiously, LFDI hadn’t been evaluated in many of the previous research.

“Previous studies had assumed that the LFDI would be ineffective, so didn’t include it,” Frost says.

The second reason that the LFDI has held on, says Frost, is that it’s simple; firefighters and planners know and understand the LFDI. With years of experience, they understand exactly what a yellow or red warning for their area means. When they’re making decisions and either can’t reach SAWS or feel they need an update, they can calculate the LFDI for themselves if they feel they need to.

The FWI, on the other hand, uses data that includes vegetation types and indexes the types of potential fuel and moisture levels over two month periods. As such, it can only be calculated once a day.

“We do need predictive models,” Frost says, “But they’ve got to be useable by an operational person in the field.”

Mobile first

The times they are a’changing, though. There are vast repositories of near real-time data to draw upon, and satellites produce a thermal map of South Africa with a resolution down to 2.5km every 90 minutes – which means they can see fires as they start and spread. In 2019, says Frost, new geostationary satellites will be launched which will produce images once every five minutes, giving us as close to realtime information about the entire country as we can get.

What’s more, we have near infinite resources of cloud computing power to make sense of the data those satellites generate, too, and this opens up a lot of possibilities. But any new national system has to be well tested, accurate, simple to use and robust enough that it won’t be changed again for a long time.

Instead of trying to replace LFDI, then, Frost says that he lobbied for the formal adoption of it as the official system in order to eliminate confusion in the network of emergency responders who need clear communication and direction (it was gazetted as such in 2010). This leaves the Meraka team freer to research better ways to predict fires that complement or run alongside LFDI.

To this end, CSIR has been running the Advanced Fire Information System (AFIS) for the last 12 years. Frost has developed into an exceptionally powerful and simple to use dashboard and mobile app that draws in data from the LFDI, but supplements it with satellite imagery from NASA, detailed reports of active fires, weather information and an alternate FDI. All the things that were missing in the Kruger fires.

So long as they have an internet connection, which isn’t always a given when it comes to areas susceptible to wildfires, response commanders on the ground can get the full suite of AFIS tools directly to their phones. If they can’t there’s the tabulated LFDI to fall back on or a call back to their HQ.

“AFIS is unique because it caters for everyone from the planner to the implementer,” Frost says, “We have very close links with our users, and it’s embedded in communities across Southern Africa as well as Chile, Brazil, Oregon, Indonesia and Portugal.”

Frost has also begun integrating the Canadian FWI indices into the AFIS dashboard for registered users. For the first time this year, Frost says he’s been working closely with the authorities in the Western Cape to use data generated by FWI as well as LFDI to pin-point fire risk areas. It’s too early to come to solid conclusions – but there’s been a lot of pressure on the fire teams in the region this year and Frost says he’s confident that the test has been successful.

Here comes the big data bit

All models can be improved, however, and the future of fire danger indices – as with so much else – lies in machine learning. Given enough computing power, FWI combines near-realtime information about current weather conditions with vegetation maps and fuel conditions.

But adding in historical data and training models over long periods of time can account for a huge number of other factors including land topology and more. In 2015 a team from Portugal and Slovenia published a paper in which they took algorithms designed to model how genes evolve and applied to fire prediction, but some of the most interesting work in this area is taking place right here, in Johannesburg.

Dr Bonolo Mathibela is a research scientist at IBM’s lab in the Tshimologong Precinct in Braamfontein. In December she presented the results of a project conducted by her team which took fire incident data from the Cape Town open data platform and overlaid it with historical weather maps from IBM’s own weather portal. Using neural network techniques and the kind of cognitive computing with which IBM is now synonymous, Mathibela says that she has developed a model which was more accurate than existing indices such as the McArthur FDI.

IBM’s dash-in-progress has great UX.

“When we did the comparisons we found that our model works incredibly well,” Mathibela says, “Especially for the high and extreme risks, we were able to predict those quite accurately.”

The next stage, Mathibela says, is to try and foster collaborations with government and bodies like CSIR in order to access live data which could be fed into systems like IBM’s Watson.

Which all means that we could, thanks to the hard work of organisations like CSIR and the resources of IBMs, be on the cusp of exciting breakthroughs in the fields of big data science and fire fighting.

In the meantime, next time you see a fire truck or helicopter loaded with water heading off to the winelands, spare a thought for a simple, 49-year-old piece of maths which helped to send that crew on its way.

[Main image: Devils Peak/Table Mountain fires 2009, CC BY Warrenski]

Adam Oxford

Adam Oxford

Adam is the Editorial Director at htxt media. He has been writing about technology for almost two full decades now. In a previous life, he was the editor of PC Format and Digital Camera Shopper in the UK, before going on to work as a freelance journalist for seven years. His work has appeared in or on Stuff, The Guardian, Linux Format, TechRadar, Wired.co.uk, PC Gamer, Green Futures, The Journalist, The Ecologist and The Review. Adam moved to South Africa in 2012 and loves 3D printers, MakerFairs and tech hubs. He hates seafood. None of his friends remember this when cooking.