A day of very high fire danger in the year 2016: the grass browning, fallen timber drying, pollen in the air, eucalyptus oil gathering in a flammable haze.
A fire starts south of Cowra, in the central west region of New South Wales. Cowra has a population of a bit under 10,000. The ignition is logged by the local Rural Fire Service (RFS) crew and by a Japanese satellite detecting infra-red from 35,000km above the earth.
At the RFS headquarters in Sydney, a bushfire prediction model begins factoring in estimates of the amount of grass and timber in the area, how the different kinds of vegetation burn, the fire history, the moisture levels, the terrain, the wind direction, wind speed, humidity, temperature.
The answer comes back in 10 minutes. The map shows how the fire would spread in 24 hours - an oval-shaped blister of red contour lines running about 30km east to west.
GIF caption: Phoenix RapidFire simulation in action. The orange-red pixels indicate potential flame height, light orange lines are the modelled fire perimeters. White bubbles are tracers showing the plume rise above the fire, responsible for distributing embers that can start spot fires.
'The state glows red'
Today a total fireban has been issued for parts of NSW and a very high fire danger warning is place for the bulk of the state. Temperatures in west Sydney are forecast to rise above 40C.
Hot summer weather is nothing new, but studies show in recent decades there has been a general trend towards higher temperatures and lower rainfall across the continent.
Veteran fire fighters are reporting the worst fire behaviour they've ever witnessed.
The bushfires season is growing longer, according to the Climate Council, while the firefighter union is calling for more resources.
At the NSW RFS headquarters, the emergency response is conducted from a central covered atrium, one wall an enormous multi-story screen showing intel from across the state: satellite mapping, television news broadcasts, a scrolling feed of typed reports from the field.
In recent years, the RFS and other Australian fire services have been investing more and more in bushfire behaviour analysis - a mixture of art and science that's being transformed by massive computing power.
Working from a small office adjacent to the headquarters' incident room, the bushfire behaviour analysis team sees what no-one else sees: the infernos that would roar through the state if the network of local fire crews were not suppressing them as soon as they broke out.
"Very quickly a lot of the state becomes coloured in with potential fires," says Dr Simon Heemstra, the head of the team and a 20-year veteran of fire fighting.
"The state glows red."
The escalating sophistication of the models is in a race to understand a new kind of bushfire. Even as the models improve, Australia's fires are becoming less predictable.
Phoenix from the ashes of Black Saturday
Australia's worst ever bush fires came on February 7, 2009. Black Saturday. Fires across Victoria claimed 173 lives and destroyed more than 2,000 houses.
Hunkered down in the state's Country Fire Authority (CFA) headquarters that day was Dr Kevin Tollhurst - a University of Melbourne academic who had spent the last five years quietly developing a computer model of bushfire behaviour.
Dr Tollhurst had partnered with an expert programmer, Dr Derek Chong, to convert the prediction models into a rudimentary software program they named Phoenix RapidFire.
"There's always been a need for software to predict fire spread, but there's never been anything available to deliver a useful output in usable time frame," Dr Chong says over Skype from his University of Melbourne office.
"Unless you provide a prediction in 10 minutes the fire managers tend to make up their own mind.
"That's the window we were really aiming for, for any size fire in the Australian landscape."
The subsequent Royal Commission into the services' response to the natural disaster cited Phoenix as an example of the potential utility of software prediction models.
This doesn't meant today's computer model could have stopped Black Saturday.
"Going back in time is always tricky," Dr Chong told Hack. "Phoenix RapidFire I don't believe would have stopped any of those fires - what it may have helped do is inform public about where the fires are going.
"But even if you do have perfect predictions, you are still dealing with a very large complex scenario with multiple fast moving fires."
After Black Saturday, Phoenix was fast-tracked from the research space and rolled out in Victoria in the summer of 2010-11. Different versions of computer prediction models are currently being used throughout Australia: Australis, Spark, Prometheus. They have grown in complexity to take into account the growing amounts of data weighed in ever more sophisticated algorithms.
NSW RFS also uses Phoenix RapidFire, though Dr Heemstra is more dubious about its accuracy. His department runs the computer modelling in parallel with more conventional manual predictions. The calculations are performed by fire behaviour analysts who take into account the same variables as Phoenix - variables like weather, fuel loads and topography.
Dr Heemstra is cautious about rolling out Phoenix too soon - he sees fire behaviour analysis as both a science and a gut feeling of where the fire will travel.
"There are things that just don't exist in the models," says Dr Heemstra.
"Nine out of ten times we still see manual calculations outperforming the computer."
'Many of the rules established no longer apply'
One of the worst fires in recent years in New South Wales came in January 2013. The fire burnt 90 per cent of Warrumbungle National Park in the central northern region.
A subsequent coronial inquiry heard the fire was so intense it became almost impossible to predict.
"The recent spate of extreme wildfires has cast some doubt upon the traditional approaches used to understand and predict bushfire behaviour," an expert bushfire scientist, Associate Professor Jason Sharples, told the inquiry.
"It seems that in the case of extreme wildfires, many of the rules established as valid on smaller fires simply no longer apply."
The column of heat above the blaze grew so large it coupled with the atmosphere, and formed a fire-generated thunderstorm, what's called a 'pyro-cumulus'.
The smoke column was 14 kilometres high (a passenger jet cruises at an altitude of about 10km). The storm threw embers and lightning far in front of the fire and started new spot fires. This made it very difficult to predict fire behaviour.
Added to this, the people in charge of managing the fire had not understood the practical limits of fire prediction maps, the inquiry heard.
The Coroner reported:
"Due to the novelty of the prediction system, the controllers certainly, and perhaps also the people running the prediction modelling, did not fully understand the vulnerabilities of the methodology especially in extreme weather conditions or where extreme fire behaviour was being experienced.
"Because so many variables are involved in these calculations, many of which are difficult to measure, there is a constant and recognised risk of inaccurate prediction."
Dr Heemstra shakes his head at the memory of the Warrumbungle pyro-cumulus.
"Phoenix underpredicted the fires," he says.
'It's all about buying time'
Whatever the relative accuracy of computer and manual predictions, it's clear where the future lies: computer modelling has seemingly infinite potential.
Dr Heemstra reckons computer predictions will become as good as manual ones within two years, and the gap will then widen further with faster computers, more complex algorithms, and more detailed, more accurate data.
Instead of one simulation per fire, there will be thousands, based on variations in temperature or wind speed. Instead of making just one prediction, the program will compute a range of probabilities that a fire will behave in a range of ways
"If we have 200 fires and want to run 1000 simulations for each fire, you actually need some serious computing grunt," says Dr Heemstra.
Another development will be fire prediction moving from fire service state headquarters and into into the hands of firefighters near the frontline.
This summer Dr Chong has been tracking every ignition in the state and monitoring the Phoenix predictions through his smartwatch.
"It calculates the number of properties the fire will potentially impact and converts that down to a series of beeps and buzzes that come out of the smartwatch.
"It's all about buying time."
Decentralising fire prediction raises plenty of issues, such as whether the models can be understood without expert analysis, and the cost of the technology.
"As we build confidence in these models they will be made available to brigades on the ground. As the fire stats they will be able to see what these fires will do," says Dr Heemstra.
He's more cautious about rolling out these models to the general public. Fire behaviour analysts are trained to understand the model's predictions are always uncertain, always based on inputs, such as weather forecasts, which are themselves imperfect. A member of the public might not appreciate this, and the prediction would give them an unrealistic impression of their fire danger.
"There's a potential of causing panic," says Dr Heemstra. "There's a potential of causing complacency.
"You've seen 20 fires that could burn you house down and the fires have never arrived, then there's one we are really worried about and we're saying you need to do something. You say, 'but there's been 20 fires that never got here.'
"There's a danger of crying wolf."
A local crew report is logged on the intel feed: the fire near Cowra has been extinguished. It was put out before it spread beyond the field of stubble where it started.