Pyrantis Fire Outlook

Pick a place

Choose a town, use your location, or click any five-kilometre square on the map to see what the model expects there this year.

Town or community

Loading the mapThe figures beside it are ready already.

No fire expected Fire before August Fire after July

Northern Territory

Why the model expects a big year

Wet-season rain

2005 to 2019 average

April greenness, every year on record

What the model was given

When the country burns

Before August

The grass still holds water and the nights are cool. Fires creep, stay low and leave patches, and many are lit on purpose to break up the fuel.

After July

The grass has cured and the wind is up. The same spark now runs for days, burns hot enough to kill canopy, and takes fences and stock with it.

Twenty-six years of fire, mapped

Each bar below is one year's burnt share. Pick a year, or press play to watch a season spread month by month.

Loading the mapIt loads when you reach it.

No fire Burnt before August Burnt after July Test year

Pick a square

Click anywhere on the map to see what burnt there that year, what the model said, and every year since 2000.

How this year was used

How fire repeats

Four questions about the same ground over 26 years. The answers are what the models had to learn.

If a place burns, will it burn again?

What happens the year after?

Can one place burn twice in a year?

Years between one fire and the next

Does burning early stop a worse fire later?

Burnt early the year before Left alone

How good is it?

Every number here comes from years the model never saw while it was learning.

Squares placed correctly, 2023 to 2025

Balanced F1, the main score Accuracy

Tested 66 times by pretending to be in the past

We cut the record at each year from 2014 to 2024, trained only on what came before the cut, and scored every year after it.

What it gets wrong

Rows add to 100. Each row is what became of the squares that really fell in that class.

What the forest leans on

How much each input helped the random forest split the data. Weather in blue, greenness in green, fire history in ember, place in grey.

Three rules it follows

These are what make the scores above mean something in a real fire season.

It never sees the year it predicts

Fire history comes only from earlier years. Two tests check this, one by scrambling a year's answers and confirming nothing the model reads moves.

It stops reading on 30 April

The fire season runs May to October. August weather would describe a fire, not forecast it, so the model can be run in May.

The month is approximate

Satellites revisit the same ground every few weeks, so a fire is dated to the window it was seen in. Close to 31 July that dating blurs.