Skip to content

Modeling Ignition Probability

Monthly vegetation ignitions

Wildfires are a seasonal pattern that typically occur after long periods of high temperatures and low humidity. This patterns is simlar for ignitions.

We've got a few notebooks for working with CALFIRE's public ignitions database, which contains point records, causes, and descriptions for 169,968 ignitions covering 2016-2019. We pulled these data in order to develop spatially-explicit models of ignition probability that vary with weather and vegetation patterns.

We used the elapid package to run Maxent species distribution models using the ignition point occurrence records, paired with PRISM climate data and Forest Observatory vegetation data as environmental features. As a result we were able to generate monthly ignition probability models for a series of fire types (grass/shrub, forest, lightning strikes, and power line).

Downloading data

There are pretty serious data requirements for building these models. This is pretty much all handled in the following notebook:

ignition-probability/notebooks/DownloadPreprocessData.ipynb

It pulls all the data from cloud storage to a local directory, then ensures the files are consistently resampled and reprojected to the same region. It also references a Google Earth Engine script used to clip and export monthly PRISM climate data.

Exploratory data analysis

We ran a series of tests to better understand the ignitions data - including the frequency and timing of ignitions from a series of different ignition classes. You can find this notebook here:

ignition-probability/notebooks/IgnitionsEDA.ipynb

Many of these ignitions have specific codes that describe the ignition cause. You can find the lookup for these codes in the ignition-probability/nfirs-incident-type-codes.pdf document.

Model training

Using the monthly climate data, annual vegetation data and the ignition point data, we trained a series of Maxent models for predicting spatial patterns of ignition probability. You can find it here:

ignition-probability/notebooks/ModelTraining.ipynb

This notebook:

  • Filters ignitions data to just points with the causes we're interested in
  • Annotates each point with the month/year of climate/vegetation data from when it occurred
  • Samples a random set of "background" samples used to define the null probability space
  • Trains and saves models for each ignition type
  • Applies the model to monthly rasters throughout the year

Much of this work got put on hold because 1) it was taking a long time to apply models to statewide datasets, and 2) it was hard to understand some of the data-driven patterns we were seeing.

I think the main issue is that there are so many ignitions, but so few of them escape containment to become major wildfires. This is difficult because we don't yet have a good way to discriminate between which of the ignitions are dangerous and which are not.

As a result, the climate conditions in the areas where the majority of fires occur are not in places we might expect. The mean/max annual temperatures are lower than average, and the vapor pressure defecits are higher than average. This is strange because we were expecting fires to disproportionally start in hotter and drier areas. And that's not the case. Why not?

Anyway, we should pick it up again soon because it's pretty cool and we're pretty far along.

Living in a garbage country

I learned a lot exploring the CALFIRE ignitions data. We start a lot of fires every year! Yet so few of them become catastrophic. But the most important thing I found was empirical evidence validating our perspective that we live in a garbage country.

There are more recorded garbage fires in California every year than there are vegetation fires. And there are a lot of vegetation fires. Look!

2016 Veg. fires   : 9,097
2016 Garbage fires: 9,456

2017 Veg. fires   : 10,222
2017 Garbage fires: 10,043

2018 Veg. fires   : 3,938
2018 Garbage fires: 5,223

2019 Veg. fires   : 7,699
2019 Garbage fires: 11,455

Garbage country!

So, what kinda garbage is burning?

There's all sorts of shit!

'Garbage fire',
'Outside trash fire',
'Dump or landfill fire',
'Construction landfill fire',
'Dumpster fire',
'Compactor/compacted trash fire',

Sweet. Garbage country.

Back to top