Skip to content

End-to-end machine learning pipeline for the prediction of extreme and dangerous wildfires.

Notifications You must be signed in to change notification settings

spaceml-org/pyrocast

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

53 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Pyrocast

Pyrocast logo

Introduction

Pyrocast is a end-to-end machine learning pipeline for the prediction of extreme and dangerous wildfires. More specifically the code in this repository allows you to find, forecast and understand the causal drivers of pyrocumolonimbus clouds, precursors the most large and unpredictable wildfires. The pipeline includes:

  • loaders to download and format the data
  • nrl_algorithm to find pyrocumolonimbus clouds and label the data
  • models to forecast the pyrocumolonimbus clouds
  • icp to understand the causal drivers of the pyrocumolonimbus clouds

The code for this repository is currently incomplete, the authors are contributing to the repository in their spare time so please be patient.

Getting started

Get in touch with [email protected] to get access to the data on Google Cloud Storage.

Data

The data is in a Zarr format, this allows us to load data that is associated to each hour of each day of each wildfire event using the ID numbers found in the wildfire_events.csv file. The data for the geostationary imagery, the pyrocast flags and masks and fuel and weather data each have their own Zarr directory.

Extracting data from a zarr folder event will yield Nx200x200 cube where N corresponds to the different wavelength channels, climate fields, etc.. These are detailed in the tables below.

Flags and masks

PyroCb_flags.zarr (array shape = 1)

N Content
0 PyroCb flag, whether or not scene contains PyroCb

PyroCb_mask.zarr (array shape = 1 x 200 x 200)

N Content
0 PyroCb mask, classification of pixel types according to NRL PyroCb algorithm

Geostationary imagery

Array shape = 18 x 200 x 200

Only some dimensions have entries which depend on the satellite source.

Himawari-8

N Channel wavelength [μm]
0 0.47
2 0.64
3 0.86
6 3.9
13 11.2
15 13.3

GOES-16 / GOES-17

N Channel wavelength [μm]
0 0.47
1 0.64
2 0.86
6 3.9
13 11.2
15 13.3

Weather and fuel

Array shape = 19 x 200 x 200

N Content
1 10m v component of wind
2 10m wind gust since previous post processing
3 boundary layer height
4 convective available potential energy
5 convective inhibition
6 geopotential
7 surface latent heat flux
8 surface sensible heat flux
9 surface vertical velocity
10 component of wind at 250hPa
11 v component of wind at 250hPa
12 fraction of high vegetation
13 fraction of low vegetation
14 type of high vegetation
15 type of low vegetation
16 relative humidity at 650hPa
17 relative humidity at 750hPa
18 relative humidity at 850hPa

About

End-to-end machine learning pipeline for the prediction of extreme and dangerous wildfires.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published