This is a machine learning application for predicting wildfire spread. The main architecture is written in PyTorch, and GeoPandas and rasterio are used for data processing.
I created this project with the intent of challenging myself to build a complete machine learning application, from the model architecture to a web application. I considered actually training and deploying a model, but was unsatisfied with the 15-kilometer constraint imposed by the dataset. Scraping historical data to create a new dataset was out of the scope of the project, so I decided against it.
The repository is broken up into several components:
train/ contains the model architecture and the training loop.
preprocessing/ describes how to download and handle the geographical data the model requires.
app/ contains a script for generating fire spread predictions, as well as code for an interactive map to view those predictions.
Each folder has a README with specifics for that aspect of the project.
Under the BSD 3 License, you are free to distribute and modify the contents of this repository to your liking.
Training data and visualization notebooks provided by Casey Graff.
Original FireCast architecture and algorithm described in [Radke et al., 2019].
Mapping demo uses Leaflet and Leaflet.heat.