Project Smokey
Wildfire Forecasting using Machine Learning
Project Smokey provides wildfire forecasts for any location in the United States selected by the user in the graphical map application, using weather forecasts as well as regional climate and vegetation information.
Technical Details
Forecasting Model
Instead of using a manually-constructed forecasting model which might not fully take into account all of the variables or underlying complexity of the application, as well as requiring extensive time, effort, and expertise to construct, Project Smokey uses a set of TensorFlow-powered neural network models, trained on a database of 1.88 million wildfires in the US taking place over roughly the two decades leading up to 2015 as well as the specific weather conditions for the days leading up to that wildfire. Essentially, given weather as an input, the models predict the chance that a wildfire will break out, and its severity. To get a wildfire prediction, then, we just need to feed weather predictions into the wildfire prediction model.
In order to account for hidden differences in each region that we can't directly pull data for that might effect how weather interacts to produce wildfires, such as specific differences in climate and vegetation that we don't have granular-enough data to account for, or regional topography, we bin (group) wildfires by climate/vegetation region, and train separate models on each of those groups of wildfires (and associated weather, of course).
User-End Application
The user-facing end of this application is written entirely in Rust, using the available GTK3 bindings. It consists of two threads, a GUI thread that runs the user interface, and a back-end thread that queues up requests for predictions at the various locations the user clicks on in the map UI, and runs the appropriate TensorFlow model on each request, returning the predictions to the user interface to be displayed. We hope that by running the TensorFlow models from Rust, and in a separate thread, some performance gains will be achieved.
Project Architecture
Overall, then, the project has three "phases":
- Creation of the user-facing application, which we can plug the models into
- Data processing, binning, and normalization, in preparation for training
- Model training