As my capstone project for the intensive data science bootcamp offered by Metis, I wanted to create a flexible tool that would be useful for wildfireprevention and resource deployment. Wildfires are a growing threat, especially on the West Coast of the United States. A tool that could predict high-risk areas would be useful so that expensive resources such as planes and equipment could be deployed efficiently.
I explored multiple possible routes to solving this problem, but eventually settled on using an Extra Trees classifier trained on a dataset of climate indicators, topography, and coarse woody debris coverage per county.
For a target feature, I leveraged a dataset of 1.88 million wildfires from Kaggle. This also allowed me to feature-engineer two lagged features to leverage past fire behavior as a predictive feature.
Using an Extra Trees classifier and oversampling the minority class, I was able to achieve precision and recall scores above 0.7 for most thresholds.
As a proof-of-concept, I attached this model to an interactive, web hosted visualization of the predictied spatial distribution of fires. I made this using Plotly and Streamlit.
If you want to check out all the steps that I took in this project, check out
fire_predition_nuts_and_bolts.ipynb.
If you're interested in how I created the interactive visualization and deployed
it to the web, check out wildfire.py.
Make sure to check out my slide deck and presentation that I gave for this project! You can also find my work in the Metis Graduate Directory.