With a warming atmosphere and expanding housing settlements, wildland fires have become more dangerous than ever. Using a dataset of 1.88 million U.S. wildfires, I try to predict what size class a fire will reach based on features known at the time of discovery. This will not replace current monitoring systems, but hopefully will help augment firefighting monitoring systems by bringing attention to potentially overlooked fires.
I choose F1 score as my metric for evaluating my models because there is a need to balance the danger to lives and property from underestimating a fire and the limited resources available for fighting wildfires.
All notebooks under the prefix of Fires2.
- Assess available columns and perform basic EDA.
- Fill in missing location data and extract additional date features.
- Look up location elevation based on lat/long.
- Query and join weather data for corresponding fires
- Clean weather data and hold out rows for modeling.
- Begin with logistic regression and random forest models.
- Further modeling: under construction