The main challenge was predicting how quickly a wildfire would reach an evacuation zone using a very small dataset of only 316 events. Traditional machine learning models overfit on this limited data and fail to provide the continuous probabilities needed for emergency planning. The goal was to build a robust model that ranks fire urgency and provides calibrated forecasts, focusing especially on the critical 48-hour window for decision making.

The analysis showed that these mathematically transformed features performed much better than raw spatial data in predicting imminent threats. By compressing complex physical data into single indicators, the model successfully ranked fire urgency and provided reliable probabilities for different time horizons. The recommendation is for emergency response teams to use these calibrated probabilities, particularly the 48-hour forecast, to trigger threshold-based evacuation plans.

I used a survival analysis pipeline with LightGBM to handle the extreme data scarcity and right-censored data. A key part of the solution was advanced mathematical feature engineering. Instead of just using raw coordinates, I created features that capture the physical momentum of the fire, such as the Dimensionless Threat Gravity Index and Eikonal Time-Field Gradients. These engineered features allowed the model to better understand the fire's speed, direction, and acceleration toward evacuation zones.