Evaluating a Probability-Based Model for Prescribed Fire Forecasting with Machine Learning

Prescribed fires are a class of burns that are intentionally lit by forest managers in order to reduce the risk of future fire hazard. These burns are common throughout the southeastern United States. A model has been developed that produces probability forecasts of these fires as a function of weather criteria, with the goal of providing improved guidance for the ignition of these burns. Such a model is also needed in order to provide more accurate emission data for air quality models, which erroneously assume persistence burning for prescribed fires. Persistence burning is a poor assumption since these fires are lit by people, generally last for much shorter time scales than wildfires, and can vary largely in their spatial extent and distribution from one day to the next.

Certain weather criteria must be met in order to safely conduct prescribed burns, and fire managers account for these criteria when deciding whether or not to burn on a given day. These include thresholds of wind speed, relative humidity (RH), and boundary layer depth. For example, fire managers rarely burn whenever the wind speeds are too high because high winds can cause a fire to grow rapidly and become difficult to manage. Likewise, it is not recommended to start fire in calm conditions because a lack of wind can cause the fire smoke to accumulate and negatively impact air quality. Hence, most experts recommend wind speed to be within a certain range that does not include extreme values at either end. Similarly, this reasoning holds for parameters such as RH.

A climatology has been developed that overlays the number of these criteria met with satellite-derived fire detections for each day within the climatology. The hypothesis is that as the number of weather criteria met increases, the number of fire counts should increase; however, this relationship has not been quantified. In order to quantify this relationship and evaluate the model’s predictive skill, a machine learning approach is being proposed. This approach will employ decision trees in xgboost. The predictor variables will include temperature, RH, 10 m wind speed, boundary layer depth, Lavdas Atmospheric Dispersion Index, daily total precipitation, and the climatological probability of fire. 5 years of data (2015 – 2019) for the monthly period January 1 – January 31 will be used. The predictor variables will be sampled at 1800 UTC, using NAM 12 km resolution and Hazard Mapping System fire detections. The algorithm will use the input data to make a prediction (i.e., the probability of fire), and the output will be compared with actual observations of fire detections.