Forecasting Daily Wildfire Activity Using Poisson Regression

Wildfires and their emissions reduce air quality in many regions of the world, contributing to thousands of premature deaths each year. Smoke forecasting systems have the potential to improve health outcomes by providing future estimates of surface aerosol concentrations (and health hazards) over a period of several days. In most operational smoke forecasting systems, fire emissions are assumed to remain constant during the duration of the weather forecast and are initialized using satellite observations. Recent work suggests that it may be possible to improve these models by predicting the temporal evolution of emissions. Here, we develop statistical models to predict fire activity one to five days into the future using Moderate Resolution Imaging Spectroradiometer (MODIS) satellite fire counts and weather data from ERA-interim reanalysis. Our predictive framework consists of two-Poisson regression models that separately represent new ignitions and the dynamics of existing fires on a coarse resolution spatial grid. We use ten years of active fire detections in Alaska to develop the model and use a cross-validation approach to evaluate model performance. Our results show that regression methods are significantly more accurate in predicting daily fire activity than persistence-based models (which suffer from an overestimation of fire counts by not accounting for fire extinction), with vapor pressure deficit being particularly effective as a single weather-based predictor in the regression approach.