Exploring the use of machine learning for daily fire growth prediction in Alaska

Abstract

Wildfire is a natural but often hazardous part of the Alaskan ecosystems. Physically based wildfire models range from simple relationships used for rapid, in-situ fire behavior analysis to complex weather models used for prediction over several days and weeks. Physical models in Alaska, however, often struggle to integrate weather forecast information to make predictions beyond just a few days. The random forest model explored here is able to leverage an array of variables to identify days of enhanced and reduced satellite fire detections. Peaks and lulls in activity are accurately identified, though exact magnitudes are often incorrect, especially when wildfire suppression efforts occurred. This study emphasizes the use of reanalysis weather variables in addition to antecedent fire activity, highlighting the usefulness of variables like vapor pressure deficit for use in quantitative prediction. By applying weather forecast data, the model generated simulated wildfire forecasts. These forecasts show some success at identifying peaks and lulls in fire activity. Effective lead time varied widely ranging between 1 and 10 days, mostly dependent on the weather model performance. By providing specific timing and using real ensemble forecasts for medium term prediction, a model likes this fills a potential open niche in fire predictive services. Machine learning may be especially useful for its relative efficiency and ease of automation.