Statistical Forecasting of the Wildfire Season in Alaska and across Northern North America Using Machine Learning

In recent years, wildfire has been increasing in the high northern latitudes, where the climate is warming more than twice as rapidly as in the global average. In Alaska and across the North American Arctic and boreal zones large fires occur in remote areas that are inaccessible by road. In Alaska, some Predictive Service Areas (PSAs), such as the Yukon-Kuskowim Delta, have recently seen a fire regime shift: large fires, which used to be rare, have become more prevalent. For resource planning purposes, fire management agencies have a need for seasonal predictions of wildfire activity and potential.

Our research project investigates the predictability of wildfire activity on a sub-seasonal to seasonal scale using a machine learning approach. We have assembled a dataset of predictor variables and fire activity related variables, and integrated them into a unified spatial framework based on the 0.25° grid of the ERA5 climate reanalysis. Meteorological variables such as the Fire Weather Indices from the Canadian Forest Fire Danger Rating System (CFFDRS), and annual snow-off date, peak snow depth and length of the snow season have been derived from ERA5. Fire activity variables include satellite-based fire detections (from MODIS TERRA) and fire perimeters provided by the Alaska Interagency Coordination Center (AICC) as well as Natural Resources Canada, re-gridded to the same ERA5 grid. Predictor variables further include sea ice concentration, atmospheric variables, sea surface temperature, teleconnection indices as well as lightning activity over Alaska.

The analytical framework is based on the ensembles of decision trees (Random Forest and Gradient Boosting) as well as an artificial neural network (ANN). The aim is to investigate the predictability of high, average, or low-intensity fire seasons before the season starts, as well as, in the short term, surges of fire activity mid-season. In spatio-temporal terms, we look at the PSAs used by the Alaska Fire Service, as well as Canadian provinces and coarse ecoregions. Sub-seasonality is defined in accordance with the four sub-seasons used in Alaska (wind-driven between season start in May and June 10, duff-driven June 11 to July 20, drought-driven July 21 to August 9, and diurnal effect between August 10 and the fire season end in September).

This research is co-produced by a team comprised of researchers from the University of Alaska Fairbanks and fire behavior and fire weather experts from the Alaska interagency fire management community. It is supported by the National Science Foundation (USA) under award #OIA-1753748 and by the State of Alaska.

Image caption: a) Average snow-off Julian day across the circumpolar and Alaska domains (source: ERA5). b) Wildfire activity by decade in Alaska domain, 1940s-present, overlaid over Alaska Predictive Service Areas (PSAs) (source: AICC, Environment Canada). c) MODIS Terra fire detections in the Yukon-Kuskokwim Delta PSA (SW Alaska) (source: AICC)