FuelNet: An Artificial Neural Network for Learning and Updating Fuel Types for Fire Research

Abstract

Wildfire is a significant driver of forest and land cover change in the central interior of British Columbia, Canada. Fuel type maps are a primary input to fire behavior calculations and simulation studies that assess wildfire threat at the landscape level. However, these thematic maps are not easily produced at the scale and speed needed to assess and mitigate wildfire threat on an annual basis. The objective of this research was to explore how an artificial neural network could be used with remotely sensed satellite imagery to map and update fuel types on an annual basis. We applied the artificial neural network over a 40 000-km2 landscape in central interior British Columbia that burned from a megafire in 2017. Fuel maps were generated for the years 2014–2018, assessed through an independent validation, and evaluated against an existing fuel type map. The highest cross-validation overall accuracy during training was 66.5% and overall accuracy from the independent validation was 63.1%. Generally, the maps had fair agreement with the existing fuel type map (circa 2016), with Cohen’s Kappa ranging from 0.28 in 2018 to 0.35 in 2015. Several recommendations are provided for future research using artificial neural networks for fuel typing such as assuring quality of training samples through rigorous standards, designing the network architecture, choosing appropriate cost functions and regularization, incorporating learning of temporal features, and identifying novel fuel types from the output activations.