Wildfire Fuel Mapping with Convolutional Neural Networks for Remote Automated Exposure Assessment

• Author / Creator

  Bennett, Liam

• While beneficial to the natural environment in many cases, wildfires become hazardous when they intersect with the built environment. As such, there is an ongoing effort to understand the fire environment, the fuels it contains, and the way that wildfire interacts with the built environment. In particular, communities and structures in the wildland urban interface (WUI) are often assessed to determine their exposure to wildfire ignition processes. Exposure assessment workflows at both scales require the identification and classification of hazard fuels. Thus, a method is presented for the remote automated detection of hazardous wildfire fuels and the application of fuel detections to community and structure scale exposure assessments. The outputs of the automated processes are intended to supplement rather than replace existing practices and form a preliminary basis of information that is both simple and rapid to collect, process, and interpret.
  Two workflows are devised to detect and classify large overstory trees from RGB imagery in the boreal, rocky mountain, and foothill natural regions of Alberta, Canada. The first workflow considers remotely piloted aircraft systems (RPAS), are used to collect RGB imagery. A convolutional neural network (CNN) is trained to detect trees in the overstory and to classify them as coniferous, deciduous, or snags. F1-scores reach 74.5% for tree detection and achieves classification F1-scores of 97.3%, 94.4%, and 90.9% for coniferous, deciduous, and snag classes respectively.
  The second workflow uses RGB satellite imagery to detect individual trees using a second trained CNN. An R2 of 0.76 is achieved comparing automated tree detection density to manual annotation density. A k-means clustering algorithm is used to determine winter ‘leaf-off’ imagery and classify trees as ‘green-in-winter’ or ‘brown-in-winter’, an indication of coniferous or deciduous trees. Classifications from satellite imagery reach an F1-score of 0.82.
  Finally, tree detections and classifications from the RPAS model are visualized around a structure in the context of a FireSmart home assessment, and it is discussed how the workflow could be used to provide informative maps of tree vegetation around a structure. At a community scale, fuel maps derived from the satellite tree detection and classification workflow are used in an existing community exposure assessment workflow to explore how fuel maps generated in this manner may be applied.

• Subjects / Keywords

  • Remote Sensing
  • Wildfire
  • Machine Learning
  • Convolutional Neural Network

• Graduation date

  Fall 2023

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-rcc5-z865

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.