Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Titre : Integrating Remote Sensing, Behavior Modeling, and Machine Learning to Better Understand the Patterns and Drivers of Wildfire

Auteur : DeCastro, Amy

Université de soutenance : University of Colorado at Boulder

Grade : Doctor of Philosophy (PhD) 2022

Résumé
Our understanding of and approach to fire is becoming increasingly important due to population increase and climate change. Current management strategies, which are focused mostly on basic resilience are not enough given the increasing wildland-urban interface and climate conditions ripe for wildfire. Informing future management strategies requires furthering our understanding of fire spread at a regional scale — the scale at which management decisions are made. The recent development of fire records cataloging thousands of fire events at daily and event-level resolutions has made it possible to investigate fire behavior at the regional scale. This research integrates remotely sensed data, wildland fire behavior modeling, and machine learning methods to further our understanding of the patterns and drivers of wildfire. High-resolution remotely-sensed datasets from several platforms were used to train random forest classification and regression models to identify drivers that matter at both the daily burned area and fire event scales, across individual case studies and collectively across thousands of events. WRF-Fire was used to investigate the intersection of drivers and how our current understanding of fire behavior, as represented by numeric models, compares to what we observe at sub-daily, daily, and event scales. Specifically it addresses data deficiencies in inhibiting the development of wildland fire behavior models, and identifies the most influential regional-scale drivers of daily fire growth. This work shows that environmental heterogeneity plays a critical role in daily fire spread ; presents a method for updating fuel data for use within wildland fire behavior models ; and bounds the uncertainty associated with inaccurate ignition data within burned area and propagation direction forecasts.

Présentation

Version intégrale