Panel Talk: Natural Language Processing and Spatiotemporal Visualization for Wildfire Event Reconstruction

Fire management is at a crisis point. A century of fire suppression, extensive urban wildland interface development, and climate change have led to nearly unstoppable fires. Attempts to avoid some of these fires using public power supply shutoffs have adversely affected tens of millions of citizens and cost billions of dollars. Major wildfire events have for many decades been documented with “after action reports.” More than 3,000 such reports are publicly available. These describe the chronology of events, including weather, fire behavior, people and equipment deployed. The impacts of each fire on lives, properties and the environmental are carefully documented. Historically, such reports have included only very limited spatial data. This makes it very difficult to comprehend narrative references to fuels, topography and local landmarks. But elsewhere within State and Federal databases, extensive spatial data are available, detailing environmental condition before, during and after both prescribed fires and wildfires. Our project is using a Natural Language Processing approach to extract relevant entities from free text, automatically linking them to spatial data, and using the combined datasets to parameterized historical fire scenarios within standard fire simulation models. We use a machine learning model (SpaCy BERT) to perform “named entity recognition,” including the identification of weather, people, equipment, institutions and geographic places as objects. Geographic places are cross referenced to fire names, dates and states within GeoMac and MTBS fire perimeter databases. These perimeters are in turn used to extract information from Landfire and analysis-ready satellite data archives suitable for fire modeling.