Tuesday, 30 January 2024
Hall E (The Baltimore Convention Center)
We describe a machine learning (ML) system under development for future prediction of pyrocumulonimbus (pyroCb). These unique storms are a dangerous and severe type of fire weather, presenting many hazards to firefighting efforts and communities along the wildland-urban interface. They also serve as a vertical transport pathway (large chimney) facilitating rapid injection of smoke into the upper troposphere and lower stratosphere. Atmospheric variables and thermodynamic profiles coinciding with fires in western North America during 2013 -2021 were collected from two global models: the NAVy Global Environmental Model (NAVGEM) and the National Centers for Environmental Prediction’s Global Forecast System (GFS). We investigate the suitability of these global models for ML applications using several common variables, such as temperature, dew point temperature and wind speed, as well as more complex derived variables related to extreme fire behavior including lapse rates, instability metrics, and other variables related to pyroCb development. Both NAVGEM and GFS carry systematic errors that further complicate the training process and affect model performance. We investigate the influence of the choice between NAVGEM and GFS variables since embedded model errors could cause significantly different modeling and feature ranking results. Emphasis is placed on developing a pipeline that integrates input data sources to assemble a training dataset, applies feature selection and data balancing techniques, and is trained with several ML algorithms. Finally, we perform initial eXplainable AI (XAI) experiments to analyze the strategies learned by the model. XAI results were also analyzed to determine which global models provide more appropriate input features to differentiate the two events, how systematic errors from NAVGEM and GFS influence ML performance, and if a combination of more complex expert-derived variables add discriminatory information. Our main goal is to investigate if the model’s learned relationships align with those known by forecasters to be associated with pyroCb development. An overview of the ML framework and a comparative assessment of all feature combinations and testing scenarios will be presented.

