Using Score-based Data Assimilation to Improve Forecast Initialization for Fire Weather Behavior

Wildfires are an imminent threat in the western united states, with wildfire activity becoming more unpredictable as climate change drives major shifts to the dynamics of forest ecosystems. For example, the rapidly spreading Camp Fire in Paradise, California in 2018 caused 85 deaths and destroyed more than 18,000 structures resulting in claims of over $12.5 billion dollars. The high cost and need for rapidly deployable and accurate forecasts have been demonstrated not only by the Camp Fire, but many other fires in the western United States and Canada over the past decade. Additionally, the use of GPU accelerated computation for machine learning based forecasting algorithms has paved the way for a new set of tools to be utilized in for studying fire weather behavior in complex and difficult to access environments, with many locations having few or no meteorological stations present. In this study, the model is initially trained on High-Resolution Rapid Refresh (HRRR) analysis wind data for the state of Montana for 2017-2023. Observational data will be taken using a combination of in-situ sites, radiosonde launches, and uninhabited aerial system (UAS) flights over the Missoula test site. After the model training on HRRR wind field data is complete, the inferencing stage is performed using any one or two in-situ station/radiosonde/UAS to generate a dense state estimation. The dense state is then compared to the unused data source for validation, as well as the HRRR. For example, if in-situ and radiosonde data are used in the inferencing step, the dense state estimation will then be compared to the UAS and HRRR data, providing insight to fire weather observations in a high-resolution grid that may be difficult or otherwise impossible to collect in real time. The Score-based Data Assimilation (SDA) method is shown to help increase the skill for the initial state of a wind forecast, in a region with topographic heterogeneity and prone to rapidly changing weather conditions. This information can then be used to enhance fire weather forecasts, to provide more accurate information for evacuations, protection of assets, and plan for wildfire mitigation strategies.