Geostationary-Satellite-based Wildfire Smoke Transport Nowcasting using Optical-Flow-GRU Coupled Scheme

Wildfire smoke, containing noxious gases and particulate matter, leads to severe effects on social activities and human health. Given increasing wildfire activities under global climate change scenario, timely and accurate forecasting of smoke transport is crucial for mitigating health risks and potential property losses. Current operational smoke prediction systems are typically subject to inherent uncertainties due to model assumptions and simplifications associated with fuel conditions, emission species, plume rise height, and so forth. In this study, we conduct wildfire smoke transport nowcasting entirely based on geostationary satellite observations, through developing a two-stage model coupling smoke motion velocity field estimation and prediction. Specifically, the wildfire hotspots are first located and tracked via GOES-17/18 fire products, and corresponding cloud-free blue-band reflectance image sequences centered on the fire spots are obtained. The horizontal motion of smoke is characterized by 2-D velocity fields, estimated through the aerosol dynamic optical flow method based on temporally continuous image sequences. These known velocity fields act as a training data for the Gated Recurrent Unit neural network which is recursively updated when the latest observations are added, while the future velocity fields are predicted based on trained model and the predictive location of smoke can be accordingly provided. As GOES-derived data is available within minutes of observation, the proposed near-real-time smoke transport nowcasting scheme offers an effective approach for alerting the public in regions impending wildfire smoke impacts 1-2 hours in advance.