Stochastic near-surface wind field estimation from sparse aerial swarm measurements to support wildfire behaviour predictions

Wildfire behaviour is influenced by weather, fuel, and topography. Detailed physics-based simulations are computationally expensive, particularly if accounting for fire–wind interactions. To address this, simplified models are developed that disregard certain physical laws and are often decoupled from the atmosphere. Efforts to improve model reliability by feeding it with near-surface wind fields using statistical and dynamical downscaling methods have two drawbacks: neglect of fire–wind interaction, and high computational demand. This study proposes combining sparse wind and temperature measurements taken by a drone swarm with convolutional neural networks to estimate near-surface fire-induced wind fields thereby capturing fire–wind interactions without solving the Navier-Stokes equations. The two-step process includes (i) super-resolution reconstruction of high-altitude wind field from sparse measurements, and (ii) high-resolution mapping from high-altitude to near-surface winds. Synthetic data are generated using Large-Eddy Simulations of grassland fires under varied wind speeds, terrain slopes, and vegetation conditions. Model performance is evaluated across different swarm sizes, spatial configurations, and measurement noise levels. The main contributions are: (1) a synthetic computational fluid dynamics dataset for fire-induced flows, (2) a two-stage deep learning framework for stochastic wind field estimation, and (3) uncertainty quantification for different swarm and noise conditions. Results show accurate near-surface wind estimation even with moderate swarm sizes, achieving an average mean absolute error and root mean square error (MAE/RMSE) of 0.849/1.323 for u, 0.672/1.022 for v, and 0.551/1.010 for w, where the (u,v,w) tuple defines the wind velocity.