Wind Field Reconstruction and Uncertainty Quantification at Wildland Fires Based on Sparse UAV-based Wind Measurements

Wildfire behaviour is highly influenced by weather, fuel and topography, resulting in highly dynamic propagation patterns. However, a detailed physics-based simulation of this dynamic behaviour can be computationally expensive and time-consuming even on small scales, particularly when accounting for fire-wind interactions. To overcome this limitation, a series of models are developed to provide rapid estimations of fire behaviour by simplifying or ignoring certain physical laws–albeit at the cost of accuracy. These models are often decoupled from the atmosphere to reduce the computational demands, which leads to increased uncertainty in their predictions. Additionally, efforts to improve model reliability by incorporating near-surface wind fields into the model using statistical and dynamical downscaling methods face two challenges, including the failure to account for fire-wind interaction and the high computational demand of dynamical methods. Consequently, this study introduces a novel framework that combines UAV-swarm-based wind and temperature measurements with convolutional neural networks (CNN), to estimate the fire-induced near-surface wind field, aiming to capture the fire-wind interaction and its effect on the fire propagation dynamics in a grassland fire without solving the complete set of Navier-Stokes equations. The framework includes a two-step process for wind field estimation, including (i) super-resolution reconstruction of the high-altitude wind field from sparse UAV-based measurements, and (ii) high-resolution estimation of the near- surface wind field based on the reconstructed high-altitude wind field. The estimated wind field could then be fed into decoupled wildfire models to replicate the effect of fire-wind interaction on fire propagation. Given the extensive data requirement of deep learning models and lack of access to real-world measured data, this study utilizes synthetic data generated from executing 150 three- dimensional Large Eddy simulations of wildfire propagation in grasslands with varying wind speeds, terrain slopes, vegetation types, and height. The accuracy and uncertainty levels of the trained models are evaluated for different UAV swarm sizes, ranging from 100 to 9 UAVs, as well as various sampling strategies, focusing on the spatial distribution of UAVs above the field. Additionally, the models’ reliability are tested under different wind measurement errors by UAV-mounted sensors, varying from 0 to 50%. The obtained results indicate that the developed framework is capable of providing accurate estimations from the near-surface wind field, even under scenarios with a limited number of UAVs, demonstrated through average MAE and RMSE values equal to 0.849 and 1.323 for the U, 0.672 and 1.022 for the V, and 0.551 and 1.01 for the W component of velocity. Uncertainty analysis indicates that even though the average performance of the model remains stable, model uncertainty increases with reducing the size of the swarm. Finally, the investigation of the effect of wind measurement errors on model accuracy and reliability indicates that increased noise levels significantly impact the model’s accuracy and uncertainty. However, increasing the swarm size helps to mitigate the effects of measurement noise to a certain extent.