New interface

# Quantification of wildland fire risk using metamodeling of fire spread

1 ANGE - Numerical Analysis, Geophysics and Ecology
Inria de Paris, LJLL (UMR_7598) - Laboratoire Jacques-Louis Lions
Abstract : This work addresses the quantification of wildfire risk by relying on simulations of fire spread. The objectives are to compute the probability distribution of burned surfaces that could result from wildfire ignition and quickly generate maps to assess which areas should receive focused protection against wildfires. This probability distribution should represent the uncertainty in the simulations. First, an ensemble of wildland fire spread simulations accounting for sources of uncertainty is generated following a Monte Carlo approach, and probabilistic evaluation of the predictions with observations is carried out. Then, the underlying probability distributions are calibrated based on the observations by adapting the Wasserstein distance to the comparison of burned surfaces to improve prediction performance in presence of uncertainty. Subsequently, a deep learning approach is followed to train a hybrid'' neural network with a convolutional part, thus building an emulator of potential'' fire size simulated by the fire spread model allowing to considerably reduce the computational time implied by the large amount of simulations required for high-resolution maps. Eventually, this emulator is applied to derive fire danger mapping from daily weather forecasts and applied to assess relatively large fire events.
Keywords :
Document type :
Theses
Domain :

https://hal.inria.fr/tel-03385307
Contributor : ABES STAR :  Contact
Submitted on : Wednesday, November 17, 2021 - 3:58:43 PM
Last modification on : Friday, August 5, 2022 - 12:02:04 PM

### File

ALLAIRE_Frederic_2021.pdf
Version validated by the jury (STAR)

### Identifiers

• HAL Id : tel-03385307, version 2

### Citation

Frederic Allaire. Quantification of wildland fire risk using metamodeling of fire spread. Silviculture, forestry. Sorbonne Université, 2021. English. ⟨NNT : 2021SORUS037⟩. ⟨tel-03385307v2⟩

Record views