Goal: We plan to design stochastic simulations, that could be used individually for any fire case, and to estimate their economic and environmental costs. Also, we want to incorporate the firefighting actions so that this work provides the bases of a decision-making platform.
Inria (Vivien, PhD) SPE (Filippi), LISA (All), CECI (All) T18 to 42
After a fire ignites, we plan to launch ensemble simulations to evaluate its potential evolution. The ignition point is uncertain, the meteorological data (in situ or from stations in the vicinity) are uncertain and may be changing quickly. The vegetation composition and state is surely uncertain. The actions of the firefighters may be reported with various degrees of accuracy. Out of all these uncertainty sources, we plan to identify the key ones with a screening sensitivity analysis (e.g., with Morris method). The most sensitive parameters will then be perturbed with carefully chosen distributions (using meteorological ensembles, expert knowledge and Bayesian calibration). The propagation model itself is a significant source of uncertainties. We will obviously use ForeFire which is modular C++ front-propagation software and allows one to plug almost any rate of spread (like Rothermel's). The strategy is to generate a multimodel ensemble with perturbations in the inputs. We will evaluate the skills of an ensemble of simulations to forecast the event: “the given location will be burned by the given time”. We will make use of indicators like the Brier score or reliability diagram. The ensemble should then be calibrated (with respect to the parameters of the inputs distributions) to obtain the best scores. The calibration must include the observations from many fires, but we will evaluate how well we can expect the ensemble to represent a single fire. The economic and environmental costs of the fire will be computed using results from Task 3.2.
2.1.1 (T42) Paper on the simulation of a few cases with documented firefighting 2.1.2 (T36) Software for multimodel and Monte Carlo simulations
Resources: PhD shared with Tasks 3.1 and 3.2. Computation on Inria computers. Risk: In a first step towards such simulations, during IDEA project, the results suggested that using several models would significantly improve the uncertainty quantification. We therefore plan to advance in this direction. Optimizing the parameters of the input distributions is a costly operation, which might be replaced with expert knowledge in cases the optimization might fail.
Dans la description du combustible à l'échelle de la Corse via les indices du Corine Land Cover (CLC), des erreurs de classification sont possibles. Par exemple, il a été constaté sur l'incendie de Suare (commune de Calenzana) du 5 août 2017 que des vignobles (221) étaient présents, alors qu'au moment de l'incendie, la végétation était différente. On peut supposer que les principales causes des erreurs de classification sont la fréquence de mise à jour des données et la difficulté d'attribution une classe pour certaines zones.
Dans le cas de la génération d'un ensemble de simulations, on propose d'attribuer une probabilité de transition d'un type n_1 à un type n_2. Par exemple, pour l'incendie de Suare, les zones de vignobles conduisaient à une propagation du feu bien plus lente que celle observée. Le fait d'attribuer au type 221 (vignobles) les propriétés du type 323 (végétation sclérophylle) pour la majorité des simulations donne lieu à un ensemble de prévision paraissant plus pertinent. En l'occurrence, ceci est effectué par le biais d'une probabilité de transition de 0.8 pour passer du type 221 au type 323. Ainsi, environ 80% des simulations de l'ensemble ont lieu “comme si les vignobles étaient de la végétation sclérophylle”.
La difficulté consiste à déterminer de “bonnes” probabilités de transition. Pour les zones où la propagation est impossible, on peut considérer une probabilité de transition nulle. Dans un premier temps, il serait intéressant d'identifier les zones où la vitesse de propagation est généralement faible et qui peuvent être mal classées. C'est le cas des vignobles (type 221). On peut également supposer que ce soit le cas des types pour lesquels le paramètre de hauteur par défaut est faible. On a donc, en plus des vignobles :
Pour le moment, on considère seulement 2 probabilités de transition suivantes :
Indépendemment du système de classification (CLC ici), il pourra être intéressant d'avoir des informations sur les erreurs/difficultés de classification si l'on souhaite attribuer des probabilités de transition.
Goal: Implement the assimilation of fire front location derived from airborne and/or spaceborne remote sensing imagery in ForeFire tool using an ensemble Kalman filter strategy.
CECI (Mélanie, all) SPE (Filippi) T12 to 36
This task consists in implementing in case of verified alert, the assimilation of geolocated front location in ForeFire with a computationally-efficient ensemble Kalman filter (EnKF) algorithm that remains compatible with operational framework. This algorithm is based on the generation of a multimodel ensemble of fire front positions accounting for the potential evolution of the burning area given the sources of uncertainties as addressed in Task 2.1. For this approach, both the environmental conditions (vegetation, near-surface wind) and the instantaneous position of the fire will be corrected (dual approach) to reduce the distance to the observed fire fronts. The EnKF correction will then be used to define the new reference fire front for further integration of the deterministic ForeFire/Meso-NH system. The key idea is to sequentially update the reference fire front in order to provide with improved accuracy, forecast of the burning area (Task 2.1) and the smoke plume (Task 2.3).
In this task, the main objective is to adapt the data assimilation prototype (http://firefly.cerfacs.fr/) developed in the frame of the ANR-IDEA and NSF-WIFIRE projects (Rochoux et al. 2013, 2014, 2015; Rochoux 2014; Zhang et al. 2016) to ForeFire (the evaluation was so far limited to controlled fires). We will implement the EnKF algorithm using the OpenPALM coupling software (CERFACS/ONERA) to efficiently integrate the ensemble members in parallel. If necessary (i.e. if model nonlinearities are strong), we will revisit the EnKF algorithm by implementing a more advanced iterative version (Bocquet and Sakov 2013). We will also investigate different strategies to reduce the number of estimation targets; we may use a metamodel in place of the fire spread model in the EnKF algorithm to save computational time (Rochoux et al. 2014). We will validate the EnKF algorithm on synthetic test cases and then on regional-scale fires. In hindcast mode, the Maido fire (La Réunion, 2011) has already been identified as an interesting test case for the evaluation of the EnKF algorithm (RapidEye-based fire fronts will be provided by SERTIT). Observations on more fire hazards will be available through aerial infrared imagery (Corsican firefighting service) and spaceborne fire products. The choice of the spaceborne products will be chosen according to their spatial and temporal resolutions through discussions with SERTIT (Hervé Yésou - external collaborator) and CNES Risk service (Claire Tinel - member of the steering committee). Products by Pléiades (example of Perthus fires, July 2012), RapidEye and the new Sentinel-3 satellite have been so far identified as possible candidates for this purpose. The complementary between airborne and spaceborne products will be analyzed. We will also investigate how the EnKF algorithm could be adapted to the Meso-NH/ForeFire coupled system. This requires a significant effort (the atmospheric state shall be corrected consistently with the fire position correction through the reconstruction of the artificial atmospheric history), but this does not constitute a barrier to the task success. Data assimilation results on ForeFire will already be valuable for the evaluation of the operational platform, in terms of computational cost, assimilation frequency and forecast accuracy.
2.2.1 (T30) Data assimilation prototype with ForeFire using remote data. 2.2.2 (T36) Technical report and publication (evaluation on a few cases). 2.2.3 (T36) Collection of post-processed appropriate data (at fire resolution).
Resources: One-year PDoc will be devoted to the task. Internal high performance computational resources at CECI will be used. PDOC fire
Risk: If no data source proves to be in useable state at high resolution, real-time MODIS/VIIRS fire detection products will be used as fire perimeter data.
Goal: Optimize the MesoNH system to better represent fire convection and smoke transport.
CNRM (Christine, all) LA/SPE (Filippi, Mari) T12 to 36
Based on the work already performed in the IDEA project, a compressible version of Meso-NH instead of the anelastic approximation will be implemented to account for the strong variations of the air density in the vicinity of the fire. This enhancement will allow to represent better fire-induced gusts and strong convection that are responsible of extreme fire behaviour, in particular the emission and transport of embers or fire jumps. Moreover, it is expected that it will help to provide a more realistic flow near the smoke injection areas in and around the fire front. Finally, a tool chain will be created that will use national analysis and create the required nesting configuration for Meso-NH/Fire to run from 2-km resolution to less than 100-m resolution (see figure). The ForeFire model will also be enhanced for these deterministic runs to account for fire-jumps and crowning.
Nested simulation and domain decomposition of the Aullene fire and corresponding plume as observed by MODIS satellites.
A reduced atmospheric smoke aerosol code will be tested for on-demand runs to decrease the computational time (e.g. by simulating only aerosol mass concentration instead of the whole chemical composition of aerosols). The emissions models will be parameterised as a function of the fire regime (e.g. flaming or smoldering stages) and fuel type (humidity, type of vegetation, …). The coupled model will allow to assess air pollution close and downwind of the fire. Sensitivity studies of smoke emission models will be performed and validated on field data from Task 1.3.
2.3 (T36) Deterministic coupled code available.
Resources: A one-year PDoc will implement the developments in MesoNH. Simulations will be performed at GENCI/CINES and on SPE supercomputer. PDOC Weather Risk: Compressible version of Meso-NH too computationally intensive for operational use. No measurement available from the prescribed field experiment to constrain emission profiles. Fallback solution: Runs will be evaluated at coarser resolution. Emission profiles will be derived from previous field campaigns (for different and similar vegetation type if possible).