Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Doctorat → Allemagne → Bayesian network models for wildfire risk estimation in the Mediterranean basin

Technische Universität München (2015)

Bayesian network models for wildfire risk estimation in the Mediterranean basin

Papakosta, Panagiota

Titre : Bayesian network models for wildfire risk estimation in the Mediterranean basin

Bayes‘sche-Netz Modelle zur Abschätzung des Brandrisikos im Mittelmeerraum

Auteur : Papakosta, Panagiota

Université de soutenance : Technische Universität München

Grade : Doktor-Ingenieurs (Dr.-Ing.) 2015

Résumé
Wildfires are common in geographic areas where the climate is sufficiently moist for vegetation growth but also features extended dry hot periods. Besides climate, human interventions either on purpose or by accident also play an important role in the occurrence of wildfires. Although wildfire incidents have always accompanied vegetation growth, there is an increase in the severity of wildfires during the past three decades with severe impacts on vegetation, animals, crops, human lives and properties. Record temperatures occurring during recent summer periods (Southeast Australia 2009, Russia 2010, South California 2014) lead to extreme wildfire events that were associated with huge socio-economical costs. In addition, scenarios of global warming suggest that wildfires will become more frequent and more intense in the future. The above stress the need for efficient wildfire risk predictive models to support the planning of precautionary, preventive and mitigating measures (e.g. danger communication, evacuation preparedness, dead fuel clearing activities, firefighting infrastructure, property insurance). In order to quantify wildfire risk, the predictive model must include models for fire occurrence, fire behavior and fire effects. Due to the randomness inherent in the wildfire process and because the modeling is subject to uncertainty in all three stages (occurrence, behavior, damages), fire risk prediction is ideally carried out in a probabilistic format. Available data and expert knowledge should be incorporated for parameter learning. Moreover, the applications need to deal with (partly incomplete) data from various sources. The aim of this thesis is to introduce a daily fire risk prediction model in the meso-scale and to produce daily fire risk maps. The modeling is carried out with Bayesian Networks (BN). BN are graphical probabilistic models that can effectively represent complex processes with multiple random variables, their interdependencies and the associated uncertainties. A probabilistic spatio-temporal BN model for fire risk prediction is presented, which predicts daily fire risk on houses and vegetated areas in the meso-scale (1 km² spatial resolution). The BN model consists of three parts : • The fire occurrence model, which involves as predictive variables weather conditions (expressed by the Fire Weather Index - FWI), land cover types, population and road density. It predicts the probability of a fire occurring daily in each 1km² and is based on the results of a Poisson regression analysis • The fire size model, triggered by the occurrence model, which includes the influence of actual and past weather conditions, fire behavior indices and topography. • The damage model, which predicts the expected losses relevant to houses and vegetated areas conditional on fire hazard. Vulnerability (resistance capacity) and exposure (values at risk) indicators are used to quantify the damage, which also depends on the fire suppression efficiency. The final outputs of the model are the expected house damage costs (the risk to houses) and the restoration costs for the vegetation (the risk to vegetated areas). The BN model is exemplarily established and predictions are made for study areas in Greece, Cyprus and France. The conditional probability distributions of the BN variables are populated with data for different time periods, regression model results and expert knowledge. The BN models are coupled with a GIS for both parameter learning and output mapping. Data from 2010 for Cyprus and from 2003 and 2010 for South France are used as verification datasets. The predictions are compared with actual losses for selected fire periods. The results are shown in daily maps with 1 km² spatial resolution.

Présentation

Version intégrale

Page publiée le 26 septembre 2017