Student Number : 0310612G - MSc research report - School of Geography, Archaeology and Environmental Studies - Faculty of Science / Wildfire prediction and management is an issue of safety and security for many rural communities in South Africa. Wildfire prediction and early warning systems can assist in saving lives, infrastructure and valuable resources in these communities. Timely and accurate data are required for accurate wildfire prediction on both weather conditions and the availability of fuels (vegetation) for wildfires. Wildfires take place in large remote areas in which land use practices and alterations to land cover cannot easily be modelled. Remote sensing offers the opportunity to monitor the extent and changes of land use practices and land cover in these areas. In order for effective fire prediction and management, data on the quantity and state of fuels is required. Traditional methods for detecting vegetation rely on the chlorophyll content and moisture of vegetation for vegetation mapping techniques. Fuels that burn in wildfires are however predominantly dry, and by implication are low in chlorophyll and moisture contents. As a result, these fuels cannot be detected using traditional indices. Other model based methods for determining above ground vegetation biomass using satellite data have been devised. These however require ancillary data, which are unavailable in many rural areas in South Africa. A method is therefore required for the detection and quantification of dry fuels that pose a fire risk. ASTER and MAS (MODIS Airborne Simulator) imagery were obtained for a study area within the Lowveld region of the Limpopo Province, South Africa. Two of the ASTER and two of the MAS images were dated towards the end of the dry season (winter) when the quantity of fuel (dry vegetation) is at its highest. The remaining ASTER image was obtained during the middle of the wet season (summer), against which the results could be tested. In situ measurements of above ground biomass were obtained from a large number of collection points within the image footprints. Normalised Difference Vegetation Index and Transformed Vegetation Index vegetation indices were calculated and tested against the above ground biomass for the dry and wet season images. Spectral response signatures of dry vegetation were evaluated to select wavelengths, which may be effective at detecting dry vegetation as opposed to green vegetation. Ratios were calculated using the respective bandwidths of the ASTER and MAS sensors and tested against above ground biomass to detect dry vegetation. The findings of this study are that it is not feasible, using ASTER and MAS remote sensing data, to estimate brown and green vegetation biomass for wildfire prediction purposes using the datasets and research methodology applied in this study. Correlations between traditional vegetation indices and above ground biomass were weak. Visual trends were noted, however no conclusive evidence could be established from this relationship. The dry vegetation ratios indicated a weak correlation between the values. The removal of background noise, in particular soil reflectance, may result in more effective detection of dry vegetation. Time series analysis of the green vegetation indices might prove a more effective predictor of biomass fuel loads. The issues preventing the frequent and quick transmission of the large data sets required are being solved with the improvements in internet connectivity to many remote areas and will probably be a more viable path to solving this problem in the near future.

Les incendies de forêt sont l'un des phénomènes les plus complexes auxquels sont confrontées nos sociétés. Le Liban, faisant partie du Moyen-Orient, est en train de perdre dramatiquement ses forêts vertes principalement en raison de graves incendies. Cette thèse étudie le phénomène des incendies de forêt. Elle propose des nouveaux modèles et méthodologies pour remédier à la crise des incendies de forêts, en particulier au Liban et en Méditerranée. Elle est divisée en deux parties principales: nouvelles approches de la prévision des incendies de forêt et développement d'un nouveau modèle de diffusion du feu plus fidèle du cas réel. La première partie est subdivisée en 3 chapitres. Le premier chapitre présente une étude analytique des modèles métrologiques les plus utilisés qui permettent de prédire les incendies de forêt. Dans le deuxième chapitre, nous appliquons cinq méthodes de techniques d’exploration de données: Réseaux de neurones, arbre de décision, floue logique, analyse discriminante linéaire et méthode SVM. Nous cherchons à trouver la technique la plus précise pour la prévision des incendies de forêt. Dans le troisième chapitre, nous utilisons différentes techniques d'analyse de données corrélatives (Régression, Pearson, Spearman et Kendall-tau) pour évaluer la corrélation entre l'occurrence d'incendie et les données météorologiques (température, point de rosée, température du sol, humidité, précipitation et vitesse du vent). Cela permet de trouver les paramètres les plus influents qui influencent l'occurrence de l’incendie, ce qui nous amène à développer un nouveau Indice Libanais de Risques d'Incendie (IL). L'indice proposé est ensuite validé à partir des données météorologiques pour les années 2015-2016. La deuxième partie est subdivisée en 3 chapitres. Le premier chapitre passe en revue les caractéristiques du comportement de feu et sa morphologie; il se concentre sur la validité des modèles mathématique et informatique de comportement de feu. Le deuxième chapitre montre l'importance des automates cellulaires, en expliquant les principaux types et examine certaines applications dans différents domaines. Dans le troisième chapitre, nous utilisons des automates cellulaires pour élaborer un nouveau modèle de comportement pour prédire la propagation de l’incendie, sur des bases elliptiques, dans des paysages homogènes et hétérogènes. La méthodologie proposée intègre les paramètres de la vitesse du vent, du carburant et de la topographie. Notre modèle développé est ensuite utilisé pour simuler les incendies de forêt qui ont balayé la forêt du village d'Aandqet, au nord du Liban. Les résultats de simulation obtenus sont comparés avec les résultats rapportés de l'incident réel et avec des simulations qu'on a iv effectuées sur le modèle de Karafyllidis et le modèle de Karafyllidis modifié par Gazmeh. Ces comparaisons ont prouvé l'ambiguë du modèle proposé. Dans cette thèse, la crise des feux de forêt a été étudiée et de nouveaux modèles ont été développés dans les deux phases: pré-feu et post-feu. Ces modèles peuvent être utilisés comme outils préventifs efficaces dans la gestion des incendies de forêt . / Wildland fires are one of the most complex phenomena facing our societies. Lebanon, a part of Middle East, is losing its green forests dramatically mainly due to severe fires. This dissertation studies the phenomenon of forest fires. It proposes new models and methodologies to tackle the crisis of forest fires particularly in Lebanon and Mediterranean. It is divided into two main parts: New Approaches in Forest Fire Prediction and Forest Fire modeling. The first part is sub-divided into 3 chapters. First chapter presents an analytical study of the most widely used metrological models that can predict forest fires. In the second chapter we apply five data mining techniques methods: Neural Networks, Decision Tree, Fuzzy Logic, Linear Discriminate Analysis and Support Vector Machine. We aim to find the most accurate technique in forecasting forest fires. In the third chapter, we use different correlative data analysis techniques (Regression, Pearson, Spearman and Kendall-tau) to evaluate the correlation between fire occurrence and meteorological data (Temperature, Dew point, Soil temperature, Humidity, Precipitation and Wind speed). This allows to find the most influential parameters that affect the occurrence of fire, which lead us to develop a new Lebanese fire danger Index (LI). The proposed index is then validated using meteorological data for the years 2015-2016. The second part is sub-divided into 3 chapters. The first chapter reviews the fire behavior characteristics and its morphology; and focuses on the validity of mathematical and computer fire behavior models. The second chapter manifests the importance of cellular automata, explains the main types of cellular automata and reviews some applications in various domains. In the third chapter, we use cellular automata to develop a new behavior model for predicting the spread of fire, on elliptical basis, in both homogeneous and heterogeneous landscapes .The proposed methodology incorporates the parameters of wind speed, fuel and topography. The developed model is then used to simulate the wildfire that swept through the forest of Aandqet village, North Lebanon. Obtained simulation results are compared with reported results of the real incident and with simulations done on Karafyllidis model and Gazmeh-Modified Karafyllidis model. These comparisons have proven the outperformance of the proposed model. In this dissertation, the crisis of forest fires has been studied and new models have been developed in both phases: pre-fire and post-fire. These models can be used as efficient preventive tools in forest fire management.