ABSTRACT Kroumiria Mountains (northwestern Tunisia) have experienced major fires, making them the main loss reason of Tunisian forested areas. The ability of accurately forecasting or modeling forest fire areas may significantly aid optimizing fire-fighting strategies. However, there are still limitations in the empirical study of forest fire loss estimation because the poor availability and low quality of fire data. In this study, a stochastic approach based on Markov process was developed for the prediction of burned areas, using available meteorological data sets and GIS layers related to the forest under analysis. The Self-organizing map (SOM) was initially used to classify spatiotemporal factors influencing the fire behavior. Subsequently, the SOM clusters were incorporated into a Hidden Markov Model (HMM) framework to model their corresponding burned areas. Results achieved using a database of 829 forest fires records between 1985 and 2016, showed the appropriateness of the HMM approach for the prediction of burned areas compared with a state-of-the art machine learning methods. The transition probability matrix (TPM) and the emission probability matrix (EPM) were also analyzed to further understand the spatiotemporal patterns of fire losses.
Read full abstract