MODELLING FOREST FIRE OCCURRENCE IN LEBANON USING SOCIO-ECONOMIC AND BIOPHYSICAL VARIABLES IN OBJECT-BASED IMAGE ANALYSIS

Landcover and land use changes driven by socio-economic changes have increased the probability of occurrence and spread, especially in the Wildland-Urban Interface. The study aimed to model the influence of both socio-economic and biophysical variables on fire occurrence in Lebanon. Forward stepwise binary logistic regression analysis of 24 socio-economic and biophysical variables was used to predict wildfire occurrence. Spearman correlation analysis was conducted to eliminate multi-collinearity between selected variables. The study area covers the country of Lebanon, divided into four distinct physiographic regions: the coastal plain, the Lebanon mountain range, the Beqaa Valley, and the Anti-Lebanon mountains. Forests and other wooded land cover 24.5% of the Lebanese territory. The climate in Lebanon is characterized by dry summers extending from June to November, with average daytime temperatures above 30°C, and little rain. Approximately 89% of the Wildland-Urban Interface in Lebanon occurs in areas classified between moderate to very high risk of fire spread. Biophysical data included slope, aspect, and vegetation fuel type. The Prometheus fuel type classification system was used for characterizing fuel type. Spatial distribution of fruit trees, mixed culture, field crop, olive trees, vineyard, and greenhouses were extracted from the landcover/land use map of Lebanon produced in 1998. Conversion of lands between 1998 and 2005 were also acquired. Animal production was obtained from the homogeneous agricultural zones data of the Ministry of Agriculture produced in 1998. In addition, population density, distance to big cities, distance to bigger cities, the agricultural-urban interface, and roads in agricultural lands were also acquired. The study employed object-based image analysis to derive a spatially explicit probability of fire occurrence across the country. The final map showed 5 different fire danger classes ranging from very low to very high. The quality of the classification results was evaluated, and under- and overestimations errors of fire occurrence were mapped. The accuracy of the fire occurrence mapping model was approximately 85% when tested on the validation data set. The probabilistic spatial output of the fire threat model was considered satisfactory given the challenges of using multi-source data in an object-based image analysis approach. Results suggest increasing the resolution of socio-economic data would improve modelling accuracy of fire occurrence in Lebanon.