Application of Bayesian networks for fire risk mapping using GIS and remote sensing data

This study estimates fire risk in Swaziland using geographic information system (GIS) and remote sensing data. Fire locations were identified in the study area from remotely sensed Moderate Resolution Imaging Spectroradiometer (MODIS) active fire and burned area data for the period between April 2000 to December 2008 and January 2001 and December 2008, respectively. A total of thirteen biophysical and socio-economic explanatory variables were analyzed and processed using a Bayesian network (BN) and GIS to generate the fire risk maps. The interdependence of each of the factors was probabilistically determined using the expectation-maximization (EM) learning algorithm. The final probabilistic outputs were then used to classify the country into five fire risk zones for mitigation and management. Accuracy assessments and comparison of the fire risk maps indicate that the risk maps derived from the active fire and burned area data were 93.14 and 96.64% accurate, respectively, demonstrating sufficient agreement between the risk maps and the existing data. High fire risk areas are observed in the Highveld particularly plantation forests and grasslands and within the Lowveld sugarcane plantations. Land tenure and land cover are the dominant determinants of fire risk, the implications of which are discussed for fire management in Swaziland. Limitations of the data used and the modeling approach are also discussed including suggestions for improvements and future research.