Prediction of the radioactive potential of fires: comparing two techniques using spatial regression

Abstract

With climate change, studies that consider elements to manage damages caused by fires in the vicinity or inside municipalities, through the radioactive potential of fire, may be useful. In this context, this work presents an approach to predict, from geographically weighted regression, the incidence of the radioactive potential of fires. For this, data extracted from the NASA’s FIRMS project and from the INMET (National Institute of Meteorology) have been acquired, cleaned, and some topographic variables have been additionally derived. The work considers the city of João Pessoa (in the northeast of Brazil) as an object of study and makes use of machine learning methods based on spatial linear regression to predict the radioactive potential of fires. To this end, it considers two scenarios whose differences lie in the extension of the samples and in the choice of the independent variables. The first scenario covers samples which are more concentrated in the city of João Pessoa, and includes climatic, satellite and topographical variables; the second scenario covers the entire mesoregion of the forest zone in Paraíba close to João Pessoa, but does not consider the climatic variables. The work has achieved promising results.