Remotely Sensed Fire Type Classification of the Brazilian Tropical Moist Forest Biome

Vegetation fires in the Brazilian Tropical Moist Forest Biome can be broadly classified into three types: i) Deforestation fires, lit to aid deforestation by burning of slashed, piled and dried forest biomass, ii) Maintenance fires, lit on agricultural fields or pasture areas to maintain and clear woody material and to rehabilitate degraded pasture areas, iii) Forest fires, associated with escaped anthropogenic fires or, less frequently, caused by lightning. Information on the incidence and spatial distribution of fire types is important as they have widely varying atmospheric emissions and ecological impacts. Satellite remote sensing offers a practical means of monitoring fires over areas as extensive as the Brazilian Tropical Moist Forest Biome which spans almost 4 million square kilometers. To date, fire type has been inferred based on the geographic context and proximity of satellite active fire detections relative to thematic land cover classes, roads, and forest edges, or by empirical consideration of the active fire detection frequency. In this paper a classification methodology is presented that demonstrates a way to classify the fire type of MODerate Resolution Imaging Spectroradiometer (MODIS) active fire detections. Training and validation fire type data are defined conservatively for MODIS active fire detections using a land cover transition matrix that labels MODIS active fires by consideration of the PRODES 120m land cover for the previous year and the year of fire detection. The training data are used with a random forest classifier and remotely sensed predictor variables including the number of MODIS Aqua and Terra satellite detections, the maximum and median Fire Radiative Power (FRP) [MW km^-2], the scaling parameter of the FRP power law distribution, the number of day and night detections, and the fire surrounding "background" surface brightness temperature [K]. In addition, the total rainfall over periods from 1 to 24 months prior to fire detection and the fire detection proximity to official and unofficial roads and navigable rivers are included as predictor variables. Results are illustrated for eight years (2003-2010) of MODIS active fire detections with a cross validation showing greater than 70% fire type classification accuracy. The spatio-temporal distribution of fire types across the Brazilian Tropical Moist Forest Biome are presented with higher incidences of deforestation fires in the "arc of deforestation" and similar proportions of forest and maintenance fires for all years except for 2007 and 2010 that exhibited a relatively higher proportion of forest fires.