Drivers of Global Wildfires : statistical analyses

Abstract: Wildfires play an important role in the earth system. Understanding the relationship between
wildfires and their drivers is critical for us to predict fire regime transformations under the
changing climate and anthropogenic influences. This study used the novel Moderate
Resolution Imaging Spectroradiometer (MODIS) burned area product MCD45A1 to
characterize wildfires. The high resolution data were regridded at 0.25° × 0.25° cellsize to
calculate the wildfire burned area ratio (BAR) and burn date, which show a new pattern of
global wildfires from April 2000 to March 2009.
Climate, land cover, topography, and various anthropogenic and natural datasets were
explored and gridded into 0.25° resolution. This study then used Pearson correlation and
generalized linear correlation analyses to estimate the relationship between the mean annual
BAR and possible fire drivers, including the mean annual surface air temperature, mean
annual rainfall, grass cover, forest cover, population density, cultivation percentage, urban
cover, nutrient availability, topographical roughness, inter-annual and intra-annual variability
of rainfall, rainfalls in fire season and non-fire season. The analyses were done at both global
and regional scales. Optimal generalized linear models (GLMs) were obtained by automatic
stepwise regression for the globe and each region. The random forest regression was also
carried out to compare the results from GLM analyses.
Among all the explanatory variables, the mean annual temperature has the closest relationship
with the mean annual BAR, and the next most important driver is the grass cover. Each region
has slightly different sequences of wildfire drivers. The regional GLMs have better prediction
performance than the global GLM and the random forest. The global random forest regression
is superior to the global GLM.