Drivers of Global Wildfires
— Statistical analyses
Master Thesis Seminar, 2010
Hongxiao Jin
Supervisor: Dr. Veiko Lehsten
Division of Physical Geography and Ecosystems AnalysisDepartment of Earth and Ecosystem Sciences
Lund University
2010.02.16Photo: www.nmcounties.org
Presentation Outline
• Introduction
• Data
• Methods
• Results
• Conclusions
Master Thesis Seminar , 2010
Introduction
Master Thesis Seminar , 2010
Impacts of wildfires in earth systemaffect ecosystemaccelerate carbon cycleinfluence climate
Wildfire driversscale-dependent
local scalelandscape scaleregional and global scale
Fire representations DGVMsDGVMs should include fire modelsbased on statistical relationsburned area
Master Thesis Seminar , 2010
Data
Fire dataCollection 5 MODIS Level 3 Monthly Tiled 500m Burned Area Product• April 2000 to March 2009• lon -180º to +180º, lat 53.22ºS to 75.55ºN• 500 m resolution• sinusoidal projection• 266 tiles• nominal size 10º10º• 2400 rows 2400 columns• tiles are overlapped• hierarchical data format (*.hdf) • 24,408 files Regridded to 0.25º0.25º to get burned area ratio (BAR) and burn date
Master Thesis Seminar , 2010
Data
Driver data
• Precipitation TRMM+NCEP
• Surface air temperature NCEP • Forest cover IIASA
• Grass cover IIASA
• Cultivation IIASA
• Urban IIASA
• Soil nutrient availability IIASA
• Population density CIESIN • Topographical roughness USGS
• Wind speed NCEP
• Air relative humidity NCEP • Soil moisture NOAA
Master Thesis Seminar , 2010
Methods
Correlation analyses
Response variable• Nine years’ mean annual BAR• Individual annual case (2004)
Explanatory variable• 13 variables for nine years’ mean annual BAR• 17 variables for individual annual case (2004)
Linear correlation• Pearson correlation• Generalized linear correlation
Master Thesis Seminar , 2010
Methods
Modelling
• Generalized linear model 14 regions and global binomial distribution logit link function linear combination of explanatory variables 1st order, 2nd order and 2-times interaction stepAIC of R 50% of samples as training (except for TENA &CEAM), world data 5% of samples as training
• Random forest regression global 500 trees 5 out of 13 variables 5% of samples as training
Master Thesis Seminar , 2010
Results and discussions
Burned area
Master Thesis Seminar , 2010
Fire seasons
Results and discussions
Master Thesis Seminar , 2010
Results and discussions
Fire seasons and peak fire month of 14 regions. Red numbers indicate the peak fire month of each region. Winter-spring fires happened between 23.5ºN and 23.5ºS and summer-autumn fires happened outside this zone. Four red rectangles had summer-autumn fire seasons, different from their main regions of the regional division scheme.
Master Thesis Seminar , 2010
Drivers
Pearson correlation coefficient
Generalized liner correlation
From the most to the least important:MeanT, IntraR, Grass, RainNoFire, InterR, RainFireSeason, Nutrient, MeanR, Topography, Forest, Population, Urban, Cultivation
Results and discussions
Master Thesis Seminar , 2010
Variable importance estimated by random forest regression. The variable importance is give by the measure of the mean squared error increasing percent (% Increase of Mean Squared Error) when that variable is permuted.
Variable importance given by randomForest
Population ?Urban ?Cultivation?
Results and discussions
Master Thesis Seminar , 2010
Models
Observed vs. Modelled
Results and discussions
Master Thesis Seminar , 2010
Observed
Global RF
Global GLM
Modelled vs. Observed
Regional GLM
Results and discussions
Master Thesis Seminar , 2010
Conclusions Burned area
3.85% of the global land area burned each year (335.74±9.18 104km2 ). Savanna fires account for 83.1% and Africa contributes 72% of the world total.
Fire seasonsGlobally 4 latitudinal zones, 2 fire seasons (August and December)
DriversMean annual temperature is the most important driver of global wildfire. The next most important driver is grass cover. Each region has slightly different sequences of wildfire drivers.
ModelsThe regional GLMs have better prediction performance than global GLM and random forest. The global random forest regression is superior to the global GLM.
Master Thesis Seminar , 2010
Thanks!