Remote Sensing and Geostatistical Analysis of Anthropogenic Biomass Burning and Forest Degradation in Madagascar

Abstract

Anthropogenic biomass burning is a long-standing concern in Madagascar. Available estimates of annual burned area consist of unsupported but widely cited guesses that 25%-33% of Madagascar burns annually, and unreliable official statistics that put the figure at only about 2.2%. Therefore, an assessment was made of the utility of Defense Meteorological Satellite Program, Operational Linescan System (DMSP-OLS) fire product for prediction of burn area. Burn scars were mapped using Landsat Enhanced Thematic Mapper (ETM) imagery and a coarsened map was compared to a DMSP-derived fire index map. Traditional linear regression indicates low correlation on a per-pixel basis, but the datasets showed extremely skewed distributions and significant auto- and cross-correlation, mandating a geostatistical approach. A linear model of correlation (LMC) shows low correlation between DMSP and burn area at fine spatial scales but increasing correlation at coarser scales. The uncertainty of burn area predictions for polygons of differing area was estimated using sequential Gaussian simulation of burn area maps, using DMSP as a secondary dataset and the variogram model of the true burn area map as a constraint. For a given time period, 100 simulations were generated and processed to produce mean and standard deviation estimates of burn area for polygons in the following GIS layers: the entire country, level 1 provinces, level 2 provinces, a JERS-derived landcover map, a protected forests map, and the outlines of the Landsat scenes used in the study. The true burn area for the calibration and validation ETM scenes was within one standard deviation of the simulationestimated mean burn area for each scene. The annual burn area estimates for the entire island were 7.0 ± 1.1% and 6.0 ± 1.1% for August-December of 1999 and 2000, respectively. Uncertainty relative to the mean increases at finer spatial scales; Madagascar preserves appear to be small enough that annual changes in burn area will not be statistically significant given the large relative uncertainty in the local DMSP-derived burn area estimate. However, as a predictor of burn area at regional (province-level) or national scales, DMSP appears to have potential.