An evaluation of di V erent bi - spectral spaces for discriminating burned shrub - savannah

We report on the numerical separation of burned and unburned vegetation classes using diVerent bi-spectral spaces, based on the analysis of spectro-radiometric data collected in situ and convolved to ve spectral bands at red to mid-infrared (MIR) wavelengths. A combination of two MIR bands was found to have strong spectral separation of burned and unburned samples. Using these bands, a spectral index was formulated which is highly sensitive to spectral changes due to burning and relatively insensitive to intrinsic variability. Results have implications for the remote sensing of burned shrub-savannah using bands available on high-and low-spatial resolution sensors, in particular, Landsat TM and MODIS. 1. Introduction Due to the substantial impacts of biomass burning, one area of remote sensing research is focused on developing algorithms to detect areas burned annually at a global scale (Pereira et al. 1997). Reliable remote detection of burned vegetation requires the use of spectral bands that are sensitive to signi cant changes in radiance in response to burning, henceforth called burn signals (Trigg and Flasse 2000), which can be isolated form spectral changes from other sources, henceforth called perturbing factors. More simply, the spectral bands used should provide a suYcient ratio of signal (due to burning) to noise (from perturbing factors) to allow reliable detection of burned areas. Perturbing factors include the geometry of illumination and observation, variable atmospheric conditions and intrinsic spectral variation at the surface, such as: