A statistical methodology for burned area estimation using multitemporal AVHRR data

A statistical methodology based on multitemporal NOAA Advanced Very High Resolution Radiometer (AVHRR) data, for identification of burned areas has been developed and tested for an area covering the Niokolo Koba National Park in the Sudanian zone of south-east Senegal. Ten NOAA AVHRR images, acquired to coincide with a field campaign, were employed in the analysis. A SPOT HRV scene, also coinciding with the field campaign, was furthermore applied and used as validation of the developed methodology. Burned areas were identified as pixels in the AVHRR data that underwent a rapid change in time in a three-dimensional feature space defined by the Normalized Difference Vegetation Index (NDVI), surface temperature and albedo. To circumvent fluctuations in the AVHRR data due to atmospheric contamination as well as the general senescence of vegetation occurring at the start of the dry season, principal component transformations of the AVHRR data were used. Following this initial identification, all identified pixels were checked against the preceding image and were confirmed as a burned area if they exhibited relatively low levels of NDVI, high surface temperature and a low albedo. All results of the method were validated against the classified SPOT HRV image and the accuracy of the method, expressed as agreement between the applied method and the classified high resolution image, was evaluated. It was found that the accuracy of burned areas is a function of the fraction of the pixel affected by fire and that for pixels burned from 75% to 100% an accuracy of 80% is found, which decreases as the fraction decreases.