A contextual approach to the mapping of burned areas in tropical Australian savannas using medium-resolution satellite data

Abstract

Our objective was to develop and test a spectral‐spatial method to detect burned areas in savanna landscapes. The hypothesis tested was that the inclusion of spatial information (context) would better detect burning than do spectral-only tests. The spectral‐spatial measure used was the standard deviation (SD) of a near-infrared (NIR) channel for a 5 × 5 pixels window because this measure exhibits a spike on burning. This measure was compared with spectral-only tests in the Australian tropical savannas during the 2000 dry season (April‐November) using Système pour l'Observation de la Terre (SPOT) VEGETATION (VGT) data. Using a population of 23 burned areas, the superiority of the SD NIR signal was demonstrated and, using videography, the influences of surface factors on the SD NIR signal were examined. A large area (29 000 km2) comparison was made of the several burning detection tests, all referred to two manually interpreted datasets. Using the criterion of user accuracy, two spatial‐spectral tests outperformed two spectral-only tests. However, the largest common class of error was omission, indicating a failure of all tests to completely capture burned areas. This deficiency is not crippling, for it can be largely overcome by spatial filtering.