An assessment of the accuracy of DOLA’s northern Australian NOAA-AVHRR Fire Affected Area (FAA) map products - appendix 2

A standard method for assessing the accuracy of fire mapping derived from NOAA-AVHRR imagery was applied to five Landsat scenes sampling different regions across northern Australia. The procedure involved the undertaking of: (1) detailed ground verification exercises (using ground and aerial transects) at the time of Landsat overpasses; (2) assessment of the accuracy of fire mapping derived from relatively high resolution Landsat imagery, using the ground-truth verification data; and (3) using the Landsat fire mapping as a basis for assessing the accuracy of fire mapping produced from more coarse-resolution NOAA-AVHRR imagery, employing a regression approach. As well, using fire map data interpreted from Landsat imagery (both MSS and TM) and NOAA-AVHRR for two of these scene areas, the size class distribution of fires was determined principally to assess the reliability of mapping small fires with NOAA-AVHRR imagery. Accuracy of fire mapping undertaken at the Landsat scale was consistently greater than 80% (range 84-99%). For assessing the accuracy of fire mapping derived from NOAA-AVHRR, a regression approach was applied using the mapped extent of burning from both Landsat and NOAA-AVHRR in 10 X 10 km grid cells. Results indicated variable, but generally high mapping accuracy, with r2 ranging from 0.41-0.81. Assessment of the size class distribution of fires indicates that mapping from NOAA-AVHRR imagery a priori misses all fires less than 1 pixel (1 km) in size; which constitute by far the great majority of individual fires. The assessment indicates that fire mapping accuracy using current semi-manual mapping procedures is generally high, but needs to be substantially automated if mapping quality is to be improved consistently.