Development of an automated method for mapping fire history captured in Landsat TM and ETM + time series across Queensland, Australia

Remote sensing can quantify past and present fire activity at spatial scales useful for a range of fire and vegetation management applications. In this study, we present a new automated approach to classifying burnt areas across the state of Queensland, Australia. The method is applied to complete time series of Landsat TM/ETM + imagery rather than single images and considers spectral (band 4, B4, and bands 4 + 5, B45), thermal, temporal and contextual information within a hierarchical framework. To maximise the available observations and the burnt area detected, we used imagery containing up to 60% cloud that was screened during pre-processing. Median filters were applied to smooth the time series and multi-date change detection used to locate negative outliers (large declines in reflectance relative to the median-smoothed time series). Watershed region growing was used to segment and map a larger spatial extent of the change while minimising commission errors. These segmented change objects were attributed as either burnt or unburnt using their thermal, reflective and contextual characteristics in a classification tree. Thermal information was found to be more important than reflective indices in the change attribution. Algorithm calibration used training data from ten Path/Rows located strategically across Queensland with four images sampled per path row (n = 40). Thresholds were optimised to maximise the burnt area detected while limiting under/over-growing of burnt area. Validation data covered a range of burnt areas from ten independent Path/Rows with ten images sampled across a range of burnt area fractions per Path/Row (n = 100). The results for burnt area mapping demonstrated an average producerʹs accuracy of 85% (range of 28 to 100% for individual images) and average userʹs accuracy of 71% (range of 4 to 99% for individual images). A morphological dilation of one pixel restricted to locations exhibiting a decline in B45 over time, increased the producerʹs accuracy by 4% but reduced the userʹs accuracy by 8%. The total accuracy for the burnt area classification was greater than 99%, however this is more a reflection of the small fraction of landscape represented by burnt area rather than the ability to detect burnt area. Areas frequently misclassified were related to areas of high spectral/land use change which included areas of cropping, frequently inundated land, and moisture/ground cover variations over dark soils. In this study, we applied a crop and water mask to minimise commission errors. Significantly, the results of this study demonstrate that an automated time series method for mapping burnt areas can be successfully applied across a diversity of land cover types. The method may be applied in similar savanna dominated environments but is likely to require modification to be applicable in other landscapes.