Burn extent and severity mapping by spectral anomaly detection in the Landsat data cube

Mapping the extent and severity of bushfires is an important part of post-fire damage assessment and contributes to the fire history of a region. This information in turn is used in estimating the gradual increase in fuel load after the fire and hence is a key variable in anticipating future fire risk. Spectral indices, i.e., linear combinations of multi-spectral bands in remote sensing imagery, are conventionally used to determine burn extent and severity by comparing differences in imagery obtained pre- A nd post-fire events. For example, the Normalised Burn Ratio (NBR) exploits differences in the relative spectral response in shortwave and near-infrared wavebands to identify areas of burnt vegetation in satellite imagery. Well-known limitations of this differencing approach include its limited consistency and applicability over large areas, the requirement of a priori knowledge of where and when a fire occurred, and the need for imagery acquired within a reasonable time before and after the burn so that seasonal changes and recovery effects are minimal. To address these challenges, we developed a combined burn extent and severity mapping approach that uses the full spectral information in time series of Landsat satellite observations available through the Digital Earth Australia archive. The method is primarily designed for perennial vegetation that does not burn frequently. The principle is to identify spectral anomalies in space and time, i.e., spectra that stand out significantly from the time series. The method quantifies the average spectral response of a pixel using the robust geometric median, which is relatively insensitive to residual atmospheric effects. Deviation of the spectral response from the geometric median is quantified through the cosine distance, a measure based on spectral similarity. Pixels with a distance greater than the equivalent of three standard deviations from the mean (i.e., statistical outliers) are identified as having changed. Subsequently, absolute and relative NBR and cosine distance changes are calculated to identify burns from other possible landcover changes. A subsequent region-growing step improves the classification by contracting pixels with below-threshold evidence of burning. In an optional post-processing step, corroborating data such as fire detections from thermal remote sensing (e.g., Geoscience Australia's Sentinel Hotspots fire detection system) and other ancillary data can be used to improve classification further. We evaluated the method for several case studies in southern Australia through a comparison with independently derived burn extent maps provided by government agencies. The results show that the fully-automated algorithm developed produces classification results that are commensurate with conventional supervised image classification methods, but with the benefit of being repeatable and fully automated.