Regional-scale fire severity mapping of Eucalyptus forests with the Landsat archive

Fire managers seek accurate methods to map fire severity for forest management efforts. Fire severity data with wide spatial and temporal coverage enhances understanding of fire's impact on vegetation and improves planning to mitigate the risk of destructive wildfires. In this study, we develop a machine learning workflow that converts a spatial database of known fire events into a collection of gridded fire severity maps representing levels of canopy scorch due to wildfires and planned fuel reduction burns (prescribed burns). This workflow uses a dataset of observed fire severity interpreted from post-fire aerial images (3–7 cm spatial resolution; ~11,649 labels from 64 unique fires) to train and test a supervised classifier that predicts fire severity using 13 Landsat-derived spectral indices. For a target fire, the classifier generates a pixel-level prediction of fire severity with classes belonging to one of five ordinal categories ranging from unburnt, three levels of canopy scorch, and canopy consumed by fire. The classifier's accuracy was 84.2% (Kappa score = 0.799) and testing was undertaken across multiple fire types (wildfire / prescribed burn), seasons, forest types and topography. MODIS thermal anomalies and a MODIS-derived burn area product were used to augment the fire severity database with start- and end-of-fire dates to enable accurate delineation of pre- and post-fire Landsat imagery; this method increased model accuracy from 78.0% (Kappa = 0.72). The model was then applied to a regional 16-year (2005 to 2020) history of fire events (n fires = 713, ~1Mha) in the sclerophyll eucalypt forests and woodlands of the Northern Jarrah Forest, Western Australia. On fires with MODIS-adjusted start and end dates, high severity fire (>80% canopy scorch) was predicted more often in wildfires (67.9%, 124900 ha) compared to prescribed burns (14.6%, 110300 ha). However, the total predicted area of high-severity fire was comparable between fire types due to prescribed burns being more frequent on the landscape. We provide the first fire severity model in Western Australia that is broadly calibrated to both prescribed burns and wildfires across multiple conditions. Our model outputs are made freely available and offer opportunities to better understand the interactions between prescribed burns and wildfires, and the effects of fire severity on the environment.