Fire-severity classification across temperate Australian forests: random forests versus spectral index thresholding

Machine learning and spectral index (SI) thresholding approaches have been tested for fire-severity mapping from local to regional scales in a range of forest types worldwide. While index thresholding can be easily implemented, its operational utility over large areas is limited as the optimum index may vary with forest type and fire regimes. In contrast, machine learning algorithms allow for multivariate fire classifications. This study compared the accuracy of fire-severity classifications from SI thresholding with those from Random Forests (RF). Reference data were from 3730 plots within the boundaries of eight major wildfires across the six temperate forest ‘functional’ groups of Victoria, south-eastern Australia. The reference plots were randomly divided into training and validation datasets (60/40) for each fire-severity class (unburnt, low, moderate, high) and forest functional group. SI fire-severity classifications were conducted using thresholds derived in a previous study based on the same datasets. A RF classification algorithm was trained to derive fire-severity levels based on appropriate spectral indices and their temporal difference. The RF classification outperformed the SI thresholding approach in most cases, increasing overall accuracy by 11% on a forest-group basis, and 16% on an individual wildfire basis. Adding more predictor variables into the RF algorithm did not improve classification accuracy. Greater overall accuracies (by 12% on average) were achieved when in situ data (rather than data from other fires) were used to train the RF algorithm. Our study shows the utility of Random Forest algorithms for streamlining fire-severity mapping across heterogeneous forested landscapes.