The ability to map fire severity is a requirement for fire management agencies worldwide. The development of repeatable methods to produce accurate and consistent fire severity maps from satellite imagery is necessary to document fire regimes, to set priorities for post-fire management responses, and for research applications. Machine learning techniques, such as random forest, have shown great promise for mapping of wildfire severity in woodland and forest ecosystems using satellite imagery. However, an assessment of the properties of training data required for automated mapping with random forest is currently lacking. This study examined how training data properties affect fire severity classification across forest, woodland and shrubland communities of southern Australia. The aims of this study were: (i) to examine how sample size (i.e. number of training points and fire events) and sample imbalance affect classification accuracy; (ii) to determine whether models were transferrable across geographic regions; and (iii) to assess the need for classifiers for prescribed burns and wildfires. We sampled 33 wildfires and 57 prescribed burns occurring across southern Australia between 2006 and 2019, to derive an extensive dataset (n = 25, 350 points) suitable for model training and validation. Five fire severity classes were mapped across the forest, woodland and shrubland communities. Using independent spatial cross validation of wildfires, we found that a minimum of 300 sample points per severity class, sampled across at least ten independent fires, was sufficient to reach the upper threshold of classification accuracy for the five severity classes. Training datasets derived across broad environmental space or close (≤100 km) to the target fire (i.e. fire to be mapped) produced better predictions than those derived regionally and far (>100 km) from the target fire. Models trained on data derived from both wildfires and prescribed burns had similar overall accuracy as those trained only on data from the fire type being predicted. However, there were significant differences in accuracy between the models trained with wildfires, prescribed burns and combined datasets within some severity classes. The random forest classifier had an overall accuracy of ~88% for wildfires and ~68% for prescribed burns across the study fires. The discrepancy in accuracy between wildfires and burns was likely due to the poorer classification performance for low fire severity classes, the dominant severity classes in prescribed burns. Overall classification accuracy was relatively consistent across forest, woodland and shrubland communities within the study fires, indicating that the method is robust across the temperate forest biome of southern Australia. Our results demonstrate that consideration of training data properties, specifically the number of points and fires sampled, sample balance and the geographic source of sample data, will be important considerations for the automated mapping of fire severity using random forest classification and Landsat imagery.