Severe fires in tropical savanna systems are recognised as incurring significant impacts on a variety of ecological attributes, including woody vegetation structure and greenhouse gas emissions. However, knowledge of the frequency and extent of severe fires is restricted given challenges associated with the development of reliable remotely sensed mapping procedures. This study takes advantage of three wildfires, 900–5300 km2 in extent, containing very severely affected areas, occurring in semi-evergreen, eucalypt-dominated, tropical Australian savanna, which resulted in significant areas of complete canopy scorch, very significant tree stem mortality (24–55%), and associated loss of living above ground biomass (47–69%) at respective sites. Although increased map scale is generally considered to improve the reliability of fire severity mapping, our analysis found > 90% agreement between Landsat and MODIS-derived burnt area mapping, and > 80% for binary (severe vs. non-severe) fire severity mapping. Mapping of internal fire (unburnt) patchiness was enhanced with finer resolution Landsat imagery, but the much longer orbital return cycle precluded its use at two of the three sites given extended cloudy conditions. Application of an automated MODIS-derived fire severity mapping algorithm (overall reliability in 2015 = 75%) calibrated for generalised north Australian savanna conditions, suggests that 15% and 12% of Australia's 1.3 M km2 tropical savannas region were burnt by severe fires in 2015 and 2016, respectively. The study illustrates the potential for MODIS-derived fire severity mapping, the impacts of very severe fires on stand structure, and ongoing challenges associated with deriving reliable fire severity mapping products in Australian savanna systems.