In November 2020, AIMS and GA collaborated to survey the largely unexplored Arafura Marine Park in northern Australia.with high-resolution multibeam sonar and towed video transects as part of the NESP Marine Biodiversity Hub. These data were used to build spatial predictive benthic habitat models – one centered on a shallow reef (Money Shoal; 10-70 m) and the other on a deep seabed feature (Pillar Bank; 180 – 200 m). Models were constructed using Random Forest implemented in Python with two-thirds of the towed video observations, and classification performance was assessed using the remaining one-third of the data. Classification performance within each study area was also modelled using locally varying versions of each of a range of metrics. We found that while global measures showed similarly high model performance in both study areas, the global performance was representative of local performance for most of Money Shoal, but only for a fraction of Pillar Bank. Managers of the Marine Park that need to make spatial decisions can thus be confident to trust the predicted benthic habitat maps at Money Shoal but should be wary of using them for all but a few selected areas of Pillar Bank.