Abstract
Amazonian dark earth (ADE) is anomalously fertile and carbon-rich soil created by past inhabitants of the Amazon basin. Despite its importance to cultural heritage and carbon sequestration, efforts to systematically map the distribution and extent of ADE are hindered by difficulties of access and field excavation. To circumvent these barriers, we use a machine-learning classifier applied to remote sensing imagery to predict the occurrence of ADE across the 26,000 km (super 2) Xingu Indigenous Territory (XIT). We compile training data derived from field excavation and mapping as well as spectrally distinct vegetation patches, with which we train a random-forest classifier on a two-season Landsat 5 composite image to produce classification maps for the XIT with predicted locations of ADE. We predict 710 km (super 2) of ADE within the XIT (approximately 2.7% of the landscape by area) and find a strong correspondence between ADE locations and topography, with sites located along the edges of bluffs adjacent to river floodplains and tributary streams. We further estimate that the XIT may hold 7 Mt of anthropogenic carbon within ADE deposits.