Abstract Forests and savannas frequently co-occur as patches within tropical landscape mosaics, yet the mechanisms controlling their spatial configuration remain unclear. The presence of both vegetation states under similar climatic conditions is often attributed to fire–vegetation feedbacks, but could also reflect variation in overlooked external drivers. In Central Africa, forest–savanna mosaics become more common with increasing topographic roughness, but how well topographic heterogeneity explains the forest–savanna configuration within mosaic landscapes is unknown. Here we address this question and examine the role of individual topographic variables that may influence tree cover by, for instance, changing water availability and fire spread. We identify mosaic landscapes from remotely sensed tree cover data and derive topographic variables from a digital elevation model. We use these variables to develop machine learning algorithms predicting vegetation state within mosaic landscapes. Models achieved an average prediction accuracy of 0.75, with local elevation (relative to the surrounding 500 m or 5000 m) emerging as the strongest predictor of vegetation state. Both model accuracy and the role of topographic predictors varied strongly among landscapes, reflecting the diverse pathways by which topography can influence tree cover. Overall, our findings indicate that topographic heterogeneity is a major driver of forest–savanna mosaics in Central Africa. Mosaic landscapes are more deterministic than previously assumed, suggesting that their response to disturbances and climate change will be spatially heterogeneous, thereby reducing the likelihood of abrupt large-scale shifts between forest and savanna states.
Zwaan et al. (2026) studied this question.