Holocene Evolution of Saharan Vegetation Revealed by Paleoclimate Simulations and Machine Learning

The Greening of the Sahara (GS) during the African Humid Period (AHP) is a striking example of past climate–vegetation interactions, yet its detailed spatiotemporal patterns across the Holocene remain insufficiently understood. This study develops a data-driven framework that combines paleoclimate simulations with machine learning to reconstruct vegetation dynamics in North Africa. Two machine learning models—an artificial neural network (ANN) and a random forest (RF)—were trained on observed nonlinear relationships between vegetation and climate variables. These models were applied to paleoclimate proxy data and the transient climate simulation TraCE-21ka to estimate Holocene normalized difference vegetation index (NDVI). The ANN model outperformed RF in representing complex vegetation–climate linkages and showed closer agreement with proxy evidence. ANN-based reconstructions indicate a rapid expansion of vegetation in North Africa and the Arabian Peninsula following the Younger Dryas (~12,000 years BP), sustained high vegetation cover during the AHP (10,000–6,200 years BP), and a gradual decline thereafter. However, the ANN underestimated both the overall vegetation cover and the abrupt decline around 6,000 years BP suggested by proxy data. Sensitivity analyses highlight monsoon-driven precipitation as the primary control on vegetation change, with temperature exerting a secondary but reinforcing influence. This machine learning–based framework provides a new perspective for investigating vegetation responses to past and future climate change.