Learning the Green Wave: A Hybrid Machine Learning Framework for Reconstructing Past Vegetation Dynamics

Understanding the seasonal timing of vegetation growth ("Green Wave") is crucial for modeling prehistoric human mobility and settlement patterns. However, high-resolution vegetation data is only available for the modern satellite era. To reconstruct these dynamics in the deep past, we present a hybrid modeling approach that combines domain-specific knowledge of seasonality with the flexibility of supervised machine learning.

Our core premise is that while we cannot observe the past directly, we can learn the rules of phenology from the present. We utilize global modern datasets to learn a mapping between climatic conditions and vegetation greenness, which can then be applied to paleoclimate simulations.

Our method decomposes the problem. First, we compress modern satellite observations into compact, interpretable parameters using a harmonic seasonal model. Second, we train a machine learning regressor to learn the complex, non-linear mapping between bioclimatic drivers and these phenological parameters. By treating the seasonal shape as a prediction target, we ensure that our reconstructions maintain structural integrity. We validate the model using spatially-disjoint cross-validation to account for spatial autocorrelation, ensuring robust generalization. The resulting framework allows us to translate paleoclimate simulations into high-resolution maps of ancient vegetation seasons, providing new quantitative inputs for archaeological hypotheses.