Global denudation rates based on machine learning

Denudation is arguably one of the most important processes in determining the functioning of the Earth’s surface, from landscape morphology to CO₂ consuming silicate weathering and soil/sediment production. While slope is the primary control on denudation rates, slope-based models explain only half of the observed variance, systematically underestimating the highest rates and overestimating the lowest rates. This discrepancy arises from the lack of other environmental factors and the decoupling of denudation rates from slope beyond certain thresholds. To address this, we propose a novel threshold-control decision tree model, incorporating 14 environmental predictors to analyze denudation rates of ~4000 river basins worldwide. Our results identify key slope thresholds at 3°, 12° and 15°. For slopes below 3°, denudation rates rarely exceed 10 mm/kyr, though high mean annual temperatures can enhance denudation by accelerating chemical weathering. As slope increases, it becomes less determinant, and denudation transitions from transport-limited to detachment-limited regimes. Climate seasonality (3°≤slope<12°), precipitation and seismicity (12°≤slope<15°), and runoff and vegetation coverage (slope≥15°) emerge as critical secondary controls. Our ensemble model of decision trees explains an additional 30% of the variation in denudation rates (R²= 0.82), enabling us to give a more accurate prediction of global denudation rates at 1-km resolution. Our results provide quantitative constraints for understanding Earth surface dynamics over the last millennia and throughout geologic history.