Vegetation and Soil Drivers of Depth-Resolved Soil Carbon Stocks in U.S. Drylands

Drylands cover more than 41% of the Earth’s terrestrial surface and provide ecosystem services to approximately 2 to 2.5 billion people. Drylands store roughly 30% of the world’s soil organic carbon (C) and exhibit high spatiotemporal variability in biogeochemical cycling, making them a critical component for accurately quantifying terrestrial carbon and nitrogen budgets. This, in turn, requires an improved understanding of the underlying biogeochemical processes. One key challenge in advancing our understanding of dryland carbon cycling is capturing processes occurring in both surface and subsurface soils. Empirical studies have provided valuable insights into how climate, vegetation, and land management shape the distribution of surface SOC, but few have explored the importance of these drivers in explaining deep SOC across diverse dryland systems. 

Here, we leverage the Rapid Carbon Assessment (RaCA), a USDA-NRCS initiative launched in 2010 to quantify SOC stocks by genetic horizon to 100 cm depth across the conterminous United States, including more than 2,400 dryland sites spanning multiple land use/land cover (LULC) types, including Rangelands (shrublands and grasslands) and Forest (deciduous, evergreen, and mixed forests). Using a multi-objective Random Forest model to predict SOC stocks at 0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm, and 60–100 cm depths, we examine the role of vegetation in explaining SOC distribution with depth, including metrics such as NDVI, net primary productivity (NPP), and plant functional cover derived from the Rangeland Analysis Platform. We additionally assess the influence of climate (e.g., aridity index) and soil properties, including texture, pH, rock fragment content, calcium carbonate concentration, and sodium adsorption ratio, obtained from the Soil Survey Geographic Database and complementary data-driven products. 

Model performance decreases with depth, with Rangelands sites performing better in the topsoil layers (R2 = 0.45, RMSE = 8.8) than Forest sites (R2 = 0.28, RMSE = 11.56). In forest systems, the highest performance was observed in the 0–5 cm layer (R² = 0.38). In contrast, rangeland systems showed their highest model performance at the 5–15 cm (R² = 0.53) and 15–30 cm (R² = 0.46) intervals, evidence of the strong link between aboveground and belowground plant production on Rangeland systems. Accumulated Local Effects (ALE) and Shapley Additive Explanations (SHAP) were used to characterize the functional form and relative contribution of individual predictors learned by the model across land-cover types and soil depths. In Rangelands, SOC predictions increase monotonically with increasing aridity and net primary productivity (NPP), whereas Forest systems exhibit saturation at high values of these predictors. This contrast highlights that the non-saturating vegetation response to water availability in dryland rangelands, and the saturation of forest productivity under high precipitation regimes, are also reflected in modeled SOC stocks. 

Addressing these questions will advance understanding of dryland biogeochemical processes and support more accurate representation of these systems in terrestrial biosphere models.