Space-Time-Depth Modelling Soil Organic Carbon Stocks Across Brazil

Digital soil mapping (DSM) is key to producing soil information. This presentation will showcase how we employed DSM to model soil organic carbon (SOC) stocks in space and time (1985-2024) across Brazilian agricultural and natural ecosystems. SOC stocks were modeled using a tree-based regression model that established statistical relationships between field data and predictor variables. Field data came from 199 public studies available in the SoilData repository. After standardization, harmonization, and gap filling, the analysis-ready dataset comprised 15,068 georeferenced points and 28,047 samples. Predictor variables were obtained from open databases to represent soil-forming factors that structure spatial variation (static predictors) and capture changes over time (dynamic variables). The trained model (ME = 0.34 t/ha, RMSE = 26.62 t/ha, MEC = 0.73) was used to produce a 40-year annual time series of SOC stock (baseline) maps (0-30 cm). Model outputs were compared with existing national and global maps to assess consistency. Spatial reliability was estimated by calculating the similarity of predictions location to training samples. Our results showed that, in 2024, Brazil stored ~37.5 Gt of SOC in the top 30 cm of soil. More than half of this stock was in the Amazon (19.5 Gt), while the Atlantic Forest exhibited the highest mean SOC stock (53.4 t ha⁻¹). SOC stocks were consistently lower in sandy than in clayey and silty soils, with contrasts approaching a twofold difference in savanna and semi-arid regions. The maps will be updated annually as more data become available and modeling strategies are refined.