Cross-Regional Digital Soil Carbon Modeling Considering Scaling Effects for Soil Carbon Assessment At Global Scale.

The implications of transposing regional digital soil models to continental and global scales are still poorly understood. Our objectives were to (i) compare the prediction performance of soil carbon models in two contrasting regions in the United States and (ii) elucidate on the scaling effects to accurately assess soil carbon employing disparate sets of environmental covariates derived at fine and coarse scales, respectively. We used soil organic carbon data in the topsoil (0-20 cm) and aggregated 0-100 cm depth across Colorado and Florida from the U.S. National Soil Survey Database (Natural Resource and Conservation Service, NRCS). Environmental covariate sets were assembled at two scales representing (i) local/regional scale ¨C high spatial resolution and high-density attribute data and (ii) continental/global scale ¨C coarse spatial resolution and generalized attribute data. The environmental covariate sets represented STEP-AWBH variables (S: soils, T: topography, E: ecology, P: parent, A: atmosphere/climate, W: water, B: biota, and H: human) as predictor variables. We used support vector machines and ensemble regression trees and various error metrics to assess the prediction performance of soil carbon. Our results suggest that spatial and attribute aggregation of environmental covariates profoundly affect the strength of relationships between STEP-AWBH predictor variables and soil carbon as well as their prediction performance. The magnitude of these effects differed among regions. These findings have implications for upscaling of regional digital soil models to continental and global scales. Scaling effects profoundly constrain the quality of universal (global) soil carbon models.