Data Mining Methods for Spatial Models of Soil Organic Carbon in Florida, USA.

Soil organic carbon (SOC) is a spatially variable component in the soil landscape. Soil carbon storage enhances regulating and supporting ecosystem services with numerous positive environmental co-effects. Geospatial models that quantify SOC are needed for the valuation and verification carbon credit markets. Hybrid geospatial models (mixed deterministic / stochastic models) can be used to make spatially-explicit estimates of SOC based on large datasets of correlated environmental variables available as geographic information system (GIS) layers. These large datasets can be analyzed for relationships with SOC with a variety of data mining techniques. The objective of this research was to develop data mining models of SOC carbon fractions for a large region in the southeastern U.S. (Florida). A regional dataset was collected (n=1014) from the top 20 cm of Florida soils (~150,000 km²). Total carbon (TC) and inorganic carbon (IC) were measured by combustion and acid reaction via a gas analyzer and SOC derived by subtraction (TC – IC). A database of environmental covariates such as digital elevation models, satellite imagery, soil map-units and map-unit attributes, land cover/land use, canopy density, and biomass were collected in a GIS. Several data mining approaches such as regression trees (single, boosting, bagging, and random forest) and multivariate adaptive regression splines were tested using a calibration/validation split. We examine the performance of individual covariates to model SOC. This research provides estimates of SOC across a large subtropical region composed of diverse soil-hydrology and land uses and identifies environmental variables that impart major control on SOC.