Regional Scale Characterization of Soil Carbon Fractions with Pedometrics

Abstract:

Regional scale characterization of the spatial distribution of soil carbon (C) fractions can facilitate a better understanding of the lability and recalcitrance of C across diverse land uses, soils, and climatic gradients. While C lability is associated with decomposition and transport processes in soils in, the stable portion of soil C persists in soil for decades to millennia. To better understand storage, flux and processes of soil C from across the soil-landscape continuum, we upscaled different fractions of soil C. Recalcitrant carbon (RC), hydrolysable carbon (HC) and total carbon (TC) were derived from the topsoil (0-20 cm) at 1,014 georeferenced sites in Florida (~150 000 km²). These were identified using a random-stratified sampling design with landuse-soil suborders strata. The Boruta method was employed for identifying all-relevant variables from the available 327 soil-environmental variables in order to develop the most parsimonious model for TC, RC and HC. We compared eight methods: Classification and Regression Tree (CaRT), Bagged Regression Tree (BaRT), Boosted Regression Tree (BoRT), Random Forest (RF), Support Vector Machine (SVM), Partial Least Square Regression (PLSR), Regression Kriging (RK), and Ordinary Kriging (OK). The accuracy of each method was assessed from 304 randomly chosen samples that were used for validation. Overall, 36, 20 and 25 variables stood out as all-relevant to TC, RC and HC, respectively. We predicted TC with a mean of 4.89 kg m^-2 and standard error of 3.71 kg m^-2. The prediction performance based on the ratio of prediction error to inter-quartile range in order of accuracy for TC was as follows: RF>BoRT>BaRT>SVM>PLSR>RK>CART>OK; however, BoRT outperformed RF for RC and HC, and the remaining order was identical for RC and HC. The best models, explained 71.6, 73.2, and 32.9 % of the total variation for TC, RC and HC, respectively. No residual spatial autocorrelation was left among the evaluated models. This indicates that the inclusion of all-relevant environmental variables explained the majority of the variation in soil C fractions.