Mapping and modeling soil organic carbon in the Eastern Allegheny Plateau and Mountains using legacy data.

摘要

The deeply dissected topography and diverse climate of the Eastern Allegheny Plateau and Mountains (Major Land Resource Area (MLRA) 127) create challenges for dynamicpedoecological modeling needed for ecosystem management adaptation to a changing climate. The spatial distribution of soil organic carbon (SOC), one of the most dynamic soil properties, has been previously estimated and mapped using the State Soil Geographical Database (STATSGO2) and the more detailed Soil Survey Geographic Database (SSURGO) for MLRA 127, estimating mean SOC to a depth of 1 m to be 2.60 and 4.40 kg m-2, respectively. Previous studies have shown that these approximations underestimate true carbon stock due to unpopulated organic horizons and inconsistencies within the databases. Between 1960 and 2009, the USDA-NRCS Kellogg Soil Survey Lab (KSSL) sampled and characterized 254 pedons within MLRA 127 based on soil survey needs. Each pedon had a site description and associated chemical and physical lab analyses to support its taxonomic classification. Data mining revealed that 13% of these 254 pedons lacked soil organic carbon data for one or more horizons and 50% lack bulk density (BD) values. Random forest (RF) and median and mean techniques were assessed, validated, and then used to populate missing BD and SOC data. Geographically weighted regression (GWR) and GWR kriging (GWRK) techniques were then used to model SOC stock in MLRA 127 using prepared and fully populated KSSL pedons and environmental covariates. The resulting SOC predictions were independentaly validated with measured Rapid Carbon Assessment (RaCA) samples and uncertainty was assessed using the fuzzy k-means with extragrades algorithm. Comparisons between GWR and GWRK models created in this study to the RaCA prediction model developed by NRCS showed that nonparametric spatial modeling techniques such as GWRK and RF are able to effectively predict SOC stock within a MLRA. The error rates calculated from the GWR, GWRK, and RaCA models were much lower than previous studies, indicating that SOC prediction by MLRA might be the most suitable way for NRCS to predict SOC stock and that GWRK should be the recommended approach for DSM of SOC. Total biosphere carbon calculated using the Forest Inventory Analysis (FIA) model and substituting GWRK soil for soil carbon and forest litter revealed that soils contain 79% of the total carbon in the terrestrial biosphere of MLRA 127. The methodology presented in this thesis, beginning with preparing KSSL data and ending with an interpolated GWRK model with 95% prediction intervals depicting the SOC stock of the upper 1 m of soil in MLRA 127, is recommended to the NRCS as a guideline for future DSM approaches. Creating, validating, and assessing uncertainty of a SOC model created from measured data and environmental covariates will enhance the understanding of terrestrial biosphere carbon and support national climate change initiatives such as the U.S. Carbon Cycle Science Program.