TY - JOUR
T1 - A hierarchical Bayesian approach to the classification of C3 and C4 grass pollen based on SPIRAL δ13C data
AU - Urban, Michael A.
AU - Nelson, David M.
AU - Kelly, Ryan
AU - Ibrahim, Tahir
AU - Dietze, Michael
AU - Pearson, Ann
AU - Hu, Feng Sheng
N1 - Funding Information:
We thank Triet Vuong for assistance with sample preparation and isotopic analysis, and Peter Kershaw, Patrick Moss, Alayne Street-Perrott, John Tibby and Dirk Verschuren for providing surface-sediment samples. We are also grateful to three anonymous reviewers and the associate editor for feedback that improved the manuscript. Funding was provided by NSF DEB-0816610 (FSH, DMN, AP) and NSF EF-1065848 (MCD). Sample collection in lakes Naivasha and Challa was done under Kenya MOEST research permit 13/001/11C to Dirk Verschuren.
Copyright:
Copyright 2013 Elsevier B.V., All rights reserved.
PY - 2013/11/15
Y1 - 2013/11/15
N2 - Differentiating C3 and C4 grass pollen in the paleorecord is difficult because of their morphological similarity. Using a spooling wire microcombustion device interfaced with an isotope ratio mass spectrometer, Single Pollen Isotope Ratio AnaLysis (SPIRAL) enables classification of grass pollen as C3 or C4 based upon δ13C values. To address several limitations of this novel technique, we expanded an existing SPIRAL training dataset of pollen δ13C data from 8 to 31 grass species. For field validation, we analyzed δ13C of individual grains of grass pollen from the surface sediments of 15 lakes in Africa and Australia, added these results to a prior dataset of 10 lakes from North America, and compared C4-pollen abundance in surface sediments with C4-grass abundance on the surrounding landscape. We also developed and tested a hierarchical Bayesian model to estimate the relative abundance of C3- and C4-grass pollen in unknown samples, including an estimation of the likelihood that either pollen type is present in a sample. The mean (±SD) δ13C values for the C3 and C4 grasses in the training dataset were -29.6±9.5‰ and -13.8±9.5‰, respectively. Across a range of % C4 in samples of known composition, the average bias of the Bayesian model was <3% for C4 in samples of at least 50 grains, indicating that the model accurately predicted the relative abundance of C4 grass pollen. The hierarchical framework of the model resulted in less bias than a previous threshold-based C3/C4 classification method, especially near the high or low extremes of C4 abundance. In addition, the percent of C4 grass pollen in surface-sediment samples estimated using the model was strongly related to the abundance of C4 grasses on the landscape (n=24, p<0.001, r2=0.65). These results improve δ13C-based quantitative reconstructions of grass community composition in the paleorecord and demonstrate the utility of the Bayesian framework to aid the interpretation of stable isotope data.
AB - Differentiating C3 and C4 grass pollen in the paleorecord is difficult because of their morphological similarity. Using a spooling wire microcombustion device interfaced with an isotope ratio mass spectrometer, Single Pollen Isotope Ratio AnaLysis (SPIRAL) enables classification of grass pollen as C3 or C4 based upon δ13C values. To address several limitations of this novel technique, we expanded an existing SPIRAL training dataset of pollen δ13C data from 8 to 31 grass species. For field validation, we analyzed δ13C of individual grains of grass pollen from the surface sediments of 15 lakes in Africa and Australia, added these results to a prior dataset of 10 lakes from North America, and compared C4-pollen abundance in surface sediments with C4-grass abundance on the surrounding landscape. We also developed and tested a hierarchical Bayesian model to estimate the relative abundance of C3- and C4-grass pollen in unknown samples, including an estimation of the likelihood that either pollen type is present in a sample. The mean (±SD) δ13C values for the C3 and C4 grasses in the training dataset were -29.6±9.5‰ and -13.8±9.5‰, respectively. Across a range of % C4 in samples of known composition, the average bias of the Bayesian model was <3% for C4 in samples of at least 50 grains, indicating that the model accurately predicted the relative abundance of C4 grass pollen. The hierarchical framework of the model resulted in less bias than a previous threshold-based C3/C4 classification method, especially near the high or low extremes of C4 abundance. In addition, the percent of C4 grass pollen in surface-sediment samples estimated using the model was strongly related to the abundance of C4 grasses on the landscape (n=24, p<0.001, r2=0.65). These results improve δ13C-based quantitative reconstructions of grass community composition in the paleorecord and demonstrate the utility of the Bayesian framework to aid the interpretation of stable isotope data.
UR - http://www.scopus.com/inward/record.url?scp=84883533174&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84883533174&partnerID=8YFLogxK
U2 - 10.1016/j.gca.2013.07.019
DO - 10.1016/j.gca.2013.07.019
M3 - Article
AN - SCOPUS:84883533174
SN - 0016-7037
VL - 121
SP - 168
EP - 176
JO - Geochimica et Cosmochimica Acta
JF - Geochimica et Cosmochimica Acta
ER -