The effects of training set selection on the relationship between pollen assemblages and climate parameters: Implications for reconstructing past climate

Bunting, M. Jane; Li, Yuecong; Liu, Jinsong; Tian, Fang; Xu, Qinghai

Abstract

494 surface soil pollen samples were collected from forests, steppes, deserts, shrubs and meadows in the northern China, and compared with modern climate data. Two approaches for estimating climate parameters from pollen data were tested independently, WA-PLS (weighted-averaging regression and calibration partial least squares) and MAT (The Modern Analogues Technique) methods. The results of cross-validation showed that pollen data could effectively estimate annual average precipitation (P-ann) (WA-PLS: R-jack(2) = 0.84, RMSEPjack = 78 mm; MAT: R = 0.91, Std. dev. = 79 mm); but were only moderately successful at estimating either mean annual temperature (WA-PLS: R-jack(2) = 0.48, RMSEjack = 2.5 degrees C; MAT: R = 0.68, Std. dev.= 2.5 degrees C) or the mean temperature of the warmest month (WA-PLS: R-jack(2) = 0.50, RMSEPjack2 = 2.6 degrees C; MAT: R = 0.72, Std. dev.= 2.6 C). One of the reasons for errors in estimating thermal parameters is the presence of some 'troublesome' samples in the dataset, samples from communities with severe human disturbance or from small patches of vegetation close to an ecotone. Removing 104 samples affected by these factors gave a 'BEST' pollen dataset, which was able to estimate thermal parameters with a greater degree of accuracy (R>0.8 or R-jack(2)>0.64). All these methods were applied to a published paleo-record, the Anyang-Chadianpo section, covering the period about 10,000 cal. B.P.-present. MAT and WA-PLS methods reconstructed similar P-ann records, but thermal parameters showed more severe and abrupt changes when reconstructed using the MAT approach rather than the WA-PLS method. Using the different pollen data sets (all 494, or 390 BEST samples) for each method led to similar paleoclimate reconstructions in most periods, apart from periods of climate transition (e.g. 6200-6800 cal. B.P., 3000-4000 cal. B.P.) or intensive human activity (after 1000 cal. B.P.), suggesting that the influence of human activities and small patches on the composition of the pollen record are more significant in these periods. (C) 2010 Elsevier B.V. All rights reserved.

Citation

Bunting, M. J., Li, Y., Liu, J., Tian, F., & Xu, Q. (2010). The effects of training set selection on the relationship between pollen assemblages and climate parameters: Implications for reconstructing past climate. Palaeogeography, palaeoclimatology, palaeoecology, 289(1-4), 123-133. https://doi.org/10.1016/j.palaeo.2010.02.024

• Journal Article Type: Article
• Acceptance Date: Feb 18, 2010
• Online Publication Date: Feb 24, 2010
• Publication Date: Apr 1, 2010
• Deposit Date: Nov 13, 2014
• Journal: Palaeogeography Palaeoclimatology Palaeoecology
• Print ISSN: 0031-0182
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Volume: 289
• Issue: 1-4
• Pages: 123-133
• DOI: https://doi.org/10.1016/j.palaeo.2010.02.024
• Keywords: Earth-Surface Processes; Palaeontology; Ecology, Evolution, Behavior and Systematics; Oceanography
• Public URL: https://hull-repository.worktribe.com/output/472846