PEAK DETECTION IN SEDIMENT CHARCOAL RECORDS USING A GAUSSIAN MIXTURE DISTRIBUTION

GAVIN, D. (1), AND HIGUERA, P. (2)

(1) Department of Plant Biology, University of Illinois, Urbana, IL 61801, (2) College of Forest Resources, University of Washington, Seattle, WA 98105

Decadal-scale resolution sediment records of macroscopic charcoal in areas with stand-replacing fire regimes have the potential to detect individual fire events occurring near the lake. Established methods decompose charcoal records by 1) removing long-term (typically = 500 yr) variation using a moving-average filter, and 2) applying a threshold value to the filtered record to separate "peaks" from "noise". In records that contain intermediate-sized peaks, the choice of a threshold is not clear, even with the aid of independent fire evidence from tree-ring studies. We present a method to objectively identify the threshold value using statistical properties of the charcoal record. We assume that for non-fire samples the variance around the moving average is Gaussian; this variance results from depositional processes and sampling effects. An information-theoretic approach is used to determine the minimum number of overlapping Gaussian distributions that ‘best’ describes the overall frequency distribution. Applied to six Holocene charcoal records, this algorithm always fits two or three Gaussian distributions, one with a mean near zero and a small variance and one or two with a large mean and variance. The threshold is set near the upper limit of the lower, noise-related, distribution. We also test this method against simulated charcoal records.

Left: Charcoal record with a 500-yr loess smoothing (gray line). Right: frequency distribution of residual CHAR from the loess curve. The Gaussian mixture algorithm identified two distributions (gray curves).