Canadian Forest Service Publications

Adaptive cluster sampling for estimation of deforestation rates. 2005. Magnussen, S.; Kurz, W.A.; Leckie, D.G.; Paradine, D. European Journal of Forest Research 124(3): 207-220.

Year: 2005

Issued by: Pacific Forestry Centre

Catalog ID: 25956

Language: English

Availability: PDF (download)

Available from the Journal's Web site.
DOI: 10.1007/s10342-005-0074-6

† This site may require a fee

Mark record


National estimates of deforestation rates may be based on a survey. Precise estimation requires an efficient design. When deforestation rates are low (<1%) large sample sizes are required with traditional sampling designs to meet a precision target. This study explores the efficiency of adaptive cluster sampling (ACS) for this estimation problem. The efficiency is assessed by simulated ACS sampling from 18,200 × 200 km populations with 78–10,742 deforestation polygons (DFP) of different shape and size and average 10-year deforestation rates between 0.2% and 1.0%. Each population is composed of four million square 1 ha population units (PU) in a regular grid. Relative root mean square errors (RMSE) of ACS were, depending on sample size, 30–50% lower than comparable errors with simple random sampling (SRS) designs. ACS achieves this advantage by adaptively adding PUs to an initial SRS sample of size n. Realized ACS sample sizes were, on average, twice the nominal size (n). Three measures of ACS efficiency indicated that the costs of adaptively increasing the sample size are critical for the effectiveness of ACS. Population effects were manifest in all estimators. Estimates of the abundance, size, and shape of DFPs will allow a prediction of these effects. Populations dominated by a few large DFPs were clearly unsuited for ACS. The performance of ACS relative to that of SRS was similar across plot sizes of 1, 10, and 40 ha. The general conclusion of this study is that the lower RMSE of ACS remains attractive when the average cost of adaptively adding a PU to the initial sample is low relative to the average cost of sampling a PU at random.