Estimating Current Forest Attributes from Paneled Inventory Data Using Plot-Level Imputation: A Study from the Pacific Northwest

Abstract

Information on current forest condition is essential to assess and characterize resources and to support resource management and policy decisions. The 1998 Farm Bill mandates the US Forest Service to conduct annual inventories to provide annual updates of each state's forest. In annual inventories, the sample size of 1 year (panel) is only a portion of the full sample and therefore the precision of the estimations for any given year is low. To achieve higher precision, the Forest Inventory and Analysis program uses a moving average (MA), which combines the data of multiple panels, as default estimator. The MA can result in biased estimates of current conditions and alternative methods are sought. Alternatives to MA have not yet been explored in the Pacific Northwest. Data from Oregon and Washington national forests were used to examine a weighted moving average (WMA) and three imputation approaches: most similar neighbor, gradient nearest neighbor, and randomForest (RF). Using the most recent measurements of the variables of interest as ancillary variables, RF provided almost unbiased estimates that were comparable to those of the MA and WMA estimators in terms of root mean square error.