Abstract:
One of the most common practices regarding estimation of forest attributes is
the partitioning of large forested subpopulations into smaller areas of interest to
coincide with specific objectives of present and future forest management. New
estimators are needed to improve estimation of selected forest attributes in small areas
where the existing sample is insufficient to obtain precise estimates.
This dissertation assessed the strength of light detection and ranging (LiDAR)
as auxiliary information for estimating plot-level forest attributes (trees/ha, basal
area/ha, volume/ha, quadratic mean diameter, Lorey's height) using intensity and nonintensity
area-level LiDAR metrics and single tree remote sensing (STRS). LiDAR
intensity metrics were useful for increasing precision for trees/ha. With the exception
of Lorey's height, STRS did not significantly improve precision for most of the
attributes.
Small area estimation (SAE) techniques were assessed for precision and bias in
estimating stand-level forest attributes (trees/ha, basal area/ha, volume/ha, quadratic
mean diameter, mean height of 100 largest trees/ha) assuming a localized
subpopulation using LiDAR auxiliary information. Selected estimation methods
included area-level regression-based composite estimators and indirect estimators
based on synthetic prediction and nearest neighbor imputation. The composite
estimators produced lower bias and higher precision than synthetic prediction and
imputation. The traditional composite estimator outperformed empirical best linear
unbiased prediction for bias but not for precision.
SAE methods were compared for precision and bias in estimating county-level
forest attributes (trees/ha, basal area/ha, volume/ha, quadratic mean diameter, mean
height of 100 largest trees/ha) assuming a regional subpopulation using Landsat
auxiliary information. Selected estimation methods included unit-level mixed
regression-based indirect and composite estimators, and imputation-based indirect and
composite estimators. The indirect and composite estimators based on linear mixed
effects models generally outperformed those based on imputation. The composite
estimators performed the best in terms of bias for all attributes.
