Year of Award

2022

Document Type

Dissertation

Degree Type

Doctor of Philosophy (PhD)

Degree Name

Forest and Conservation Science

Department or School/College

W.A. Franke College of Forestry and Conservation

Committee Chair

David Affleck

Commitee Members

C. Kenneth Brewer, Jon Graham, Christopher Keyes, Carl Seielstad

Keywords

bias estimation, forest inventory, forest regeneration, monitoring trends in burn severity, small area estimation, wildland fire

Publisher

University of Montana

Subject Categories

Forest Sciences

Abstract

Wildfire activity in the western United States is expanding and concern for the declining extent of postfire tree cover in many western forests is mounting. Accurate estimates of postfire seedling, sapling, and large tree density following wildfire are critical for postfire forest management planning and monitoring forest dynamics. National forest inventory programs, such as the US Forest Service Forest Inventory and Analysis (FIA) program, can provide vegetation data for direct spatiotemporal domain estimation of postfire tree density, but sample observations within domains of administrative utility are often few to none. This research investigates indirect domain estimators, which borrow sample data from outside the domain to increase precision of domain estimates. Domains consist of National Forest System (NFS) lands burned in a particular US state and over a particular burn period, at varied times–since–burn. On the basis of estimated standard error, a strategy for augmenting domain samples with observations proximate in time proves superior to a strategy that borrows observations proximate in space when using FIA sample data alone. However, estimators of the mean squared error (MSE) of indirect domain estimators prove frequently negative and too highly variable for operational utility in this context. Relationships are therefore explored between observations of postfire tree density and a broad suite of geospatial explanatory variables in efforts to reveal trends and identify candidate auxiliary variables for model–assisted domain estimation. Algorithmic and parametric modeling techniques identify a multispectral satellite-based tree cover product and climate variables as the most important predictors. Yet poor overall performance suggests that a single model of tree regeneration throughout the entire western US is not feasible. Finally, model–assisted small area estimators are compared in a design–based inferential framework. In particular, k nearest neighbor–based (kNN) and linear regression–based small area estimators are developed and compared on the basis of domain–level standard error in domains spanning burned NFS lands within individual US states and decades. A kNN–based technique using only spatial coordinates as predictors yields the lowest standard errors at the domain level, indicating that none of the model–assisted approaches investigated here could be leveraged to better effect than to simply average the nearest observations in space, irrespective of time–since–disturbance.

Download

Share

COinS
 

© Copyright 2022 George Chilton Gaines III