Date of Award

Spring 5-12-2018

Level of Access Assigned by Author

Open-Access Dissertation

Degree Name

Doctor of Philosophy (PhD)


Forest Resources


Steven Sader

Second Committee Member

Aaron Weiskittel

Third Committee Member

Robert Seymour

Additional Committee Members

Jeremy Wilson

Daniel Harrison

William Halteman


The sustainable management of forest landscapes requires an understanding of the functional relationships between management practices, changes in landscape conditions, and ecological response. This presents a substantial need of spatial information in support of both applied research and adaptive management. Satellite remote sensing has the potential to address much of this need, but forest conditions and patterns of change remain difficult to synthesize over large areas and long time periods. Compounding this problem is error in forest attribute maps and consequent uncertainty in subsequent analyses. The research described in this document is directed at these long-standing problems.

Chapter 1 demonstrates a generalizable approach to the characterization of predominant patterns of forest landscape change. Within a ~1.5 Mha northwest Maine study area, a time series of satellite-derived forest harvest maps (1973-2010) served as the basis grouping landscape units according to time series of cumulative harvest area. Different groups reflected different harvest histories, which were linked to changes in landscape composition and configuration through time series of selected landscape metrics. Time series data resolved differences in landscape change attributable to passage of the Maine Forest Practices Act, a major change in forest policy. Our approach should be of value in supporting empirical landscape research.

Perhaps the single most important source of uncertainty in the characterization of landscape conditions is over- or under-representation of class prevalence caused by prediction bias. Systematic error is similarly impactful in maps of continuous forest attributes, where regression dilution or attenuation bias causes the overestimation of low values and underestimation of high values. In both cases, patterns of error tend to produce more homogeneous characterizations of landscape conditions. Chapters 2 and 3 present a machine learning method designed to simultaneously reduce systematic and total error in continuous and categorical maps, respectively. By training support vector machines with a multi-objective genetic algorithm, attenuation bias was substantially reduced in regression models of tree species relative abundance (chapter 2), and prediction bias was effectively removed from classification models predicting tree species occurrence and forest disturbance (chapter 3). This approach is generalizable to other prediction problems, other regions, or other geospatial disciplines.