Date of Award

Summer 8-23-2019

Level of Access

Open-Access Thesis



Degree Name

Doctor of Philosophy (PhD)


Forest Resources


Daniel Hayes

Second Committee Member

Aaron Weiskittel

Third Committee Member

Shawn Fraver

Additional Committee Members

Bruce Cook

John Kershaw


For two decades Light Detection and Ranging (LiDAR) data has been used to develop spatially-explicit forest inventories. Data derived from LiDAR depict three-dimensional forest canopy structure and are useful for predicting forest attributes such as biomass, stem density, and species. Such enhanced forest inventories (EFIs) are useful for carbon accounting, forest management, and wildlife habitat characterization by allowing practitioners to target specific areas without extensive field work. Here in New England, LiDAR data covers nearly the entire geographical extent of the region. However, until now the region’s forest attributes have not been mapped. Developing regional inventories has traditionally been problematic because most regions – including New England – are comprised of a patchwork of datasets acquired with various specifications. These variations in specifications prohibit developing a single set of predictive models for a region. The purpose of this work is to develop a new set of modeling techniques, allowing for EFIs consisting of disparate LiDAR datasets. The work presented in the first chapter improves upon existing LiDAR modeling techniques by developing a new set of metrics for quantifying LiDAR based on ecological ii principles. These fall into five categories: canopy height, canopy complexity, individual tree attributes, crowding, and abiotic. These metrics were compared to those traditionally used, and results indicated that they are a more effective means of modeling forest attributes across multiple LiDAR datasets. In the following chapters, artificial intelligence (AI) algorithms were developed to interpret LiDAR data and make forest predictions. After settling on the optimal algorithm, we incorporated satellite spectral, disturbance, and climate data. Our results indicated that this approach dramatically outperformed the traditional modeling techniques. We then applied the AI model to the region’s LiDAR, developing 10 m resolution wall-to-wall forest inventory maps of fourteen forest attributes. We assessed error using U.S. federal inventory data, and determined that our EFIs did not differ significantly in 33, 25, and 30/38 counties when predicting biomass, percent conifer, and stem density. We were ultimately able to develop the region’s most complete and detailed forest inventories. This will allow practitioners to assess forest characteristics without the cost and effort associated with extensive field-inventories.