Date of Award

8-2009

Level of Access

Campus-Only Thesis

Degree Name

Master of Science (MS)

Department

Forest Resources

Advisor

Steven A. Sader

Second Committee Member

William Halteman

Third Committee Member

Jeremy Wilson

Abstract

Forests are classified into smaller units, called stands, which are made up of trees of similar species, size, and spacing to allow foresters to more easily understand the complex landscapes that they manage and aid in their decision making. Forest stand type maps historically have been produced by trained photo interpreters who manually interpret high resolution aerial imagery. These stand maps are known for being highly variable due to interpreter experience, and costly due to the large amount of manual labor that is put into their generation. Recent advances in computer technology and availability of high resolution digital imagery has facilitated research using computer-automated systems that attempt to reduce the cost and variability of stand mapping; however, little previous research has been attempted in the northeastern forests of the United States. This research creates an initial framework for an automated interpretation system that will be capable of handling the complex forests of the northeast United States. This automated stand mapping is accomplished by using local maxima analysis, region-growing segmentation to delineate tree crowns, classification of these initial objects, and further grouping of tree objects into forest stands using a process rooted in pyramid node linking. Some new techniques as well as existing methods adapted to this region have been integrated into algorithms using custom code written in the Interactive Data Language (IDL) (ITTVIS, 2009) to create a prototype system designed to take digital multispectral sensor data with a ground resolution of 1 meter or less and produce stand boundaries with labels as a final product. The mapping system can operate on standard desktop computers to process study areas upwards of a million hectares. Initial tree objects were compared to ground data and the metrics of crown area, percent crown cover, and trees per acre were found to be comparable for seven out of 10 of the stands sampled. Forest type estimates were found to be comparable for fifty percent of the sampled stands, with best results for softwood dominant stands. Stand boundaries were visually inspected, but had some delineations that could be improved. An error matrix was created using existing ground data for the study area to validate the stand labels. Forest type and height-based size class labels had an overall accuracy of fifty percent, diameter-based size labels had an overall accuracy of sixty percent, and the density label had an overall accuracy of thirty percent. Initial development and testing has identified some components in the algorithms, such as the development of forest type labels and density class categorization, that will need to be augmented and amended before the system can approach operational status. Also, the methodology needs to be tested in a wider range of northern forest conditions to examine the behavior of the algorithms and to determine the robustness and adaptability of the mapping system. Although more work and improvements in the stand mapping algorithms are needed in future work, the evaluation of the prototype system has shown potential for computer-automated mapping in the complex northeastern forest conditions.

Files over 10MB may be slow to open. For best results, right-click and select "save as..."

Share