Graduate Thesis Or Dissertation


Classification strategies for unbalanced binary maps : Finding Ponderosa Pine in the Willamette Valley Public Deposited

Downloadable Content

Download PDF


Attribute NameValues
  • Forest species classifications are becoming increasingly automated as advances are made in machine learning. The algorithms used to identify tree species range from simple decision trees to intricate neural networks, and often excel in accurately delineating tree species. However, complex algorithms can have high input costs, including the cost of high resolution data and time required to train both the researcher and the algorithm. Furthermore, the proofs that support the performances of such algorithms frequently use small study areas while attempting to classify down to the individual tree level. Such precise techniques may be impractical for large-scale classifications, which instead may benefit from simpler, conventional methods of classification, such as maximum likelihood (ML). In this study, I compared the results of a human-guided maximum likelihood (HGML) classification with the results of an automated random forests (ARF) classification of ponderosa pine stands in the southern Willamette Valley, Oregon. The study area is approximately 816,600 ha, significantly larger than most study areas used for presenting techniques for tree species classification, which typically occur on a scale of a few thousand hectares. The goal of the classification was to identify contiguous stands of ponderosa pine against a mix of forest and non-forest background. The classifications were performed using 1m resolution RGBI aerial imagery from the USDA’s National Agriculture Imagery Program. Following classification, I applied a height mask from a 3m resolution lidar-derived canopy height model (CHM). The lidar data was acquired from the Oregon Lidar Consortium. The masks reduced the pine class commission error for mature mixed-conifer forests and non-forest vegetation areas. The HGML classifications achieved an average accuracy of about 63% (kappa 0.48), with a pine class producer accuracy (non-omitted pixels) of 74%, and user accuracy (non-committed pixels) of 60%. With CHM masks, the HGML classifications increased to 92% overall accuracy (kappa 0.60) with a 61% producer accuracy and 97% user accuracy for the pine class. The ARF classification achieved an 86% overall accuracy (kappa 0.46), with a pine class producer accuracy of 82% and user accuracy of 85%. With the application of a CHM mask, the ARF accuracies increased slightly: it reached 90% overall accuracy with a kappa of 0.49, with a 65% producer accuracy and a 93% user accuracy. These results suggest that while multiple classification methods can reach high accuracies in tree species classifications, the HGML method, when enhanced with a CHM mask, is more suitable for large-scale classifications of RGBI imagery, as is evident by its high kappa coefficient.
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Committee Member
Academic Affiliation
Rights Statement
Peer Reviewed
Embargo reason
  • Ongoing Research
Embargo date range
  • 2018-06-05 to 2019-01-06



This work has no parents.

In Collection: