Author ORCID Identifier

https://orcid.org/0009-0004-0856-3569

Semester

Fall

Date of Graduation

2025

Document Type

Thesis

Degree Type

MS

College

Davis College of Agriculture, Natural Resources and Design

Department

Forest Resource Management

Committee Chair

Paul Kinder

Committee Co-Chair

Jamie Schuler

Committee Member

Michael Strager

Abstract

Invasive species are introduced species that cause detrimental impacts to the native ecosystem or cause harm to human health. The work necessary for land managers to survey for and treat invasive species takes up limited time, labor, and financial resources. This study assesses how Unmanned Aerial Systems (UAS) and machine learning models can be used to make the identification, mapping, and treatment of invasive vegetation more efficient and effective. A UAS-mounted multispectral camera captured images of our target species, Rosa multiflora (multiflora rose), throughout its growing season in southwest Pennsylvania. A random forest machine learning model was trained on these images for the identification of multiflora rose. An iterative drop-out test of the flight dates was able to identify the phenological period during which multiflora rose was best identified by the model. The results of the drop-out tests were used to train 3 limited-data random forest models, of which the single flight, early leaf-on model performed the best. The single flight model produced a multiflora rose class F1 score of 0.41 for the test dataset. A field assessment of the single flight results identified the species present in false positive polygons and identified differences in characteristics between true positive and false negative multiflora rose polygons. With additional multiflora rose shrubs identified by the model, the single flight multiflora rose class F1 score was 0.5, with the most common species identified in the false positive polygons being invasive shrub honeysuckle (Lonicera spp). While the species level classification accuracies from this model fall below successful accuracies, understanding common sources of error within the model provide an informed base model to build off of for general invasive species classification or improved species level classification. The random forest training polygons were also used to guide targeted UAS-delivered herbicide treatment of the identified shrubs.

Share

COinS