Author ORCID Identifier

https://orcid.org/0009-0009-3329-6307

Semester

Spring

Date of Graduation

2025

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Xin Li

Committee Co-Chair

Natalia Schmid

Committee Member

Natalia Schmid

Committee Member

Matthew Valenti

Committee Member

Kevin Bandura

Committee Member

Yong-Lak Park

Abstract

Forest and agricultural ecosystems are increasingly at risk due to invasive species, pests, and diseases, necessitating scalable, automated, and intelligent monitoring solutions. Traditional field based forest and agriculture health assessments are limited by cost, time, and spatial coverage. This dissertation presents a multiscale deep learning framework that automates forest and agriculture health monitoring using drone imagery and computer vision techniques. The system operates across three spatial levels: forest level, tree level, and leaf level, combining object detection, segmentation, and classification models to support large scale ecological assessment.

At the forest level, high-altitude drone imagery is processed using object detection and segmentation models, leveraging the state-of-the-art Mask2former architecture to detect and map large-scale ecological threats, such as invasive species and pest infestations. At the tree level, two complementary approaches are employed: a semi-supervised learning pipeline using a novel Gaussian Mixture Model (GMM) for limited labeled data, and a deep learning segmentation model optimized for large datasets to provide fine-grained spatial mapping of invasive species presence. Our semi-supervised GMM combines supervised and unsupervised learning, incorporating user input for manual region selection and data augmentation to enhance model robustness. At the leaf level, a classification model based on transformer architectures is used to differentiate invasive species, enabling species-level identification based on visual patterns.

Our framework integrates deep learning, semi-supervised learning, and efficient data handling to enable large-scale ecological surveillance. By combining supervised and semi-supervised learning, it adapts to varying data availability, providing accurate monitoring of forest and agriculture health. This scalable system enhances automated ecological surveillance, offering valuable insights for forest conservation and ecosystem resilience.

Download

Share

COinS