Date Approved

8-19-2025

Graduate Degree Type

Thesis

Degree Name

Applied Computer Science (M.S.)

Degree Program

School of Computing and Information Systems

First Advisor

Rahat Ibn Rafiq

Second Advisor

Byron DeVries

Third Advisor

Lawrence Burns

Academic Year

2024/2025

Abstract

Forests are critical ecosystems, delivering services such as biodiversity conservation, climate regulation, timber production, and recreation. However, they face increasing threats from pathogens like Bretziella fagacearum, which causes Oak Wilt, a lethal disease that disrupts water transport in oak trees, leading to canopy dieback and eventual death. Traditional detection methods rely on manual ground surveys, which are labor-intensive, time-consuming, and prone to error, particularly in early disease stages.

This research presents an automated, scalable, high-precision Oak Wilt detection system using Unmanned Aerial Vehicles (UAVs) combined with deep learning-based computer vision. Expanding on earlier work with a lightweight CNN achieving an F1-score of 79.8%, this study compares twelve state-of-the-art architectures—traditional CNNs, mobile-optimized models, hybrid networks, and vision transformers—across multiple evaluation metrics. A UAV-acquired dataset of 11,474 images from Michigan and Minnesota forests was used, with 759 images expert-labeled as “Oak Wilt” or “Not Oak Wilt.” The dataset captures varied canopy structures, seasonal changes, and lighting conditions. Data augmentation techniques such as random cropping, rotation, flipping, and normalization were applied. Models were trained for 25 epochs with AdamW (learning rate 3e-4, weight decay 1e-2) and optimal input sizes (224×224 or 256×256).

Among tested models, SwinV2-Tiny achieved the best results—98.68% Accuracy, 98.61% Precision, Recall, and F1—attributed to its hierarchical window-based attention mechanism capturing both local and global canopy features. Other strong performers included BEiT V2 (F1: 91.30%), ViT-B16 (F1: 91.67%), and Swin-Tiny (F1: 91.85%). Hybrid networks such as ConvNeXt (F1: 89.86%) and DenseNet201 (F1: 91%) also performed well. Traditional CNNs like ResNet-101 and ResNet-50 achieved F1-scores of 85% and 83% respectively, while MobileNet, though computationally efficient, exhibited lower precision due to overprediction in complex backgrounds. A Flask and VueJS-based web interface was developed for real-time predictions and geospatial mapping, designed for Michigan DNR deployment. The study also outlines future enhancements, including integrating Reinforcement Learning from Human Feedback (RLHF) for adaptive retraining.

This work demonstrates the potential of transformer-based architecture for precision forestry, providing a foundation for broader invasive pest detection under accelerating climate change.

Comments

For Further Discussions regarding this topic, please send an email at bismoym@mail.gvsu.edu

Download

Share

COinS