Dylan Niner

Date of Award

2025

Document Type

Thesis

Degree Name

Bachelors

Department

Natural Sciences

First Advisor

Toro-Farmer, Gerardo

Area of Concentration

Marine Biology

Abstract

Coastal ecosystems are essential for biodiversity and ecological services; they are increasingly under threat from human activity. Within these systems, macroalgae (submerged aquatic vegetation, or SAV) play a vital role. For that reason, assessing the condition and presence of SAV is important for understanding coastal ecosystem health. Traditional field surveys for SAV are limited in spatial and temporal resolution. In contrast, high-resolution Aerial (AUV) imagery, when combined with Convolutional neural networks, provides a great opportunity for analyzing this critical habitat at the necessary scales. Neural networks are computer models of interconnected nodes that learn to map inputs to outputs by adjusting connection weights during training. However, the challenge of analyzing the resulting large datasets presents a bottleneck for many potential coastal applications. A dataset was compiled from Aerial imagery captured over Lido Key, Sarasota, Florida. Images were preprocessed into standardized patches and manually labeled and sorted into three categories: SAV, Sand/Water, and Terrestrial Vegetation . The performance of two neural network architectures was comparatively evaluated for this 3-class classification task: a baseline Multi-Layer Perceptron (MLP) and a Convolutional Neural Network (CNN) inspired by the LeNet-5 architecture with modern enhancements. The results showed that the MLP was severely overfitting during training and had poor generalization to the unseen test set (Accuracy ≈ 0.582). This meant that the MLP was not suitable for reliable SAV detection. The opposite was true for the modified LeNet-5 CNN; it performed effectively on the test set (Accuracy ≈ 0.997). For the target class, 'SAV,' the CNN had perfect precision (1.00) and near-perfect recall (0.99). This means that the modified LeNet-5 CNN with transfer learning was extremely good at accurately identifying and detecting patches containing SAV with few false positives and negligible false negatives. The findings underscore the ability of convolutional neural networks (CNNs) to better utilize the spatial nature of image data than multilayer perceptrons (MLPs) for this ecological task. The CNN developed here is a presence detection filter that can significantly reduce the effort needed to screen the numerous standard Aerial images for the presence of coastal macroalgae. This automated step can direct the relevant expert’s attention to the right places and make the workflow much more efficient and scalable.

Download

Share

COinS