Paper Type

Master's Thesis


College of Computing, Engineering & Construction

Degree Name

Master of Science in Computer and Information Sciences (MS)



NACO controlled Corporate Body

University of North Florida. School of Computing

First Advisor

Dr. Xudong Liu

Second Advisor

Dr. Indika Kahanda

Third Advisor

Dr. Sandeep Reddi

Fourth Advisor

Dr. Karthikeyan Umapathy

Department Chair

Dr. Sherif Elfayoumy

College Dean

Dr. William F. Klostermeyer


A marsh is a wetland dominated by various species of grasses rather than trees. It supports an important ecosystem, providing habitats for many kinds of animals and posing a crucial impact on coastal climates. Marshes are rapidly changing, and it is vital for scientists to track these changes to understand the health of them. To do so, biologists perform vegetation monitoring to estimate the coverage of vegetation in an area of marsh. This task often calls for extensive human labor carefully examining pixels in photos of marsh sites to calculate vegetation density for a variety of marsh grass species. This is a very time-consuming process. In this thesis work, I designed a computational framework that automatically predicts the vegetation density using deep learning models, in particular, convolutional neural networks (CNNs). Next, to obtain data for the purpose of training and testing of the predictive models, I developed a labeling software and used it to collect labeled image snippets from photoquadrats provided by biologists at Guana Tolomato Matanzas Estuarine Research Reserve (GTMNERR). With a set of university volunteers, 77,530 labeled images were collected. Using this collected dataset, I experimented with CNNs, such as LeNet-5, AlexNet, and VGG-16, to train effective models for predicting whether an input snippet is (1) unvegetated or not, (2) one of the five considered marsh grass species, or (3) either unvegetated or one of the five vegetated classes for a total of six classes. These models were applied in two variant frameworks: a single model framework with only the six-class CNN model, and a chained classifier that first classifies a point as vegetated or unvegetated and then classifies it as belonging to a specific species if it is vegetated. We found the chained VGG-16 classifier to be the most accurate with a test accuracy of 84%. Finally, integrating the pre-trained CNN models in the back-end, I implemented MarshCover, a web-based system that automates the entire process of monitoring vegetation density.