Date of Award
Doctor of Philosophy (PhD)
Electrical & Computer Engineering
Chung Hao Chen
Seagrass forms the basis for critically important marine ecosystems. Seagrass is an important factor to balance marine ecological systems, and it is of great interest to monitor its distribution in different parts of the world. Remote sensing imagery is considered as an effective data modality based on which seagrass monitoring and quantification can be performed remotely. Traditionally, researchers utilized multispectral satellite images to map seagrass manually. Automatic machine learning techniques, especially deep learning algorithms, recently achieved state-of-the-art performances in many computer vision applications. This dissertation presents a set of deep learning models for seagrass detection in multispectral satellite images. It also introduces novel domain adaptation approaches to adapt the models for new locations and for temporal image series. In Chapter 3, I compare a deep capsule network (DCN) with a deep convolutional neural network (DCNN) for seagrass detection in high-resolution multispectral satellite images. These methods are tested on three satellite images in Florida coastal areas and obtain comparable performances. In addition, I also propose a few-shot deep learning strategy to transfer knowledge learned by DCN from one location to the others for seagrass detection. In Chapter 4, I develop a semi-supervised domain adaptation method to generalize a trained DCNN model to multiple locations for seagrass detection. First, the model utilizes a generative adversarial network (GAN) to align marginal distribution of data in the source domain to that in the target domain using unlabeled data from both domains. Second, it uses a few labeled samples from the target domain to align class-specific data distributions between the two. The model achieves the best results in 28 out of 36 scenarios as compared to other state-of-the-art domain adaptation methods. In Chapter 5, I develop a semantic segmentation method for seagrass detection in multispectral time-series images. First, I train a state-of-the-art image segmentation method using an active learning approach where I use the DCNN classifier in the loop. Then, I develop an unsupervised domain adaptation (UDA) algorithm to detect seagrass across temporal images. I also extend our unsupervised domain adaptation work for seagrass detection across locations. In Chapter 6, I present an automated bathymetry estimation model based on multispectral satellite images. Bathymetry refers to the depth of the ocean floor and contributes a predominant role in identifying marine species in seawater. Accurate bathymetry information of coastal areas will facilitate seagrass detection by reducing false positives because seagrass usually do not grow beyond a certain depth. However, bathymetry information of most parts of the world is obsolete or missing. Traditional bathymetry measurement systems require extensive labor efforts. I utilize an ensemble machine learning-based approach to estimate bathymetry based on a few in-situ sonar measurements and evaluate the proposed model in three coastal locations in Florida.
Islam, Kazi A..
"Deep Learning Approaches for Seagrass Detection in Multispectral Imagery"
(2021). Doctor of Philosophy (PhD), Dissertation, Electrical & Computer Engineering, Old Dominion University, DOI: 10.25777/gct9-yr76