Monitoring Coastal Environments using UAS Imagery and Deep Learning

Hobley, Brandon (2023) Monitoring Coastal Environments using UAS Imagery and Deep Learning. Doctoral thesis, University of East Anglia.

Preview

PDF Download (205MB) | Preview

Abstract

Coastal monitoring is a complex mapping problem for environments that exhibit distinct physical variations through the energy expended from water and sediment movement. In recent years, sensor platforms that capture imagery from these environments have reached centimeter level pixel resolution, which allowed object-based image processing methods to become a standard mapping tool. However, this tool still adheres to shallow machine learning methods, whereby the construction of a learning system is broken into two steps: feature extraction and machine learning model optimisation.

In the last decade, deep learning and convolutional neural networks have established state-of-the-art performance on a myriad of computer vision applications. However, deep learning models perform best with large, labelled, training datasets. For coastal monitoring, ground-truth observations can be acquired either in-situ or through post-processed imagery, but both avenues require manual process in producing the ground-truth annotations. In turn, this requires laborious and expensive efforts with domain expertise of coastal processes, posing a bottleneck and challenge for accurate coastal monitoring.

In this thesis, practical applications of coastal monitoring using deep learning and convolutional neural networks are discussed. These methods attempt to improve the performance and generalisation of convolutional neural networks with limited amounts of labelled data, which could ease costs of producing ground-truth annotations. A number of approaches are described that reduce the effort required to produce them, or analyse the feasibility of non-domain expert labels.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Chris White
Date Deposited:	12 Jun 2024 08:43
Last Modified:	12 Jun 2024 08:43
URI:	https://ueaeprints.uea.ac.uk/id/eprint/95589
DOI:

Downloads

Actions (login required)

View Item