E-Thesis 17 views 6 downloads
Mass labelling Remote Sensing Images: A visualisation and machine learning combinatory approach / TULSI PATEL
Swansea University Author: TULSI PATEL
-
PDF | E-Thesis – open access
Copyright: the author, Tulsi Patel, 2026. Distributed under the terms of a Creative Commons Attribution 4.0 License (CC BY 4.0).
Download (34.11MB)
DOI (Published version): 10.23889/SUThesis.71617
Abstract
Deep learning and computer vision have made significant advancements in the field of the automated analysis of Remote Sensing imagery. However, in order to maximise the utility of satellite derived Remote sensing imagery, numerous challenges must be addressed. Firstly, the complexity of the Earth’s su...
| Published: |
Swansea
2025
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Jones, M. W. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa71617 |
| Abstract: |
Deep learning and computer vision have made significant advancements in the field of the automated analysis of Remote Sensing imagery. However, in order to maximise the utility of satellite derived Remote sensing imagery, numerous challenges must be addressed. Firstly, the complexity of the Earth’s surface, in conjunction with its atmosphere, introduces variation within data, requiring expert knowledge to interpret, label and ensure accuracy. Secondly, a large volume of imagery data is generated every day from a single sensing instrument and there are multiple instruments continuously monitoring the earth’s surface. Only a very small percentage of data has been accurately labelled, and that which has been is targeted towards a specific domains and of limited temporal resolution. In this thesis, we identify the need for and explore a new remote image labelling framework. We utilise visualisation, unsupervised ma-chine learning and remote sensing methodologies to enable users to create large-scale, labelled remote sensing datasets. Our approach combines three main components - Deep convolutional auto-encoders, manifold learning and an interactive data labelling application. The deep convolutional auto-encoder learns a condensed and informative representations satellite imagery.The manifold learning condenses this representation down into two dimensions, which then supports visualisation techniques that help to convey patterns and representations to a labeller.Graph neural networks are then introduced to further enhance the spatial encoding of geographical features within imagery, and the use of super-pixel representations allow users to create full segmentation labels. Throughout the thesis, we present an evolving visualisation application in which we explore the feature encodings provided by our deep learning pipelines. We introduce a novel methodology for branching and merging datasets, providing a more fine-grained and expressive labelling experience for users. Overall, this thesis presents novel and innovative methods for labelling remote sensing images, using techniques from visualisation, computer vision and the remote sensing domains. |
|---|---|
| Keywords: |
Remote Sensing, Manifold Learning, Visualisation, Labelling |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC Doctoral Training Grant |