University of Southampton
GlobeNet & GBNet: Deep self-supervised learning for global and national remote sensing applications
Abstract
dc:description.abstractMaintaining the topographic data of National Mapping Agencys (NMAs) is a consistent challenge in a changing world, especially for the NMA of Great Britain , the Ordnance Survey (OS), who are charged with detecting any topographic changes. Despite concerted efforts, this is still a largely manual process involving experts pouring over aerial imagery and surveyors being sent out into the field to map these changes. The leaps in capability offered by Deep Learning (DL) over the last ∼15 years presents an opportunity to alleviate this burden. One such innovation has been a form of Self-Supervised Learning (SSL): Siamese Learning. As there is often an abundance of imagery in the geospatial domain but a lack of corresponding verification data, Siamese learning presents an ideal opportunity to train a geo-foundation model to extract topographic features on a global scale of data. This work introduces GlobeNet: a remote-sensing focussed, RGB-NIR, fully-convolutional, semantic segmentation based approach to Siamese learning. Trained on the SSL4EO-S12 [1] curation of Sentinel-2 imagery, downstream evaluation across a spectrum of Remote Sensing (RS) tasks<br/>suggested GlobeNet had a performance on par with ImageNet pre-trained weights. Examination of qualitative results suggests the GlobeNet proposal produces encoders capable of superior predictions when part of a semantic segmentation model. Separate work on OS’s proof-of-concept approach to geo-foundation models, GBNet, is also presented. By leveraging expertly curation and very high quality aerial imagery, GBNet outperforms ImageNet by 3.4pp ± 0.3pp ( linear-probing) / 7.3pp ± 0.6pp ( fine-tuning) on overall mIoU score. Probable evidence of the effect of data quality on the stability of Siamese learning is also presented, and arguments towards greater focus and quantification of data quality made. Finally, the code created and used in this work is introduced and can be accessed in the minerva repository on GitHub.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Baker, Harry James
- Advisors dc:contributor.advisor
-
- Hare, Jonathon
- Sargent, Isabel
- Prugel-Bennett, Adam
- Coupland, Steven
- ORCID person:identifier.orcid
Rights
- Language dc:language
- English, eng
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5258/SOTON/PG/T267