This website stores cookies on your computer. These cookies are used to collect information about how you interact with our website and allow us to remember your browser. We use this information to improve and customize your browsing experience, for analytics and metrics about our visitors both on this website and other media, and for marketing purposes. By using this website, you accept and agree to be bound by UVic’s Terms of Use for web and social media privacy.  If you do not agree to the above, you can configure your browser’s setting to “do not track.”

Skip to main content

Piper Steffen

  • BSc (Queen’s University, 2022)

Notice of the Final Oral Examination for the Degree of Master of Science

Topic

Satellite-based coastline extraction across diverse coastal environments and low tide conditions in British Columbia, Canada

Department of Geography

Date & location

  • Tuesday, December 1, 2026

  • 12:00 P.M.

  • Virtual Defence

Reviewers

Supervisory Committee

  • Dr. Maycira Costa, Department of Geography, University of Victoria (Supervisor)

  • Dr. Eva Kwoll, Department of Geography, UVic (Member) 

External Examiner

  • Dr. Tom Bell, School of Marine Science and Technology, UMass Dartmouth 

Chair of Oral Examination

  • Dr. Kieka Mynhardt, Department of Mathematics and Statistics, UVic 

Abstract

Accurate, scalable methods for extracting coastlines from optical satellite imagery are essential for monitoring coastal habitats, coastal erosion, disaster management, and sea-level rise. However, the performance of established coastline extraction approaches in geomorphologically complex, heterogeneous coastal environments remains uncertain, as prior studies have largely focused on sandy beaches. In this thesis, we present the first low tide satellite-derived coastline (LT-SDC) of the Pacific coast of British Columbia (BC), Canada, demonstrating a scalable coastline extraction method applicable to complex coastal environments using Sentinel-2 imagery. For this, we evaluate three deep-learning based coastline extraction models across a wide range of coastal landforms and apply the top-performing model to low tide image mosaics of the BC coast to define LT-SDC position, with an associated spatial confidence rating informed by coastal geomorphology. 

The coastline extraction models consist of a convolutional neural network (CNN) with a U-Net architecture trained for per-pixel water/non-water classification, from which the boundary between the classes is defined to represent instantaneous coastline position. We compare the performance of 1) the BC model, trained on a new, manually annotated dataset, the BC dataset, specific to the BC coast and encompassing a diverse range of coastal geomorphologies; 2) the Global model, trained on the globally distributed Sentinel-2 Water Edges Dataset (SWED); and 3) the BC-FT model, leveraging both datasets through finetuning-based transfer learning. Models are evaluated in terms of standard segmentation performance metrics to assess CNN classification ability, and coastline position error (CPE) to measure the difference between a model output and reference coastline across geomorphologically diverse validation sites. Although the BC-FT model achieved marginally higher IOU than the BC model (0.988 and 0.987, respectively), the BC model demonstrated superior performance in terms of CPE, with an RMSE of 7.71 m compared to an RMSE of 9.82 m for the BC-FT model, indicating that negative transfer may have diminished the performance of the BC-FT model. Conversely, the Global model performed substantially worse, with an IOU of 0.927 and an RMSE of 26.63 m, corroborating the challenges of developing geographically generalizable deep learning algorithms established in prior studies. The BC model is therefore selected as the top-performing model and evaluated according to geomorphological factors of shore material, exposure, and slope. Rocky, protected to semi-exposed coastlines of all slopes resulted in the lowest CPE (RMSE = 5.57 m), whereas sediment, exposed, inclined coastlines resulted in the highest CPE (RMSE = 14.18 m). 

Finally, the BC model is applied to cloud-free, low tide Sentinel-2 images covering the BC coast, selected using Google Earth Engine and a global tide model to filter 8 years of available imagery, to extract a LT SDC at the provincial scale. A spatial confidence rating indicating relative uncertainty in SDC position is informed by coastal geomorphology, with the highest confidence assigned to predominantly protected, rock or mixed rock and sediment coastlines, and the lowest confidence assigned to exposed sediment coastlines with gradual slopes. The resulting LT-SDC undergoes quality control to remove erroneous coastlines and clearly visible artifacts, such as those arising from inland lakes, watercraft, and terrain shadows. The final LT-SDC of coastal BC, measuring over 36,000 km, provides a robust, reliable spatial data product to help coastal managers, researchers, and communities monitor coastal ecosystems in BC and demonstrates how deep learning-based coastline extraction models trained on local data can be applied effectively to complex coastal regions at broad scales.