Skip to content

Sensor-agnostic coastline delineation using foundational models

Type of project

M.Sc.

Contact

Nicklas Simonsen (nisi@dhigroup.com)

Topic

Remote sensing provides a unique opportunity to observe coastlines at a large scale and throughout history to analyze changes in shape and structure. These coastlines may be used to assess stability, the effect of coastal protection and/or nature-based solutions, the impact of land cover changes, etc.  

Several tools currently exist to extract coastlines from single cloud-free images from specific sensors and resolutions. However, most of these do not generalize well outside of very specific domains or provide accurate results in the presence of clouds. Additionally, cloud-masking algorithms typically do not capture clouds very well close to the water and often incorrectly remove bright sandy shores. 

Foundational models are generalized across a vast number of different tasks and sensors and have shown improved accuracy over task-specific models when only a limited number of labels are available. This project proposes utilizing these foundational models to create a robust sensor-agnostic model that can accurately separate land, clouds, and water in marine environments globally.

Relevance to DHI

The DHI earth observation center of excellence is often applying machine learning and deep learning methods to map bathymetry, coastlines and other coastal habitats. This project would explore the newest state of the are models and techniques to improve current workflows and products.

What is the student's contribution

  • Catalogue available labeled datasets.
  • Explore the capabilities of different foundational models for geospatial computer vision.
  • Develop a classification head that can accurately segment land, clouds, and water across different sensors.
  • Optional: Train and test the applicability of VHR sensors (GSD 0.3-1m).

References