Project description
Improving coastal mapping with deep learning
Much of the ocean floor is unmapped, and many shallow coastal areas that are most impacted by climatic and anthropogenic pressures are poorly updated. The EU-funded MagicBathy project hopes to fix this through a new deep learning algorithm that can make better use of satellite and unmanned aerial vehicles (UAV) imagery. Currently, the optics of water can affect UAV image quality, and correction is expensive. Satellite images can suffer from low resolution. The MagicBathy project will use machine learning to correct the spatial resolution of satellite images and the bathymetric maps that result from them. Researchers will also develop a special resolution improvement algorithm for shallow water imagery.
Objective
Accurate, detailed and high-frequent bathymetry, coupled with the important visual and semantic information, is crucial for the undermapped shallow coastal areas being affected by intense climatological and anthropogenic pressures. Regular UAV and satellite imagery have the potential to frequently and consistently map those areas to different extents and detail, providing ground breaking key information. However, optical properties of water severely affect images and refraction is the main factor affecting their geometry. Current Structure from Motion (SfM) based solutions for refraction correction are slow and costly. Satellite Derived Bathymetry (SDB) methods deliver faster results over huge shallow areas albeit in lower spatial resolution, failing to handle non-homogeneous seabeds. Recent methods based on Convolutional Neural Networks (CNNs) deliver either only the bathymetry or the semantics of the scene, tackling those problems separately and in one scale/modality at a time. They are mostly dedicated to satellite images, failing to address the challenges of shallow waters, being also inefficient for UAV images, preventing higher resolution results. MagicBathy will establish an advanced deep learning framework for low-cost shallow water mapping by developing a novel boundary-aware multitask, multiscale and multimodal learning approach for bathymetry and semantics together, exploiting single either UAV or satellite imagery. To overcome the domain gap, generalize and improve performance, self-supervised in-domain representation learning will be performed. To enhance the spatial resolution of low resolution satellite images and hence of the resulting bathymetric/semantic maps, a conditional generative adversarial network (cGAN)-based Super Resolution framework will be developed, dealing with the special challenges of shallow water imagery. Frameworks, models and results will be published in open access, enabling the rapid progress in shallow water mapping worldwide
Fields of science
- engineering and technologymechanical engineeringvehicle engineeringaerospace engineeringsatellite technology
- natural sciencescomputer and information sciencesartificial intelligencegenerative artificial intelligence
- natural sciencescomputer and information sciencesartificial intelligencemachine learningdeep learning
- natural sciencesmathematicspure mathematicsgeometry
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Keywords
Programme(s)
- HORIZON.1.2 - Marie Skłodowska-Curie Actions (MSCA) Main Programme
Funding Scheme
HORIZON-TMA-MSCA-PF-EF - HORIZON TMA MSCA Postdoctoral Fellowships - European FellowshipsCoordinator
10623 Berlin
Germany