Thesis subject
MSc thesis topic: Implicit Neural Representation of the Earth Gravity Field
The Earth is a potatoe: The Earth’s gravity field is not homogeneous and varies by about 80 meters compared to the WGS80 reference ellipsoid. These Geoid undulations are determined from satellite missions and stored as spherical harmonic coefficients. Calculating spherical harmonics up to a high resolution (i.e., many harmonics) can be computationally expensive. Storing the same geospatial information in neural networks, i.e., location encoders (Rußwurm et al., 2024), as an “implicit neural representation” can be a computationally more efficient alternative. This Master thesis will evaluate the potential of storing geospatial data like spherical harmonics in neural networks.
The core research question of this thesis will be: Can a neural network location encoder produce a sufficiently similar geoid representation than a classic weighted spherical harmonic representation by combining spherical harmonics with sinusoidal representation networks (SIREN). This will involve reproducing the Earth Gravitational Model 2008 (EGM2008) (degree of 2160) and training a location encoder with fewer harmonics in the positional encoder and a small (SIREN) neural network as a surrogate model. Further questions involve how to implement uncertainties (EGM harmonic coefficients are given with variances) and a temporal component.
Relevance to research/projects at GRS or other groups
This Thesis is supervised by Marc Rußwurm who will provide expertise in machine learning and deep location encoding.
Objectives
- Reproduce the Earth Gravitational Model 2008 (EGM2008) in python by using the public spherical harmonic coefficients
- Train a surrogate neural network model to reproduce the values of the EGM 2008 model
- Compare the computational efficiency and identify potentials and shortcomings of storing Geoid information in neural networks
Requirements
- recommended: deep learning course
- recommended: interest in data representation
- recommended: interest in math and Geodesy
Expected reading list before starting the thesis research
- Rußwurm, M., Klemmer, K., Rolf, E., Zbinden, R., & Tuia, D. (2024). Geographic location encoding with spherical harmonics and sinusoidal representation networks. International Conference on Learning Representations (ICLR).
- Genchen Mai et al., (2024). A Review of Location Encoding for GeoAI: Methods and Applications
- https://en.wikipedia.org/wiki/Earth_Gravitational_Model
Theme(s): Modelling & visualisation