Background: Satellite data are usually stored as a grid with one value per pixel. Implicit neural representations (INRs) instead treat the data as a function of location and time and train a neural network to learn it. You give the network a location and a date, and it returns the values for that place and time. This can save a lot of storage, but the network only stores an approximation. Its quality is usually judged by a single number, the average reconstruction error. What this number hides, which information is lost and where, is largely unknown.
Source:
https://gkvoelkl.github.io/python-streamlit-neural-earth-field/en/01-geodaten-als-funktion.html |
Short
description: The thesis trains an INR on a Sentinel-2 time
series of a chosen region and compares the reconstruction with the original
data, pixel by pixel and date by date. The aim is to find out what gets lost.
Possible candidates are sudden changes (e.g., a meadow mown on a specific date
that shows a sharp drop in NDVI), small objects such as single buildings or
ponds or extreme values such as the peak of the growing season. The result is a
clear picture of which information an INR keeps and which it does not.
Suggested reading
Sitzmann, V., Martel, J., Bergman,
A., Lindell, D., & Wetzstein, G. (2020). Implicit
neural representations with periodic activation functions. Advances in Neural
Information Processing Systems, 33, 7462–7473.
Madadikhaljan,
M., Prexl, J., Wittmann, I., Albrecht, C. M., & Schmitt, M. (2026). Location
Is All You Need: Continuous spatiotemporal neural representations of Earth
observation data. arXiv:2604.07092.
https://gkvoelkl.github.io/python-streamlit-neural-earth-field/en/01-geodaten-als-funktion.html
Start: Anytime
Prerequisites/qualification: This
topic is technically demanding and suits students who enjoy coding and have
trained neural networks before.
·
Remote
Sensing
·
Very good Python skills and
experience with PyTorch
·
Experience training neural networks
No comments:
Post a Comment