About us

In Earth Observation (EO), large amount of satellite data is available and can be used for many applications such as biomass estimation, crop type classification, or crop yield prediction. These tasks benefit from a series of satellite images with different timestamps, to observe e.g. plant dynamics, to improve the prediction performance. However, downloading long time series of images for an entire region and to process them with AI models, such as IBM’s Prithvi EO model [ 1 ], is challenging due to the increasing costs and volume for data transfer and compute.

Neural compression [ 2 ] can mitigate these challenges by creating compressed embedding representations of the satellite data. A data provider can encode satellite images using pre-trained compression models. Users would only have to download the compressed embeddings and run lightweight decoders for their tasks. So far, only individual images have been compressed, which would require late-fusion approaches in the decoder for temporal tasks. Including temporal aggregation in the encoding process could further improve the compression and additionally encode temporal changes. General video compression models already deal with many images as input and their aggregation. This type of models can be the base-line for EO applications.

Goal

In this master thesis, you will combine the concept of temporal aggregation of satellite images with neural compression enabling the sharing of embeddings to reduce on data transfer time and cost. Specifically, you will (1) perform a solid review of existing video models and video/EO compression literature, (2) develop of a multi-temporal compression model, and (3) evaluate the model on multi-temporal downstream tasks.
The work is an important step to enable embedding sharing in EO which can improve many environmental applications.

Requirements

• Enrolled in a master' program in computer science, scientific computing or similar disciplines
• Python, PyTorch, GitHub/Lab
• Problem-solving mindset: ability to explore technical literature, cross-referencing relevant information, synthesis of novel ML architectures
• Ways to stand out: hands-on experience with HPC systems and job scheduling, prior encounters with large-scale transformer architectures, prior experience with multi-temporal data and/or compression algorithms
  
  

Diversity

IBM is committed to diversity at the workplace. With us you will find an open, multicultural environment. Excellent flexible working arrangements enable all genders to strike the desired balance between their professional development and their personal lives.

How to apply

If you are interested to apply for this position, please submit your application through the link below.