Earth observation satellites generate vast amounts of multi-temporal imagery, making storage and transmission a major challenge. This thesis investigates how well different compression methods handle remote-sensing time series — from traditional approaches (JPEG2000, 3D wavelets) to modern video codecs (HEVC, AV1) and recent neural compression methods. The goal is to compare them in terms of compression ratio, speed, and reconstruction quality, and assess their impact on downstream tasks such as land-cover change detection.

Possible tasks include:

• Implementing and benchmarking classical, video-based, and neural compression techniques.

• Evaluating compression factor, encoding/decoding speed, and fidelity (PSNR, SSIM, spectral metrics).

• Analyzing trade-offs and recommending suitable approaches for different use cases.

Requirements for students:

• Solid Python skills and experience with machine learning (PyTorch preferred).

• Familiarity with remote sensing data handling (e.g., Sentinel-2) or motivation to learn.

• Interest in running large-scale experiments and comparing modern compression methods.

In case you are interested, please contact jan.pauls@uni-muenster.de