Student: “Myrto Potamiti”

Program: “Data Science and Information Technologies”

Title: “MSTOK: Adapting Discrete Visual Tokenization to Multispectral Earth Observation Imagery”
Abstract:
Recent advances in generative modelling for Earth observation highlight the importance of compact image representations that preserve spatial and spectral information. This thesis presents MSTok, an adaptation of the Open-MAGVIT2 visual tokenizer using lookup-free quantization (LFQ) for 12-band Sentinel-2 Level-2A imagery from TerraMesh. Experiments investigate reflectance preprocessing, discrete code-space capacity, ImageNet-pretrained initialization, architectural adaptations, and joint ImageNet–TerraMesh training. Reconstruction quality is assessed
on 89,088 validation samples at 224×224 resolution using multispectral and RGB perceptual metrics. After one epoch, the standard model trained from scratch outperformed its
ImageNet-initialized counterpart across the reported reconstruction metrics, achieving an FID of 61.38 versus 69.53. A recurring trade-off emerged between multispectral accuracy and RGB perceptual quality: extending linear-adapter training reduced all-band MAE from 0.0480 to 0.0469, while increasing FID from 29.26 to 39.18. High code utilization and accurate standalone adapter reconstruction did not ensure effective end-to-end tokenization. The from-scratch configuration was retained as the preferred overall model within the available training budget.
These findings demonstrate the feasibility of LFQ-based multispectral tokenization and the need to assess multispectral reconstruction fidelity and visible-band perceptual quality jointly.

Date/Time: October 16, 2026 – 14:00 PM.
Examination Committee:
”
Dr. Bill Psomas
Dr. Stavros Perantonis
Dr. Ioannis Kakogeorgiou
“
Presentation link:
https://meet.google.com/dvs-gzfy-nkq