PRACTICAL EMBEDDING WORKFLOWS WITH TERRATORCH, THE GEOSPATIAL FINE-TUNING TOOLKIT

Romeo Kienzler, Isabelle Wittmann, Paolo Fraccaro

Abstract:

Geospatial foundation models pretrained on large-scale Earth observation archives
offer strong transfer capabilities across remote sensing tasks, but practical adop-
tion remains challenging due to heterogeneous data formats, complex fine-
tuning pipelines, and inconsistent evaluation protocols. We present TerraTorch,
a configuration-driven toolkit for reproducible adaptation and benchmarking of
geospatial foundation models.
This workshop contribution complements earlier TerraTorch system work by fo-
cusing specifically on embedding-centric workflows: (i) generic embedding gen-
eration from pretrained encoders and (ii) downstream learning on top of frozen
embeddings for semantic segmentation. All demonstrations are linked to ex-
ecutable repository examples, lowering the barrier for ML4RS researchers and
practitioners to apply foundation models in real-world Earth observation settings.

Submission Number: 6