Chromik, Niklas (2026) Benchmarking Any-Sensor Foundation Models for Earth Observation. Masterarbeit, Ostbayerische Technische Hochschule Regensburg.
Dieses Archiv kann nicht den Volltext zur Verfügung stellen.
Kurzfassung
Early foundation models for Earth observation were tied to fixed sensor configurations. More recently, sensor-agnostic models have emerged that promise to support arbitrary sensors without retraining. So far, this claim has mostly been validated by the models' authors themselves under inconsistent evaluation protocols. We conduct a unified evaluation of six publicly available sensor-agnostic models and simple baselines using frozen encoders across 23 tasks and 11 sensors covering multispectral, synthetic aperture radar and hyperspectral data. We find that Panopticon is the only model that performs consistently well across all modalities. However, the models show no consistent advantage over a supervised UNet or randomly initialized encoders. Experiments with modified input bands and corrupted wavelength metadata further reveal that DOFA and DOFA+ are largely insensitive to incorrect metadata and do not consistently benefit from additional spectral bands, which raises questions about whether their hypernetworks meaningfully interpolate to unseen wavelengths. We also find that decoder choice changes the model rankings, while differences in patch size and pretraining data limit the interpretation of performance differences between models. These findings underscore the need to evaluate sensor-agnostic foundation models across diverse sensors and against simple baselines, and to test explicitly whether their mechanisms respond to changes in spectral inputs and sensor metadata.
| elib-URL des Eintrags: | https://elib.dlr.de/228744/ | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Dokumentart: | Hochschulschrift (Masterarbeit) | ||||||||||||
| Titel: | Benchmarking Any-Sensor Foundation Models for Earth Observation | ||||||||||||
| Autoren: |
| ||||||||||||
| DLR-Supervisor: |
| ||||||||||||
| Datum: | 2026 | ||||||||||||
| Open Access: | Nein | ||||||||||||
| Seitenanzahl: | 77 | ||||||||||||
| Status: | veröffentlicht | ||||||||||||
| Stichwörter: | self-supervised learning, foundation models, earth observation, benchmarking, hyperspectral imagery | ||||||||||||
| Institution: | Ostbayerische Technische Hochschule Regensburg | ||||||||||||
| HGF - Forschungsbereich: | Luftfahrt, Raumfahrt und Verkehr | ||||||||||||
| HGF - Programm: | Raumfahrt | ||||||||||||
| HGF - Programmthema: | Erdbeobachtung | ||||||||||||
| DLR - Schwerpunkt: | Raumfahrt | ||||||||||||
| DLR - Forschungsgebiet: | R EO - Erdbeobachtung | ||||||||||||
| DLR - Teilgebiet (Projekt, Vorhaben): | R - Optische Fernerkundung, R - Fernerkundung u. Geoforschung, R - Künstliche Intelligenz | ||||||||||||
| Standort: | Oberpfaffenhofen | ||||||||||||
| Institute & Einrichtungen: | Institut für Methodik der Fernerkundung > EO Data Science | ||||||||||||
| Hinterlegt von: | Banze, Aaron | ||||||||||||
| Hinterlegt am: | 10 Nov 2026 15:02 | ||||||||||||
| Letzte Änderung: | 10 Nov 2026 15:02 |
Nur für Mitarbeiter des Archivs: Kontrollseite des Eintrags