Have Foundation Models Seen Satellite Images?
Investigating the Generalization of Pre-Trained Models on Geospatial Data
This study evaluates the zero-shot performance of pre-trained foundation models on remote sensing tasks. The research analyzes whether these models have previously encountered satellite imagery and assesses their adaptability to standard benchmarks like EuroSAT and BigEarthNet-S2. Additionally, it explores the impact of geospatial domain-specific textual descriptions compared to standard class-based prompts.
- Assesses foundation models' zero-shot classification abilities in remote sensing.
- Compares standard class-based prompts with geospatial domain-specific descriptions.
- Evaluates model performance on EuroSAT and BigEarthNet-S2 datasets.
Zero-Shot Learning
Foundation Models
Remote Sensing