Strategic Data Sampling for Spatial Generalization of Earth Observation Foundation Models

Deep learning has transformed Earth Observation research, but the scarcity and uneven distribution of labeled data remain fundamental bottlenecks. Geospatial foundation models circumvent this by pre-training on large unlabeled datasets to extract broadly useful features, then fine-tuning on (limited) labeled data for specific tasks. However, how small labeled datasets should be chosen, and whether models generalize to unseen regions, remains underexplored. We address this through a series of ablation studies simulating label-scarce scenarios, and assess how well models generalize. We first pre-train a vision transformer-based masked autoencoder on Sentinel-1 Synthetic Aperture Radar data to learn feature-rich embeddings, and fine-tune it for terrestrial downstream tasks. We use vegetation prediction as our primary case study, and show that pre-training reduces the requirement for labels by >90% compared to training from scratch, highlighting that high-quality predictions can be achieved with very few labels. We also investigate sampling strategies for selecting small labeled datasets, comparing random sampling against stratified approaches based on land cover, elevation, climate, or diversity of embeddings. We find that stratified sampling using task-relevant characteristics strongly improves cross-region generalization compared to random sampling. We also show that global sampling, even with small regional footprints, yields better cross-region generalization than models trained with 100 times more data from a single region. Our findings suggest that when labeling resources are limited, even simple, interpretable sampling schemes can significantly improve model robustness and spatial generalization. Furthermore, this demonstrates that intelligent data sampling is critical for deploying Earth Observation foundation models in operational settings.