When GeoFMs Meet VHR: Benchmarking Foundation Models on Very High Resolution Earth Observation Imagery

TL;DR

The study introduces VHR-Bench, a benchmark for very high resolution Earth observation imagery at 1.5m GSD or finer, evaluating twelve encoders across three pretraining tiers. Results show that both natural-image and VHR-specific pretraining transfer well, with SkySense excelling under label scarcity and DINOv3 ViT-L outperforming at full labels. Continuing DINOv3 ViT-L's self-supervised pretraining on VHR data improves performance without increasing inference cost.

Why It Matters

Many real-world remote sensing applications depend on very high resolution imagery, yet most foundation models are pretrained on coarser data. This work provides a standardized evaluation framework and demonstrates that enhancing general-purpose models with a short VHR-specific pretraining phase yields measurable gains, enabling better performance with minimal additional cost.

Key Findings

• Pretraining on large natural-image datasets transfers effectively to VHR tasks, especially when labels are limited.
• Encoders pretrained solely on Sentinel-scale imagery perform worse on VHR data under RGB input.
• Continuing self-supervised pretraining of DINOv3 ViT-L on 1.02 million VHR tiles improves performance across label fractions without increasing inference complexity.