Geospatial models pretrained on very high resolution images improve remote sensing tasks

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

Computer Vision and Pattern Recognition

Summary

Satellite images are used to study the Earth, but most computer models are trained on lower-resolution images that blur important details. The authors checked how well different pretrained models work on very high resolution images that show details like buildings and roads. They found some models trained on general photos and some on high-res images each have strengths. By continuing to train a strong model on high-res data, they improved its performance on tasks with fewer labels. This shows starting with a broad training and then specializing on high-resolution images helps tackle detailed Earth observation tasks.

What this means in practice

  • For remote sensing teams: Improve detection and mapping tasks by using models further trained on high-resolution satellite and aerial images for better detail recognition.
  • For urban planning departments: Use specialized models to enhance identification of buildings, roads, and amenities from very high resolution imagery for improved city management.

Authors

Kshitij Raj Sharma, Pierre Adorni, Minh-Tan Pham, Dirk Tiede, Sébastien Lefèvre

Abstract

Geospatial foundation models are a widely adopted starting point for remote sensing tasks. Most are pretrained on Sentinel-1/2 and Landsat imagery at a ground sampling distance (GSD) of 10m or coarser. Many applications instead rely on very high resolution (VHR) imagery, where individual buildings, roads, and amenities are resolved. We study which pretraining regimes transfer well to this scale. We introduce VHR-Bench, a benchmark suite of ten datasets at 1.5m GSD or finer spanning five task families, eight application domains, and both satellite and aerial platforms. We evaluate twelve encoders in three pretraining tiers with frozen backbones at two training-data fractions. We find that both pretraining on large natural-image collections and VHR-specific pretraining transfer effectively to VHR tasks, with complementary strengths. SkySense leads when labels are scarce; at the full label fraction the web-pretrained DINOv3 ViT-L edges ahead, with the two close together there. Encoders pretrained only on Sentinel-scale imagery transfer least well under RGB input. Guided by these complementary results, we continue the self-supervised pretraining of DINOv3 ViT-L on 1.02 million held-out VHR tiles and evaluate the resulting encoder without changing the downstream protocol. The continued model gives a clear improvement over the released backbone when labels are scarce and a modest one at the full fraction, at no increase in inference parameters or compute. These findings support a generalize-then-specialize strategy: already strong, broad natural-image pretraining can be further improved by a short VHR continuation stage while preserving its general visual features.