Visual grounding benchmark for geospatial AI
How well AI models do visual grounding on satellite, aerial and map data: 1 set with results, scored the same way for every model.
Sets
| Set | Index | Tasks of this type | Best score on the set | Models |
|---|---|---|---|---|
| BigEarthNet.txt | Spatial Intelligence Index | 60 | 60 | 16 |
Top models
Best score on BigEarthNet.txt, the set most models have run, 0 to 100.
| Rank | Model | Developer | Score |
|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | 60 |
| 2 | Claude Sonnet 5.5 | Anthropic | 57 |
| 3 | Claude Haiku 5.5 | Anthropic | 54 |
| 4 | Claude Opus 5 | Anthropic | 53 |
| 5 | Claude Opus 4.7 | Anthropic | 49 |
| 6 | Step 5 Preview | StepFun | 49 |
| 7 | Claude Opus 4.8 | Anthropic | 48 |
| 8 | Qwen 3.8 27B | Alibaba (Qwen) | 48 |
| 9 | Claude Sonnet 5 | Anthropic | 47 |
| 10 | Nemotron 3.5 Super VL 120B | NVIDIA | 47 |
Every task type · Every evaluation · Methodology
Frequently asked questions
Which AI model is best at visual grounding on geospatial data?
Claude Opus 5.5 leads, on the sets below. Best score on BigEarthNet.txt, the set most models have run, 0 to 100.
What is this visual grounding benchmark?
Every task in our sets that asks for visual grounding, in GIS and Earth observation work, with each set's own score. It is a view over the sets, not an index of its own.
Which index does it count in?
Each set counts in one index: the Spatial Intelligence Index, as the table shows.
