MBZUAI/OpenEarthAgent
收藏Hugging Face2026-03-03 更新2026-04-05 收录
下载链接：
https://hf-mirror.com/datasets/MBZUAI/OpenEarthAgent
下载链接
链接失效反馈官方服务：
资源简介：
---
license: apache-2.0
---
**A Tool-Augmented Geospatial Reasoning Dataset for Multimodal Earth Observation Agents**
---
## Dataset Summary
**OpenEarthAgent Dataset** is a large-scale, tool-augmented geospatial reasoning corpus designed to train and evaluate multimodal agents on structured, multi-step Earth Observation (EO) tasks.
Unlike conventional remote sensing datasets focused on perception (classification, detection, segmentation), this dataset enables **interpretable multi-step reasoning** through explicit tool interactions over:
- Optical satellite imagery
- SAR imagery
- GIS vector layers
- Georeferenced rasters (GeoTIFF)
- Spectral index layers (NDVI, NBR, NDBI, etc.)
Each instance includes:
- A natural-language query
- Multimodal geospatial inputs
- A structured reasoning trajectory
- Explicit tool calls with arguments
- Intermediate tool observations
- Final grounded answer
The dataset supports research in:
- Tool-augmented LLMs
- Geospatial reasoning
- Multimodal agents
- Interpretable EO workflows
- Structured planning with spatial grounding
---
## Dataset Statistics
| Split | Instances | Reasoning Steps | Avg. Steps per Query |
|--------|------------|----------------|-----------------------|
| Train | 14,538 | 100,656 | 6.92 |
| Test | 1,169 | 7,064 | 6.04 |
Total reasoning steps across the corpus exceed **107K structured thought-action-observation transitions**.
---
提供机构：
MBZUAI
