README.md

---
license: cc-by-nc-nd-4.0
task_categories:
- text-to-image
---

The **Git-10M** dataset is a global-scale remote sensing image-text pair dataset, consisting of over **10 million** image-text pairs with geographical locations and resolution information.

## CC-BY-NC-ND-4.0 License: This dataset is not allowed to be modified or distributed without authorization!

<h1>
<a href="https://chen-yang-liu.github.io/Text2Earth/">Project Page: https://chen-yang-liu.github.io/Text2Earth/ </a>
</h1>

<div align="center">
  <img src="https://github.com/Chen-Yang-Liu/Text2Earth/raw/main/images/dataset.png" width="1000"/>
</div>


## View samples from the dataset 
```python
from datasets import load_dataset
import math

def XYZToLonLat(x,y,z):
    # Transform tile-location to (longitude,latitude)
    n = 2**z*1.0
    lon = x / n * 360.0 - 180.0  # longitude
    lat = math.atan(math.sinh(math.pi * (1 - 2.0 * y / n)))
    lat = math.degrees(lat)  # latitude
    return lon,lat


# load dataset
save_path = 'xxxxx'
ds = load_dataset.load('lcybuaa/Git-10M', cache_dir=save_path)
train_dataset = ds["train"]


for i, example in enumerate(train_dataset):
    # PIL image:
    image = example["image"]
    # filename of the image:
    img_name = example["img_name"]
    # visual quality score as shown in Fig. 5 of the paper.
    img_quality_score = example['img_quality_score']
    # caption of the image
    caption = example['caption']
    # word length of the caption as shown in Fig. 6 of the paper.
    caption_length = example['caption_length']
    # image spatial resolution as shown in Fig. 4 of the paper.
    resolution = example['resolution']
    # image Geolocation as shown in Fig. 3 of the paper.
    Google_location = example['Google_location']
    Level_TileZ, TileX, TileY = Google_location.split('_')
    longitude, latitude = XYZToLonLat(TileX, TileY, Level_TileZ)

    # More Tips:
    # Resolution = 2 ** (17 - Level_TileZ)

```

## Git-RSCLIP: Remote Sensing Vision-Language Contrastive Pre-training Foundation Model
Git-RSCLIP is pre-trained using the contrastive learning framework on the **Git-10M dataset**.
Git-RSCLIP is here：[[Huggingface](https://huggingface.co/lcybuaa/Git-RSCLIP) | [Modelscope](https://modelscope.cn/models/lcybuaa1111/Git-RSCLIP)]

Compare the Top1-Acc of Zero-shot classification on multiple image classification datasets:

| Method | OPTIMAL31 | RSC11 | RSICB128 | WHURS19 | RS2800/RSSCN7 | CLRS | Average score |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| CLIP | 0.6 | 0.45 | 0.25 | 0.77 | 0.52 | 0.56 | 0.52 |
| RemoteCLIP | 0.82 | 0.67 | 0.34 | 0.93 | 0.52 | 0.66 | 0.65 |
| GeoRSCLIP | 0.83 | 0.67 | 0.35 | 0.89 | 0.63 | 0.69 | 0.68 |
| SkyCLIP50 | 0.77 | 0.60 | 0.38 | 0.78 | 0.55 | 0.61 | 0.62 |
| (Git-RSCLIP) Ours | **0.95** | **0.67** | **0.52** | **0.94** | **0.64** | **0.65** | **0.73** |



# BibTeX entry and citation info

```bibtex
@ARTICLE{Text2Earth,
  author={Liu, Chenyang and Chen, Keyan and Zhao, Rui and Zou, Zhengxia and Shi, Zhenwei},
  journal={IEEE Geoscience and Remote Sensing Magazine}, 
  title={Text2Earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model}, 
  year={2025},
  volume={},
  number={},
  pages={2-23},
  doi={10.1109/MGRS.2025.3560455}}
```