pawlo2013/EarthLoc_2021_Database

🌍 Sentinel 2021 Image WebDataset

This dataset contains 150k 1024x1024 satellite images accross (9,10,11) zooms and bounded within (-60,60) latitude. Stored using the WebDataset format. The data is sharded across 11 .tar archives, and each sample contains:

• A JPEG image .jpg
• A unique key __key__ corresponding to the original image path
• Companion .index file for fast random access stored next to the shard

This key encodes all relevant metadata about the image, including its bounding box, nadir, zoom, area parameters.

Note this is not the whole data used to train the model. 3 more datasets are needed for years 2018,2019,2020 you can download them from https://github.com/gmberton/EarthLoc if you want to train from scratch.

---

🗂️ Dataset Structure

Each shard (e.g., shard-000000.tar) contains up to 15,000 samples. Every sample includes:

• jpg: The image bytes (decoded automatically)
• __key__: A string derived from the image's original path

The key format encodes metadata in the filename using the following structure:

@ lat1 @ lon1 @ lat2 @ lon2 @ lat3 @ lon3 @ lat4 @ lon4 @ image_id @ timestamp @ nadir_lat @ nadir_lon @ sq_km_area @ orientation @.jpg

Commas (,) in the keys have been used to replace dots (.) due to WebDataset's format. You’ll need to reverse this replacement (replace(',', '.')) when decoding.

For more details on how the key structure encodes geospatial metadata, refer to the EarthLoc repository.

---

🧪 Example: Displaying the First Image in a Shard

You can inspect a sample using the following code in a Jupyter notebook:

import webdataset as wds
from PIL import Image
import matplotlib.pyplot as plt

# Path to a WebDataset .tar file
tar_path = "./shard-000000.tar"

# Create a WebDataset iterator
dataset = wds.WebDataset(tar_path).decode("pil")

# Load the first sample
sample = next(iter(dataset))

# Access image and key
image = sample["jpg"]  # PIL Image
key = sample["__key__"].replace(',', '.')

# Display the image
plt.imshow(image)
plt.axis("off")
plt.title(f"Key: {key}")
plt.show()

# Print the key (for metadata parsing)
print(f"Key: {key}")