Dataset Card for SATIN

Dataset Description

• Homepage: https://satinbenchmark.github.io
• Repository:
• Paper: SATIN: A Multi-Task Metadataset for Classifying Satellite Imagery using Vision-Language Models
• Leaderboard: SATIN Leaderboard

Dataset Summary

SATIN (SATellite ImageNet) is a metadataset containing 27 constituent satellite and aerial image datasets spanning 6 distinct tasks: Land Cover, Land Use, Hierarchical Land Use, Complex Scenes, Rare Scenes, and False Colour Scenes. The imagery is globally distributed, comprised of resolutions spanning 5 orders of magnitude, multiple fields of view sizes, and over 250 distinct class labels. Presented at ICCV '23 TNGCV Workshop.

Dataset Structure

The SATIN benchmark is comprised of the following datasets:

Task 1: Land Cover

• SAT-4
• SAT-6
• NASC-TG2

Task 2: Land Use

• WHU-RS19
• RSSCN7
• RS_C11
• SIRI-WHU
• EuroSAT
• NWPU-RESISC45
• PatternNet
• RSD46-WHU
• GID
• CLRS
• Optimal-31

Task 3: Hierarchical Land Use

• Million-AID
• RSI-CB256

Task 4: Complex Scenes

• UC_Merced_LandUse_MultiLabel
• MLRSNet
• MultiScene
• AID_MultiLabel

Task 5: Rare Scenes

• Airbus-Wind-Turbines-Patches
• USTC_SmokeRS
• Canadian_Cropland
• Ships-In-Satellite-Imagery
• Satellite-Images-of-Hurricane-Damage

Task 6: False Colour Scenes

• Brazilian_Coffee_Scenes
• Brazilian_Cerrado-Savanna_Scenes

For ease of use and to avoid having to download the entire benchmark for each use, in this dataset repository, each of the 27 datasets is included as a separate 'config'.

Example Usage

from datasets import load_dataset


hf_dataset = load_dataset('jonathan-roberts1/SATIN', DATASET_NAME, split='train') # for DATASET_NAME use one of the configs listed above (e.g., EuroSAT)
features = hf_dataset.features
class_labels = features['label'].names 
#class_labels = features['label'].feature.names # for the Complex Scenes datasets
#class_labels_1 = features['label_1'].names # for the Hierarchical Land Use datasets, the label field is replaced with label_1, label_2, ...

random_index = 5
example = hf_dataset[random_index]
image, label = example['image'], example['label']

Data Splits

For each config, there is just the single, default 'train' split.

Source Data

More information regarding the source data can be found in our paper. Additionally, each of the constituent datasets have been uploaded to HuggingFace datasets. They can be accessed at: huggingface.co/datasets/jonathan-roberts1/DATASET_NAME.

Dataset Curators

This dataset was curated by Jonathan Roberts, Kai Han, and Samuel Albanie

Licensing Information

As SATIN is comprised of existing datasets with differing licenses, there is not a single license for SATIN. All of the datasets in SATIN can be used for research purposes; usage information of specific constituent datasets can be found in the Appendix of our paper.

Citation Information

@article{roberts2023satin,
  title        = {SATIN: A Multi-Task Metadataset for Classifying Satellite Imagery using Vision-Language Models},
  author       = {Jonathan Roberts, Kai Han, and Samuel Albanie},
  year         = {2023},
  eprint       = {2304.11619},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV}
}