ARAS400k: A Large-Scale Remote Sensing Dataset Augmented with Synthetic Data for Segmentation and Captioning

Alternative title: Grounding Synthetic Data Generation With Vision and Language Models

ARAS400k contains 100,240 real and 300,000 synthetic remote sensing images (400,240 in total). Every image is 256×256 pixels and is paired with a semantic segmentation mask and five descriptive captions (2,001,200 captions in total).

This repository mirrors the official release on Zenodo: 10.5281/zenodo.18890661 (v1, published March 9, 2026). Code: https://github.com/caglarmert/ARAS400k

• Authors: Ümit Mert Çağlar (ORCID), Alptekin Temizel (ORCID) — Middle East Technical University
• License: CC BY 4.0
• Keywords: Semantic Segmentation, Image Captioning, Synthetic Data Generation, Remote Sensing, Deep Learning

Splits

Split (Zenodo folder)	HF split	Images
train	train	80,192
val	validation	10,024
test	test	10,024
synth	synth	300,000
Total		400,240

The train / validation / test splits are real images; synth holds the synthetic images.

Data format

On Zenodo each subset is a folder with images/, masks/ and captions.csv. Here, to stay within Hugging Face repository file-count limits and to enable the dataset viewer, the same content is stored as sharded Parquet files under data/. The PNG image and mask bytes are embedded unchanged (lossless, byte-identical to the Zenodo files), and each row is one image / mask / captions record taken from captions.csv.

Column	Description
image	256×256 PNG image
mask	256×256 PNG semantic segmentation mask
filename	Original file name (identical for the image, the mask and the captions.csv row), e.g. N36E025_000_002.png (real) or 0000.png (synthetic)
split	Original subset name (train, val, test, synth)
Tree, Shrub, Grass, Crop, Built-up, Barren, Water	Percentage of the image covered by each of the seven land cover classes
hybrid_gemma3-4b, hybrid_qwen3-vl-8b, text_qwen3-4b, vision_gemma3-4b, vision_qwen3-vl-8b	The five captions per image, one per caption-generation method

Caption methods: vision_* captions are produced from the image by a vision-language model, text_* from the class percentages by a language model, and hybrid_* from both.

Usage

from datasets import load_dataset

ds = load_dataset("caglarmert/ARAS400k", split="validation")
ex = ds[0]
ex["image"], ex["mask"], ex["hybrid_gemma3-4b"]

Use streaming=True to avoid downloading the full ~40 GB (the synth split alone is ~33 GB).

Citation

@article{ccauglar2026grounding,
  title={Grounding synthetic data generation with vision and language models},
  author={{\c{C}}a{\u{g}}lar, {\"U}mit Mert and Temizel, Alptekin},
  journal={arXiv preprint arXiv:2603.09625},
  year={2026}
}

Please also cite the Zenodo record (DOI 10.5281/zenodo.18890661) when using the data.