hackelle/BigEarthNetV2-Lithuania-Summer-LMDB

TU Berlin	RSiM	DIMA	BigEarth	BIFOLD

---

reBEN — Lithuania Summer Subset (pre-converted to LMDB)

⚠️ Unofficial mirror. This is an unofficial, community-provided pre-conversion of a subset of the BigEarthNet v2.0 (reBEN) dataset into LMDB format. It is provided as a convenience for researchers who wish to get started quickly without running the full conversion pipeline. In case of any discrepancy, the original publication and the original files always take precedence. Please refer to the authoritative sources listed below.

---

Overview

This dataset card describes a pre-converted LMDB version of a subset of BigEarthNet v2.0 (also known as reBEN — Refined BigEarthNet), a large-scale, multi-label remote sensing benchmark dataset. The subset contains patches from Lithuania acquired during the summer months (June, July, August), identified by filtering on the acquisition month index embedded in the patch ID.

The original dataset was converted to LMDB format using rico-HDL, which is the recommended conversion tool for reBEN. The LMDB file stores Sentinel-1 and Sentinel-2 patches as serialized SafeTensors entries, keyed by patch ID.

The accompanying metadata_lithuania_summer.parquet file provides all patch-level metadata (labels, split assignments, geographic information, etc.) for the included patches.

---

Authoritative Sources

Please always consult the following primary resources:

Resource	Link
BigEarthNet project page	bigearth.net
BigEarthNet image–text dataset (txt.bigearth.net)	txt.bigearth.net
Original reBEN files (Zenodo)	zenodo.org/records/10891137
reBEN training scripts (official repository)	git.tu-berlin.de/rsim/reben-training-scripts
Pretrained model weights	BIFOLD-BigEarthNetv2-0 on Hugging Face

---

Dataset Details

Subset Criteria

Patches were selected from the full reBEN dataset using two criteria applied to the metadata:

1. Country: Lithuania (i.e., metadata.country == 'Lithuania')
2. Season: Summer, determined by the month index in the patch ID — specifically, patches where the character at position 5 of the third underscore-separated token in the patch ID is one of {'6', '7', '8'} (corresponding to June, July, and August).

In code:

metadata_lithuania_summer = metadata[
    (metadata.country == 'Lithuania') &
    (metadata.patch_id.apply(lambda x: x.split('_')[2][5] in {'6', '7', '8'}))
]

LMDB Structure

Each entry in the LMDB file is a SafeTensors-serialized object, keyed by either the Sentinel-2 patch ID or the corresponding Sentinel-1 patch name (s1_name). This matches the format produced by rico-HDL.

• Sentinel-2 entries contain bands: B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12
• Sentinel-1 entries contain bands: VV, VH

Metadata

The metadata_lithuania_summer.parquet file is a direct subset of the full reBEN metadata parquet, filtered to the patches described above. It retains all original columns including patch IDs, labels, split assignments, and geographic metadata.

---

Usage

The recommended way to use the dataset is with configILM, which directly supports pytorch dataloader and lightning datamodules with optimized multi-threaded loading. Please follow the documentation for details. Alternatively, to load individual patches from the LMDB file, you need lmdb and safetensors:

pip install lmdb safetensors pandas pyarrow

import lmdb
import pandas as pd
from safetensors.numpy import load as safetensor_load

lmdb_path = "path/to/BENv2_lithuania_summer.lmdb"
metadata_path = "path/to/metadata_lithuania_summer.parquet"

metadata = pd.read_parquet(metadata_path)
lmdb_env = lmdb.open(lmdb_path, map_size=1024, max_dbs=False, readonly=True)

# Load a Sentinel-2 patch
patch_id = metadata.patch_id.iloc[0]
with lmdb_env.begin(write=False) as txn:
    data = txn.get(patch_id.encode())
    tensor = safetensor_load(data)

# Access individual bands
r, g, b = tensor["B04"], tensor["B03"], tensor["B02"]

# Load the corresponding Sentinel-1 patch
s1_name = metadata.s1_name.iloc[0]
with lmdb_env.begin(write=False) as txn:
    data = txn.get(s1_name.encode())
    tensor = safetensor_load(data)

vv, vh = tensor["VV"], tensor["VH"]

---

Conversion Details

The LMDB was generated from the original reBEN dataset files (downloaded from Zenodo) using rico-HDL, which is the officially recommended conversion tool for reBEN. The metadata subset was extracted using pandas and saved as a Parquet file.

---

License

The underlying data is licensed under the Community Data License Agreement — Permissive, Version 1.0 (CDLA-Permissive-1.0), consistent with the license of the original BigEarthNet v2.0 dataset. This pre-converted version inherits the same license.

---

Citation

If you use this dataset in your research, please cite the original reBEN publication and the ConfigILM library:

@inproceedings{clasen2025refinedbigearthnet,
  title={{reBEN}: Refined BigEarthNet Dataset for Remote Sensing Image Analysis},
  author={Clasen, Kai Norman and Hackel, Leonard and Burgert, Tom and Sumbul, Gencer and Demir, Beg{\"u}m and Markl, Volker},
  year={2025},
  booktitle={IEEE International Geoscience and Remote Sensing Symposium (IGARSS)},
}

@article{hackel2024configilm,
  title={ConfigILM: A general purpose configurable library for combining image and language models for visual question answering},
  author={Hackel, Leonard and Clasen, Kai Norman and Demir, Beg{\"u}m},
  journal={SoftwareX},
  volume={26},
  pages={101731},
  year={2024},
  publisher={Elsevier}
}

The preprint for reBEN is also available on arXiv:

K. Clasen et al., "reBEN: Refined BigEarthNet Dataset for Remote Sensing Image Analysis", arXiv:2407.03653, 2024. https://arxiv.org/abs/2407.03653