metadata

license: cc-by-nc-4.0
task_categories:
  - text-to-image
language:
  - en
tags:
  - climate
  - biology
  - remote-sensing
  - geospatial
  - multimodal
  - crisis-management
  - climate
pretty_name: CrisisLandMark
size_categories:
  - 100K<n<1M

CrisisLandMark

CrisisLandMark is a large-scale, multimodal corpus for Text-to-Remote-Sensing-Image Retrieval (T2RSIR). It contains over 647,000 Sentinel-1 (SAR) and Sentinel-2 (multispectral optical) images enriched with structured textual and geospatial annotations. The dataset is designed to move beyond standard RGB imagery, enabling the development of retrieval systems that can leverage the rich physical information from different satellite sensors for applications in Land Use/Land Cover (LULC) mapping and crisis management.

Curated by: Daniele Rege Cambrin
License: Creative Commons Attribution Non Commercial 4.0
Repository: GitHub
Paper: Arxiv

Getting Started

To work with the dataset, you will need the h5py and hdf5plugin packages. You can try to use the HF dataset implementation, but I strongly suggest using the GitHub implementation.

Dataset Structure

The dataset is distributed in HDF5 format. The main file, crisislandmark.h5, contains external links to data shards, so all .h5 files must be kept in the same directory for access. Each key in the HDF5 file corresponds to a unique sample. Each key contains the following matrices:

Key	Shape	Data Type	Description
`image`	`(B, 120, 120)`	`float32`	The image data. B is the number of bands: 2 for Sentinel-1 (VV, VH) or 12 for Sentinel-2.
`coords`	`(2, 120, 120)`	`float32`	Contains the x and y coordinates for each pixel of the image.

and the following attributes:

Key	Shape	Data Type	Description
`crs`	`(1)`	`float32`	The EPSG code for the Coordinate Reference System.
`timestamp`	`(1)`	`float32`	Contains the associated timestamp if available.
`labels`	`(L)`	`list[str]`	Contains the list of labels associated with the image.

The metadata.parquet files contain, for each key, the associated split and the labels from the original source (either CLC or DW). The queries.jsonl file contains the queries and their IDs. The qrels for each query can be found in the qrels folder.

Data split

The dataset is divided into two splits based on a stratified multi-label sampling strategy to ensure similar label distribution:

Training Set: 20% of the data, intended for model training.
Corpus Set: 80% of the data, intended for retrieval and evaluation.

Source Data and Annotations

The dataset is built from satellite images sourced from Sentinel-1 and Sentinel-2 missions. The raw images were drawn from five existing public datasets: re-BEN, CaBuAr, QuakeSet, MMFlood, and Sen12Flood. All images were processed to a uniform 10-meter spatial resolution and divided into 120x120 pixel patches.

The annotations are created from the following sources:

Land Use/Land Cover (LULC): Annotations were derived from the CORINE Land Cover (CLC) system for European regions and the global, near-real-time Dynamic World (DW) system for crisis-event images. The script mapping.py in the repository details the mapping between CLC and DW.
Crisis Events: For images from crisis-focused datasets, original event tags like "wildfire", "flooding", and "earthquake" were retained.
Geospatial: Every image patch is annotated with its geographic coordinates.

Citation

@misc{cambrin2025texttoremotesensingimageretrievalrgbsources,
      title={Text-to-Remote-Sensing-Image Retrieval beyond RGB Sources}, 
      author={Daniele Rege Cambrin and Lorenzo Vaiani and Giuseppe Gallipoli and Luca Cagliero and Paolo Garza},
      year={2025},
      eprint={2507.10403},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.10403}, 
}

Licensing

The data in this dataset is a compilation of multiple sources, each with its own license. For detailed information on the licensing of each component, please see the NOTICE.md file.