TerraMesh
=========

:doc:`← Back to geoai <../geoai/index>`

.. list-table:: Dataset Metadata
   :header-rows: 1

   * - Field
     - Value
   * - **Folder**
     - ``/datasets/geoai/ibm-esa-geospatial/TerraMesh``
   * - **Discipline**
     - GeoAI / Remote Sensing / Earth Science
   * - **DOI**
     - `10.48550/arXiv.2504.11172 <https://doi.org/10.48550/arXiv.2504.11172>`__
   * - **Link**
     - `Access Data <https://huggingface.co/datasets/ibm-esa-geospatial/TerraMesh>`__
   * - **Public**
     - ``True``
   * - **Publication Date**
     - 2025-09-05
   * - **Downloaded**
     - 2025-09-05
   * - **Data Type**
     - tar
   * - **Dataset Size**
     - 31T
   * - **Number of Files**
     - 12618
   * - **Usage Policy Link**
     - https://choosealicense.com/licenses/cc-by-sa-4.0/

Description
-----------
TerraMesh<br><br>Dataset Summary<br>TerraMesh is a planetary-scale, multimodal analysis-ready dataset for Earth Observation foundation models. It merges Sentinel-1 SAR, Sentinel-2 optical, Copernicus DEM, NDVI, and land-cover sources into more than nine million co-registered patches for large-scale representation learning.<br><br>Dataset Structure<br>The dataset includes two top-level splits (train/ and val/), each containing sub-folders for modalities: DEM, LULC, NDVI, S1GRD, S1RTC, S2L1C, S2L2A, and S2RGB. Each folder contains up to 889 shard files, each storing up to 10,240 samples as compressed Zarr archives.<br><br>Data Characteristics<br>Each sample contains seven spatially aligned modalities (optical, radar, topographic, vegetation, and land-cover). Metadata fields include center latitude/longitude, timestamps, CRS, and cloud masks.<br><br>Intended Use<br>TerraMesh enables multimodal pretraining, global geospatial feature extraction, and benchmarking of foundation models for planetary surface understanding.<br><br>Performance & Evaluation<br>Pretraining on TerraMesh led to TerraMind-B achieving 66.6% mIoU across PANGAEA benchmark tasks, outperforming CROMA and SSL4EO-S12 models.<br><br>Acknowledgments<br>Developed under ESA Φ-Lab’s FAST-EO project. Source data include SSL4EO-S12 (CC-BY-4.0), MajorTOM-Core (CC-BY-SA-4.0), and Copernicus DEM (© DLR / Airbus / ESA).

Usage
-----
.. code-block:: bash

   $ module avail
   $ module load datasets
   $ module load geoai/ibm-esa-geospatial/TerraMesh/2025-09-05

Usage Policy
------------

See also: https://choosealicense.com/licenses/cc-by-sa-4.0/

Citation
--------
Blumenstiel, B., Fraccaro, P., Marsocci, V., Jakubik, J., Maurogiovanni, S., Czerkawski, M., Sedona, R., Cavallaro, G., Brunschwiler, T., Bernabe-Moreno, J., et al. (2025). TerraMesh: A planetary mosaic of multimodal Earth observation data. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops.

BibTeX
------
.. code-block:: bibtex

   @article{blumenstiel2025terramesh,
     title={Terramesh: A planetary mosaic of multimodal earth observation data},
     author={Blumenstiel, Benedikt and Fraccaro, Paolo and Marsocci, Valerio and Jakubik, Johannes and Maurogiovanni, Stefano and Czerkawski, Mikolaj and Sedona, Rocco and Cavallaro, Gabriele and Brunschwiler, Thomas and Bernabe-Moreno, Juan and others},
     journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
     year={2025},
   }