Papers

Cesar Aybar, Julio Contreras, Chen Ma, Oscar J. Pellicer-Valero, Gonzalo Mateo-García, Luis Gómez-Chova, Gustau Camps-Valls, Nils Lehmann, Mikolaj Czerkawski, David Montero, Miguel D. Mahecha, Ingrid Aybar

TerraBytes, ICML 2025 Workshop

Abstract

Geospatial communities have long relied on standardised formats for raster and vector data, enabling interoperability, stable tool development, and long-term data preservation. In contrast, Artificial Intelligence (AI)-ready datasets, particularly those derived from Earth Observation (EO), lack equivalent conventions. As a result, data producers often adopt ad hoc file structures, loosely defined formats, and inconsistent semantic encodings. This fragmentation hinders interoperability, complicates reuse, and undermines reproducibility. We argue that the lack of a standard format represents a structural bottleneck to scalable scientific progress, especially in the era of foundation models, where diverse datasets must be combined for effective training and performance evaluation in downstream tasks. To address this, we introduce TACO: a comprehensive specification that defines a formal data model, a cloud-optimized on-disk layout, and an API for creating and accessing AI-ready EO datasets.

Oscar J. Pellicer-Valero, Cesar Aybar, Mikolaj Czerkawski, Carmen Oliver, Kevin Monsálvez, Julio Contreras, Gustau Camps-Valls

EGU General Assembly 2026 · Vienna, Austria, 3-8 May 2026

Abstract

The field of Artificial Intelligence for Earth Observation (AI4EO) currently suffers from significant data friction, especially when moving Petabyte-scale archives from cloud object storage to High Performance Computing (HPC) nodes. We present TACO (Transparent Access to Cloud-Optimized datasets), a production-grade standard designed to replace file-centric legacy workflows with a high-throughput streaming paradigm. We showcase the practical implications of this architecture for the deployment of geospatial Foundation Models (FMs), by running pretrained FMs on downstream inference tasks (such as semantic segmentation or land-cover classification) directly on arbitrary samples of arbitrary cloud-hosted datasets, quickly, and without the need for local staging or any specific preprocessing.

David Montero, César Aybar, Miguel D. Mahecha, Luis Gómez-Chova

ISPRS Archives, L-4/W1-2026, 167-174

Abstract

Earth observation (EO) is essential to understanding the Earth system, enabling the transformation of planetary properties into measurable variables that can be analysed, compared, and modelled. In recent decades, EO capabilities have grown rapidly, accompanied by an even faster expansion in the number and variety of available EO instruments. However, despite this proliferation, users often lack a single, reliable source describing their existence and key characteristics. Here we present Awesome Earth Observation Instruments, an open, standardized, and community-oriented registry providing machine-readable metadata for EO instruments, hosted on GitHub and validated against a common schema.

Julio Contreras¹*, Carlos Giner¹*, Cesar Aybar¹, Gonzalo Mateo-García², Javier Gorroño³, David Montero⁴,⁵, Miguel D. Mahecha⁴,⁵, Luis Guanter³, Luis Gómez-Chova¹

Scientific Data · in preparation

• ¹ Image and Signal Processing Group, Universitat de València, Spain
• ² International Methane Emissions Observatory, United Nations Environment Programme, France
• ³ Universitat Politècnica de València, Spain
• ⁴ Remote Sensing Centre for Earth System Research (RSC4Earth), Leipzig University, Germany
• ⁵ German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig, Germany
• * Equal contribution.

Abstract

Detecting methane plumes from space at high spatial resolution relies on two broad classes of instruments that sample the shortwave infrared window: multispectral radiometers such as Sentinel-2 and Landsat 8/9, and imaging spectrometers such as EMIT, each with distinct data requirements. Multispectral radiometers possess only two broad shortwave infrared bands, so robust detection requires temporally paired acquisitions to isolate methane absorption. Imaging spectrometers sample the same window at much finer spectral resolution, giving high sensitivity to the methane signature and enabling detection from a single acquisition; their processing approaches differ, from matched-filter retrievals to deep learning, but all require calibrated radiance as input.

No existing resource standardizes data across both instrument classes or provides the input diversity needed to train and benchmark detection models jointly. MethaneSET addresses this gap with datasets for Sentinel-2, Landsat 8/9, and EMIT under a unified metadata schema. The multispectral components include expert-verified plume masks and confirmed plume-free reference scenes. The hyperspectral EMIT component is, to our knowledge, the first to distribute calibrated radiance hypercubes alongside plume annotations, adding precomputed matched-filter products and two independently derived plume segmentation masks per scene. A synthetic plume bank of over one million enhancements from WRF-LES simulations enables both controlled data augmentation and physical sensitivity experiments across wind speeds, solar geometries, and source configurations. All datasets are distributed as Parquet catalogs with Cloud-Optimized GeoTIFFs.

Citation

@article{contreras2026methaneset,
  title   = {MethaneSET: Unified Multi-Sensor AI-Ready Datasets
             for Satellite-Based Methane Plume Detection},
  author  = {Contreras, Julio and Giner, Carlos and Aybar, Cesar
             and Mateo-Garcia, Gonzalo and Gorrono, Javier
             and Montero, David and Mahecha, Miguel D.
             and Guanter, Luis and Gomez-Chova, Luis},
  journal = {Scientific Data},
  year    = {2026}
}