Date : 2025-07-18 => 2025-07-19
Lieu : Vancouver, Canada
ICML 2025 workshop on global datasets and models for Earth Observation
When: 18/19 July 2025
Where: Vancouver, Canada
Submission website: https://openreview.net/group?id=ICML.cc/2025/Workshop/TerraBytes
Important dates
Papers submission deadline: May 16th, 2025
Papers acceptance notification: June 6, 2025
Workshop: July 18/19, 2025
Earth Observation (EO) presents unique challenges and opportunities that set it apart from other fields of machine learning (ML) and computer vision (CV) (Rolf et al., 2024). EO data is abundant, repeatedly covering a large but limited environment: our planet. The colocation and evolution of these observations is a rich, emergent source of information, of multimodal and multitemporal nature. However, due to both local and global changes, especially climate change, the statistical distribution of EO data is inherently non-stationary (Tarasiou et al., 2023). These properties break some of the usual assumptions of ML, and EO data require special care for data handling and modeling (Mai et al., 2022). In addition, current EO datasets are sampled with spatio-temporal biases. Some areas, e.g., the global South, are strongly under-represented within EO datasets (Cornebise et al., 2022). In optical imagery, cloud cover is undesirable, thus leading to datasets that remove cloudy images at the risk of biasing their geographical coverage (Tiede et al., 2021). Addressing these distributional biases is of primary importance, as they have an impact on the performance and reliability of models for downstream applications in ecology, geosciences, agriculture, urban planning, etc. (Kattenborn et al., 2022).
TerraBytes is an initiative to address these challenges. At the intersection of data curation, data archiving, and representation learning, this workshop will foster a holistic discussion covering major steps in the EO from downlinked satellite data, training paradigms to downstream applications.
Call for submissions
We invite submissions on the following topics:
Large-scale Earth observation datasets and benchmarks:
curation of new large-scale vision datasets for EO,
augmentation techniques for EO data and their impact on learning, data traceability and expansion of existing datasets,
re-use and adoption of existing datasets in new scenarios,
impact of training data quality and size on downstream tasks under domain shifts.
Efficient and continual representation learning for remote sensing:
weakly and self-supervised learning for EO data,
domain adaptation strategies to deal with temporal, geographical or sensor gaps,
continual and online learning, to update models when facing distribution shifts,
active learning and reinforcement learning.
Real-world applications:
early detection, monitoring and assessment of disasters e.g. floods, landslides, wildfires,
large-scale Land Cover and Land Use mapping,
2D and 3D change detection in urban and rural areas,
estimation of biophysical parameters (biomass, biodiversity, soil stress, etc.).
Notre site web : www.madics.fr
Suivez-nous sur Tweeter : @GDR_MADICS
Pour vous désabonner de la liste, suivre ce lien.