Workshop 2025

Session 1: Foundation Models in Practice: Transfer and Domain Generalization Challenges in EO

Foundation models are reshaping Earth Observation (EO) by enabling cross-task transfer, domain adaptation, and zero-shot inference. Yet, applying these models in EO remains nontrivial: domain shifts, spatiotemporal variability, and data sparsity challenge both generalization and deployment. This session invites contributions that explore the development, evaluation, and real-world use of foundation models in EO settings, such as spanning pre-training strategies, fine-tuning paradigms, and evaluation mechanisms. We welcome both theoretical insights and application-driven case studies.

1) How can foundation models trained on global EO data be adapted to highly localized or domain-specific tasks (e.g., urban flood detection, crop type mapping)?

2) What are effective strategies for ensuring generalization across geographical, seasonal, and sensor shifts in EO applications?

3) Can foundation models in EO reduce the reliance on labeled data—and if so, under what limitations or risks?

Session 2 : Change Detection and Time Series Analysis

Earth Observation time series data provide vital insights into dynamic environmental processes, yet analyzing these data requires specialized methods to handle irregular sampling, data gaps, and complex temporal patterns. This session focuses on the latest deep learning approaches with applications for EO time series data. We welcome contributions addressing anomaly detection, change classification, forecasting, and rigorous validation of temporal models. Papers discussing how to handle clouds, missing data, and irregular observation intervals in EO time series are also encouraged.

1）Which machine learning architectures are promising to capture temporal dependencies in EO time series?

2）How can irregular time series and data gaps be effectively handled in time series tasks?

3）How should temporal models be validated to ensure reliability in real-world EO monitoring?

Session 3: Multisensor Data Fusion in Earth Observation: Methods, Challenges, and Applications

Integrating data from diverse Earth Observation sensors—such as SAR, optical, LiDAR, hyperspectral, and BIM—enables more comprehensive and accurate environmental monitoring and analysis. This session focuses on the latest advances in multisensor data fusion, covering techniques at both feature and decision levels, sensor-specific preprocessing strategies, and innovative architectures for combining heterogeneous data sources. We invite contributions that address practical challenges, methodological innovations, and application-driven case studies that demonstrate how fusion enhances EO insights. Discussions on limitations, sensor-specific considerations, and benchmarking are also welcome.

1）What fusion techniques are promising across different sensor combinations and EO applications?

2）How can preprocessing be optimized to harmonize data with diverse characteristics and resolutions?

3）What are the main challenges in aligning and integrating satellite-based and in-situ data sources?

Session 4:  Reasoning, Uncertainty Quantification, and Explainable AI for Trustworthy Earth Observation

As AI becomes increasingly integral to Earth Observation and environmental sciences, ensuring that models provide not only accurate predictions but also understandable, trustworthy, and actionable insights is crucial. This session explores methods for uncertainty quantification, explainable AI (XAI), and hybrid modeling approaches that combine data-driven and physical knowledge. Contributions addressing how to build AI systems with traceability and trust, and how these advances foster scientific reasoning and decision-making, are encouraged.

1）How can or should we quantify and communicate uncertainties in AI predictions for EO applications?

2）What explainability techniques provide real insights beyond black-box outputs?

3）How can hybrid models that incorporate physical knowledge enhance AI interpretability and robustness?

4）What design principles ensure trust, traceability, and accountability in AI systems?

Session 5: Data, Platforms, and Infrastructure for Scalable Earth Observation

The success of AI-driven Earth Observation (AI4EO) hinges not only on algorithmic advances but critically on the availability, accessibility, and interoperability of diverse datasets and the supporting platforms and infrastructures. This is underpinned by community driven open-source efforts that make software tools available to researchers and practitioners. We encourage papers exploring technical, organizational, and normative aspects to accelerate efficient AI4EO workflows. Discussions on benchmarks, data standards, community-driven validation, and sustainability of platforms are also highly welcomed.

1) What are the current bottlenecks in accessing and harmonizing EO datasets across national, commercial, and open sources?

2) How can federated platforms and data spaces be designed to support scalable, interoperable AI4EO applications?

3) What infrastructure (HPC, cloud, repositories) is essential to enable reproducible and scalable EO AI workflows?

4) Which open datasets, benchmarks, and metadata standards are most needed to advance AI4EO research?

5) How can the EO community collaboratively validate AI models and ensure sustainable platform development?