EARTHVISION 2025
June 11th, Nashville, TN, USA
in conjuction with the Computer Vision and Pattern Recognition (CVPR) 2025 Conference
Aims and Scope / Call for Papers
Earth Observation (EO) and remote sensing are ever-growing fields of investigation where computer vision, machine learning, and signal/image processing meet. The general objective of the domain is to provide large-scale and consistent information about processes occurring at the surface of the Earth by exploiting data collected by airborne and spaceborne sensors. Earth Observation covers a broad range of tasks, from detection to registration, data mining, and multi-sensor, multi-resolution, multi-temporal, and multi-modality fusion and regression, to name just a few. It is motivated by numerous applications such as location-based services, online mapping services, large-scale surveillance, 3D urban modeling, navigation systems, natural hazard forecast and response, climate change monitoring, virtual habitat modeling, food security, etc. The sheer amount of data calls for highly automated scene interpretation workflows.
Earth Observation and in particular the analysis of spaceborne data directly connects to 34 indicators out of 40 (29 targets and 11 goals) of the Sustainable Development Goals defined by the United Nations ( https://sdgs.un.org/goals ). The aim of EarthVision to advance the state of the art in machine learning-based analysis of remote sensing data is thus of high relevance. It also connects to other immediate societal challenges such as monitoring of forest fires and other natural hazards, urban growth, deforestation, and climate change.
A non exhaustive list of topics of interest includes the following:
Super-resolution in the spectral and spatial domain
Hyperspectral and multispectral image processing
Reconstruction and segmentation of optical and LiDAR 3D point clouds
Feature extraction and learning from spatio-temporal data
Analysis of UAV / aerial and satellite images and videos
Deep learning tailored for large-scale Earth Observation
Domain adaptation, concept drift, and the detection of out-of-distribution data
Data-centric machine learning
Evaluating models using unlabeled data
Self-, weakly, and unsupervised approaches for learning with spatial data
Foundation models and representation learning in the context of EO
Human-in-the-loop and active learning
Multi-resolution, multi-temporal, multi-sensor, multi-modal processing
Fusion of machine learning and physical models
Explainable and interpretable machine learning in Earth Observation applications
Uncertainty quantification of machine-learning based prediction from EO data
Applications for climate change, sustainable development goals, and geoscience
Public benchmark datasets: training data standards, testing & evaluation metrics, as well as open source research and development.
All manuscripts will be subject to a double-blind review process. Accepted EarthVision papers will be included in the CVPR2024 workshop proceedings (published open access on the Computer Vision Foundation website) and submitted to IEEE for publication in IEEEXplore. Publication in IEEEXplore will be granted only if the paper meets IEEE publication policies and procedures.
Important Dates
All deadlines are considered end of day anywhere on Earth.
| March 3, 2025 | Submission deadline | |
| March 31, 2025 | Notification to authors | |
| April 7, 2025 | Camera-ready deadline | |
| June 11, 2025 | Workshop |
Program
9:00-9:15 | Welcome and Awards Announcement |
9:15-10:00 | Keynote 1 – “From the Research Lab to a Global Map: Balancing Innovation with Operational Reality” Over the past decade, the field of Earth Observation has witnessed a remarkable transformation in how large-scale applications like global crop type mapping are approached. From carefully crafted, expert-driven systems to pipelines powered by geospatial foundation models, the pace of innovation has reshaped both expectations and possibilities. WorldCereal exemplifies this shift—evolving from a classification system built around stable, interpretable feature engineering to embracing pretrained models and their self-supervised learning capabilities. Yet, this evolution brings a persistent tension: staying at the forefront of scientific and technological progress while maintaining the operational robustness required to deliver global, real-world results. This talk reflects on that journey from a practitioner’s perspective, and the considerations that guide adoption in an ever-accelerating landscape. Kristof Van Tricht is a senior researcher at VITO Remote Sensing, where he specializes in agricultural applications. With a focus on both global and local agriculture monitoring, he leverages advanced remote sensing technologies to address food security challenges. His work spans the development and deployment of crop mapping algorithms, and the development of the CropSAR technology, which fuses Sentinel-1 and Sentinel-2 observations to provide cloud-free data streams. His research projects cover a wide spectrum, from regional farming studies to large-scale initiatives like the ESA WorldCereal project, always aiming to bridge the gap between cutting-edge machine learning technologies, the reality of large-scale operationalization, and the science of remote sensing. |
Kristof Van Tricht, VITO10:00-10:30 | Morning Coffee Break |
10:30-11:15 | Keynote 2 – “Scaling Ground Truth to Advance Earth Observation” The rapid progress of machine learning, exemplified by breakthroughs like large language models, has been powered by massive, diverse datasets that capture human knowledge—such as the vast troves of text data available on the internet. We hypothesize that similarly transformative potential exists in Earth observation (EO), where deep learning models can unlock insights from complex real-world imagery, provided they have access to large, richly labeled datasets. However, creating such datasets presents unique challenges in scale and specificity. This talk explores a few approaches to scaling ground truth for EO applications, including using roadside imagery for crop type mapping, leveraging simulations for GHG plume detection, and mining OpenStreetMap annotations for object detection. Together, these examples illustrate how innovating methods for dataset creation can drive new insights and capabilities in EO. Bio Sherrie Wang is an Assistant Professor at MIT in the Department of Mechanical Engineering and Institute of Data, Systems, and Society. Her research uses novel data and computational algorithms to monitor our planet and enable sustainable development. Her focus is on improving agricultural management and mitigating climate change, especially in low- or middle-income regions of the world. To this end, she frequently uses satellite imagery, crowdsourced data, LiDAR, and other spatial data. Due to the scarcity of ground truth data in these regions and the noisiness of real-world data in general, her methodological work is geared toward developing machine learning methods that work well with these constraints. |
Sherrie Wang, Massachusetts Institute of Technology11:15-11:45 | Posters spotlights I
|
11:45-13:15 | Lunch Break |
13:15-13:45 | Posters spotlights II
|
13:45-14:15 | Best papers presentations
|
14:15-15:15 |
15:15-15:45 | Afternoon coffee break |
15:45-16:30 | Keynote 3 – “Closing the Gap: Building Planetary AI Systems for Real-World Users” Earth observation (EO) data is indispensable for tackling urgent global challenges such as rising sea levels, wildfires, food security, climate change, and disaster relief. Yet building EO-based models that are both high-performing and globally applicable remains technically difficult. These models often lack robustness across diverse real-world tasks and are rarely updated or maintained. We will describe how AI2 delivers performant, resource-efficient, and reliable EO models in near real-time using a diverse set of sources—from NASA and ESA satellites to streaming GPS data. We’ll explore what it means to build research from a production mindset, and how this approach helps close the gap between the lab and real-world impact by centering users from the start. Finally, we will introduce Earth System, a new initiative at AI2 focused on large-scale, highly multimodal planetary foundation models, and the infrastructure needed to tune and deploy them at scale. Earth System will be a fully open-source platform designed to support the entire lifecycle of planetary modeling and deployment—enabling resource-efficient research, scalable development, and impactful real-world applications. Patrick Beukema is a Principal Engineer and Scientist leading AI efforts within the Applied Science division at the Allen Institute for Artificial Intelligence (Ai2). He directs initiatives focused on large-scale, mixed-modality planetary modeling, supporting real-world applications in climate science, wildfire management, agriculture, maritime intelligence, conservation, and ecosystem health. Patrick holds a Ph.D. in neuroscience from the Center for the Neural Basis of Cognition (Carnegie Mellon and University of Pittsburgh), where he studied neural plasticity and neural decoding. Prior to Ai2, he worked across startups, tech, and academia on topics including recurrent neural networks, causal inference, and applied machine learning. Affiliation: Allen Institute for Artificial Intelligence (Ai2). |
Patrick Beukema, 16:30-17:30 | Poster Session (poster boards #419 – #443) |
19:00-on | Workshop dinner |
Challenge
This year’s data challenge is organized by the Embed2Scale consortium and revolves around Lossy Neural Compression for Geospatial Analytics. Specifically, participants are asked to develop an encoder to compress SSL4EO-S12 data cubes down to 1024 features, a.k.a. embeddings, that are then evaluated on a hidden set of applications, a.k.a. downstream tasks as in the literature of foundation models.
You will have three weeks in March 2025 to develop your encoder and test the quality of your embeddings by interacting with this challenge portal Eval.AI submitting embeddings for data available at HuggingFace. The 1st week of April will allow you three submissions in three days based on a separate dataset made public on HuggingFace two days before. This phase determines the final leaderboard and challenge winner who will present their solution at the workshop in Nashville or remotely as per availability.
Important Dates:
Development phase: Mar 10-31
Testing phase: Apr 3-5
Winner presentation: At the EarthVision WS
Webpage: eval.ai/web/challenges/challenge-page/2465
Challenge Data in HuggingFace
pretraining: huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1
development and evaluation: huggingface.co/datasets/embed2scale/SSL4EO-S12-downstream
Development Phase ended, Final Submission for Ranking ahead!
The development phase closed. Now get ready to compress approx. 90GB from huggingface.co/datasets/embed2scale/SSL4EO-S12-downstream/tree/main/data_eval (available for download Apr 1) for submission starting Apr 3 through Apr 5.
Your scores will determine the final ranking to find our winners (cf. below). We are excited about your solution, all the best luck!
About CVPR EARTHVISION winners
In addition to the winner according to github.com/DLR-MF-DAS/embed2scale-challenge-supplement?tab=readme-ov-file#leader… we decided to also invite the solution with the highest `q_mean` score to present their solution at the CVPR EarthVision workshop. Each of the winning teams will receive a cash prize of 1k EUR as support to come to Nashville, TN, USA for the presentation. On April 7, 2025 we’ll get in touch with the two winning teams through the email ID they provided to Eval.AI .
And even should you not win the thing …
… we are so grateful you participated! We are also very happy to link the code to your solution. If you’d like so, pls open a corresponding issue in github.com/DLR-MF-DAS/embed2scale-challenge-supplement/issues.
Organizers
- Ronny Hänsch, German Aerospace Center, Germany,
- Devis Tuia, EPFL, Switzerland,
- Jan Dirk Wegner, University of Zurich & ETH Zurich, Switzerland,
- Nathan Jacobs, Washington University in St. Louis, USA
- Loïc Landrieu, ENPC ParisTech, France
- Charlotte Pelletier, UBS Vannes, France
- Hannah Kerner, Arizona State University, USA
Technical Committee
- Aayush Dhakal, Washington University in St Louis
- Aimi Okabayashi, Université Bretagne Sud
- Alex Levering, VU Amsterdam
- Alexandre Xavier Falcao, IC-UNICAMP
- Amanda Bright, National Geospatial-Intelligence Agency
- Anastasia Schlegel, DLR
- Ankit Jha, The LNM Institute of Information Technology, Jaipur
- Antoine Bralet, UBS/IRISA
- Begum Demir, TU Berlin
- Biplab Banerjee, Indian Institute of Technology, Bombay
- Caleb Robinson, Microsoft
- Camille Couprie, Facebook
- Camille Kurtz, Université Paris Cité
- Christian Heipke, Leibniz Universität Hannover
- Christopher R Ratto, JHUAPL
- Claudio Persello, University of Twente
- Clement Mallet, IGN, France
- Conrad M Albrecht, German Aerospace Center
- Corentin Dufourg, Univ. Bretagne Sud / IRISA
- Dalton Lunga, Oak Ridge National Laboratory
- Damien Robert, University of Zurich
- Daniel Iordache, VITO, Belgium
- Diego Marcos, Inria
- Dimitri Gominski, University of Copenhagen
- Elliot Vincent, Ecole nationale des ponts et chaussées / Inria / IGN
- Emanuele Dalsasso, EPFL
- ESTHER ROLF, Harvard University
- Ewelina Rupnik, Univ Gustave Eiffel, LASTIG, ENSG-IGN
- Ferda Ofli, Qatar Computing Research Institute, HBKU
- Franz Rottensteiner, Leibniz Universitat Hannover, Germany
- Gabriel Tseng, NASA Harvest
- Gabriele Moser, Università di Genova
- Gedeon Muhawenayo, Arizona State University
- Gemine Vivone, CNR-IMAA
- Gencer Sumbul, EPFL
- Georgios Voulgaris, University of Oxford
- Guillaume Astruc, ENPC/IGN
- Gülşen Taşkın, İstanbul Teknik Üniversitesi
- Gustau Camps-Valls, Universitat de València
- Helmut Mayer, Bundeswehr University Munich
- Islam Mansour, German Aerospace Center (DLR) & ETH Zurich
- Jacob Arndt, Oak Ridge National Laboratory
- Jakob Gawlikowski, German Aerospace Center (DLR)
- Javiera Castillo Navarro, Cnam
- Joëlle Hanna, University of St. Gallen
- Jonathan Giezendanner, MIT
- Linus M. Scheibenreif, University of St. Gallen
- Lukas Drees, University of Zurich
- M. Usman Rafique, Kitware Inc.
- Marc Rußwurm, Wageningen University
- Mareike Dorozynski, Institute of Photogrammetry and Geoinformation
- Martin Weinmann, Karlsruhe Institute of Technology
- Matt Leotta, Kitware
- Michael Mommert, Stuttgart University of Applied Sciences
- Michael Schmitt, University of the Bundeswehr Munich
- Miguel-Ángel Fernández-Torres, Universidad Carlos III de Madrid
- Minh-Tan Pham, IRISA-UBS
- Myron Z Brown, JHU
- Nicolas Audebert, IGN
- Nikolaos Dionelis, ESA
- Ragini Bal Mahesh, German Aerospace Center – DLR
- Ramesh S. Nair, Planet Labs PBC
- Roberto Interdonato, CIRAD
- Scott Workman, DZYNE Technologies
- Sophie Giffard-Roisin, Univ. Grenoble Alpes
- Srikumar Sastry, Washington University in St Louis
- Subit Chakrabarti, Floodbase
- Sylvain Lobry, Université Paris Cité
- Tanya Nair, Floodbase
- Tatsumi Uezato, Hitachi, Ltd
- Teng Wu, Univ Gustave Eiffel, ENSG, IGN, LASTIG, F-94160 Saint-Mandé
- Thibaud Ehret, AMIAD, Pôle recherche
- Utkarsh Mall, Columbia University
- Valerio Marsocci, KU Leuven
- Yohann PERRON, LIGM ENPC
- Zhijie Zhang, University of Arizona
- Zhuangfang Yi, Regrow
Challenge Sponsors
Affiliations
.
Submissions
1. Prepare the anonymous, 8-page (references excluded) submission using the ev2025-template and following the paper guidelines.
2. Submit at cmt3.research.microsoft.com/EarthVision2025. Please submit via this link only.
Policies
A complete paper should be submitted using the EarthVision templates provided above.
Reviewing is double blind, i.e. authors do not know the names of the reviewers and reviewers do not know the names of the authors. Please read Section 1.7 of the example paper EarthVision2025.pdf for detailed instructions on how to preserve anonymity. Avoid providing acknowledgments or links that may identify the authors.
Papers are to be submitted using the dedicated submission platform: cmt3.research.microsoft.com/EarthVision2025.
The submission deadline is strict.
By submitting a manuscript, the authors guarantee that it has not been previously published or accepted for publication in a substantially similar form. CVPR rules regarding plagiarism, double submission, etc. apply.
The Microsoft CMT service was used for managing the peer-reviewing process for this conference. This service was provided for free by Microsoft and they bore all expenses, including costs for Azure cloud services as well as for software development and support.
CVPR 2025
CVPR is the premier annual computer vision event comprising the main conference and several co-located workshops and short courses. With its high quality and low cost, it provides an exceptional value for students, academics and industry researchers.
