Abhishek Singh
MSc graduate in Data Science, University of Europe for Applied Sciences
Vision Foundation Models for Earth Observation, Weak Supervision, and Cross-Dataset Evaluation
Benchmark & Hackathon
Neural
Earth Fields
Learning continuous representations of our planet.
7–8 December 2026 · University of Bonn
24-hour overnight sprint · 5 teams · English
The application deadline passed on 16 October 2026. Applications are now closed.
Mission
Rethinking how we represent the Earth
Most geospatial information is stored as rasters, grids, vector layers, or other discrete formats. These representations are powerful, but often resolution-dependent, storage-intensive, and difficult to combine across heterogeneous data sources.
Neural Earth Fields explores a different paradigm: encoding geospatial information in neural-network parameters and querying it continuously from geographic coordinates.
Meet the participants
Participants
Ana Gabriela Mantilla Dulcey
Research Scientist at Universidad Industrial de Santander, with an MSc in Geophysics
Physics-Guided Machine Learning and Multiphysics Geophysical Inversion
Dheeraj Parkash
Machine learning researcher with research experience at INRIA/CNRS
Multimodal Learning and Continuous Representations for Scientific and Earth Data
Andrea Federici
PhD researcher at UiT The Arctic University of Norway
Adapting Earth Foundation Models for Weather Downscaling and Regional Generalization
Frederick Schindlegger
MSc student at the University of Münster and visiting researcher at IBM Research Europe
Geospatial Foundation Models, Dense Representations, and Sustainable Development Applications
Moa Sjöberg
PhD researcher in Meteorology at Uppsala University
Satellite-Based Sea Ice Analysis and Machine Learning
Sacha Lewin
PhD researcher at the University of Liège
Continuous Generative Models for Spatiotemporal Data and Global Weather Modeling
Daniel Rodrigues Perazzo
ELLIS PhD researcher at the Italian Institute of Technology, affiliated with the University of Genoa
Implicit Neural Representations, Neural Fields, and Geometric Modeling
Ananda Krishnan
Researcher at the Christian Doppler Laboratory GEOHUM, Salzburg, Austria
Self-Supervised Learning and Diffusion-Based Thermal Downscaling
Haoran Zhang
MPhil/PhD researcher at University College London
Geographic Domain Shifts and Spatial Representation Learning
Moritz Lucas
PhD researcher at the Joint Lab for Artificial Intelligence and Data Science, University of Osnabrück
Deep Learning for Tree Detection and Mapping in Aerial Imagery
Titus Top
PhD researcher in the EarthMapps research group at KU Leuven
Graph Neural Network Emulators for Ice-Sheet Dynamics and Sea-Level Projections
The idea
Earth data as a continuous, queryable function
Interdisciplinary teams will build compact neural models that take longitude and latitude as input and reconstruct multiple global geospatial layers. You bring the architecture; we provide the data, an evaluation harness, and a starter kit to get you running from day one. Benchmark results and reproducible submissions remain publicly available after the event, contributing to an open ecosystem for the community.
01
Represent
Store diverse geospatial fields implicitly in neural-network weights.
02
Query
Reconstruct information continuously at geographic coordinates.
03
Compare
Evaluate reconstruction quality together with parameter efficiency.
Benchmark
One model, diverse Earth data
Every team receives the same curated global dataset and a common evaluation framework. The initial benchmark brings together three complementary views of the planet.
Physical Earth
Water storage anomaly
A time series of total water storage anomaly maps derived from gravity observations.
Earth observation
ESA WorldCover
A high-resolution global land-cover map representing discrete surface classes.
Living Earth
iNaturalist
Biodiversity observations describing the spatial distribution of species.
Event
24 hours in Bonn
- Date
- 7–8 December 2026
- Format
- Noon-to-noon, overnight hackathon
- Location
- IMPULSE – House for Intellectual Innovation and Creativity
University of Bonn · Adenauerallee 131 · Bonn - Participants
- A deliberately small cohort: 10 participants in 5 two-person teams, matched by the organizers across Geo and AI expertise
- Language
- English
Event location
A house built for exchanging ideas
IMPULSE is the University of Bonn’s House for Intellectual Innovation and Creativity—a place for inspiration, intellectual encounter, and creative collaboration.
Its mission is grounded in the idea that academic work thrives through receiving and sharing impulses. As a dedicated hub for inter- and transdisciplinary innovation, it is a natural home for teams bringing geospatial science and AI together.
Adenauerallee 131 · 53113 Bonn
Learn more about IMPULSE ↗
Program
Learn, build, present
Day 1 · 12:00–16:00
Kickoff
Introductory modules, benchmark briefing, team formation, and technical setup.
Overnight · 16:00–08:00
Hackathon sprint
Code through the night with mentors on hand and the leaderboard live. Basic sleeping arrangements provided — bring a sleeping bag.
Day 2 · 09:00–12:00
Finals
Consolidation, team presentations, jury review, awards, and closing.
Who should apply?
Build a Geo + AI team
We are looking for Master's students and PhD candidates with a background in either geospatial science or machine learning/AI — you don't need both. The organizers match participants into two-person teams across complementary expertise. Selection aims for disciplinary, international, and gender diversity.
See full criteria and timeline →Applications
Interested in Neural Earth Fields?
We are looking for early-career researchers — Master's students and PhD candidates at any stage — with strong expertise in either geospatial science or machine learning/AI. Teams are intentionally matched across both domains, so you do not need to cover both.
Good candidates bring
- Domain knowledge in geodesy, geography, Earth observation, geoinformatics, or a related field; or experience in ML, deep learning, computer vision, or applied mathematics
- Genuine curiosity about neural representations of geospatial data
- Readiness to work through an intensive overnight coding sprint with your team
- Their own laptop — software tools and a starter kit will be provided
What we cover
- Travel costs within Europe (train preferred)
- All food and drinks throughout the event (day and night)
- Very basic rest and sleeping arrangements at IMPULSE House — the sprint runs through the night; bring a sleeping bag and prepare for little sleep
- Networking opportunities with fellow participants, mentors, and organizers
Top teams receive a certificate and a commemorative trophy. All reproducible submissions remain on the public benchmark leaderboard after the event.
The application deadline passed on 16 October 2026. Applications are now closed.
- 16 October 2026 Application Deadline
- 30 October 2026 Acceptance Notification
- Early November 2026 On-boarding Meeting Online
- 7–8 December 2026 Event — University of Bonn
Team
Organizers
Marc Rußwurm
University of Bonn
Jürgen Gall
University of Bonn
Jürgen Kusche
University of Bonn
Konstantin Klemmer
UCL
Alistair Francis
Asterisk Labs
Isaac Corley
Taylor Geospatial
Kyle Gao
Aalto University
Supported by
Partners supporting Neural Earth Fields
The benchmark and hackathon are made possible by partners across the University of Bonn and the international geospatial research community.
