ELLIS UnConference Workshop - AI for Earth and Climate Sciences
Copenhagen (Denmark), December 2, 2025
This ELLIS UnConference workshop organized by the ELLIS Unit Jena is co-located with EurIPS Copenhagen and will take place on December 2, 2025.
The workshop aims to foster a vibrant dialogue at the intersection of computer science, Earth system science, and environmental sciences, bringing together experts who share a vision of leveraging artificial intelligence to tackle some of the planet’s most pressing challenges.
Participants will exchange ideas on how advanced AI methodologies—from machine learning and deep neural networks to explainable AI and physics-informed modeling—can open new frontiers in climate prediction, environmental monitoring, and ecosystem understanding. A special emphasis will be placed on integrating physical knowledge with data-driven approaches, ensuring that models not only learn from vast Earth observation datasets but also remain faithful to underlying scientific principles. The workshop will also highlight the importance of data quality, robustness, and interpretability, reflecting the need for trustworthy and transparent AI systems in high-stakes societal and ecological contexts.
Beyond technical discussions, the workshop seeks to cultivate a collaborative spirit: connecting communities that often work in parallel, inspiring joint projects, and identifying the scientific and societal questions where AI can have the greatest impact. By bringing together insights from computer science and Earth system research, we aim to illuminate pathways toward innovative tools and practices that advance sustainable development and deepen our understanding of our changing planet.
Venue
Bella Center Copenhagen
Martha Christensens Vej 10, Indgang 3, 2300 København S, Denmark
Room 18, Seat Capacity: 97
Call for Submissions
We warmly invite submissions to the AI for Earth and Climate Sciences workshop. Accepted contributions will be selected for presentation as a talk or poster on December 2, 2025. Submissions may include both previously published and new, unpublished work. Please note that this workshop is separate from the ELLIS UnConference poster session.
Submission Guidelines
- Application: Please submit your abstract via the following form: https://survey.academiccloud.de/index.php/979696?lang=en
- Review: Submissions will be reviewed for relevance and quality. Accepted abstracts will be selected for presentation as posters or short talks. You may indicate your preferred presentation format, and we will do our best to accommodate your preference.
- Archival Policy: This workshop is non-archival. Authors are encouraged to submit preliminary, in-progress work, or recently published material.
Important Dates
- Submission Deadline: October 24, 2025
- Notification of Acceptance: October 30, 2025
- Workshop: December 2, 2025
Workshop Agenda (9:00 – 15:00)
09:00 – 09:15 Welcome and Introduction
ELLIS Unit Jena and GENAI-X
The ELLIS Unit Jena brings together leading researchers in machine learning, computer vision, and computational life sciences from the University of Jena, the Max Planck Institute for Biogeochemistry, and the German Aerospace Center (DLR). With a strong focus on reliable, data-efficient, and science-informed AI methods, the unit aims to bridge theoretical advances with real-world impact across medicine, biology, and environmental sciences. GENAI-X, Jena’s forthcoming initiative for next-generation AI and data-driven scientific discovery, complements this mission by fostering cross-disciplinary collaboration, large-scale datasets, and cutting-edge computational infrastructure. Funded with €6 million by the Carl Zeiss Foundation for the period April 2026–March 2031, GENAI-X addresses one of AI’s fundamental challenges: achieving robust generalizability in non-stationary environmental systems where conditions vary across space and evolve unpredictably. Focusing on hydro-climatic extremes—including floods, landslides, droughts, and late-frost events—and their associated impacts, GENAI-X advances AI methods that adapt to shifting data patterns and uncertainties. The project integrates hybrid modeling, causal modeling, equation discovery, dimension reduction, and uncertainty quantification to improve environmental understanding and prediction. Beyond methodological innovation, GENAI-X develops AI-driven tools for risk assessment, interpretation of environmental change, and decision support.
Profile and Publications
https://inf-cv.uni-jena.de/home/staff/denzler/
https://scholar.google.com/citations?user=bhpi3vgAAAAJ&hl=de
09:15 – 09:30 Impulse Talk
09:30 – 10:00 Keynote Talk
AI for Action: Geospatial Intelligence for Humanitarian Response and Climate Resilience
Artificial intelligence has the potential to transform how we respond to climate impacts and humanitarian crises—especially when combined with Earth observation, open geospatial data, and local knowledge. This keynote will explore how AI-powered tools can support decision-making in challenging environments, from disrupted infrastructure after disasters to urban areas facing health and climate risks. Focusing on applied, real-world use cases, the talk will highlight how integrating physical models with machine learning enables more robust, interpretable, and adaptive systems—for example, in routing under disaster conditions, participatory risk mapping, or estimating exposure to climate-sensitive health risks such as vector-borne diseases. The presentation will emphasize collaboration across disciplines and sectors, showing how AI can contribute not just to understanding our changing planet, but to actively shaping more resilient futures.
Profile and Publications
https://www.geog.uni-heidelberg.de/de/personen-am-institut/dr-steffen-knoblauch
https://scholar.google.com/citations?user=bnVdx2YAAAAJ&hl=de
10:00 – 10:30 Short Talks
Thematic Group A [Earth Observation, Environmental Monitoring & Biodiversity] (Part 1)
Spring barley yield prediction in southern Sweden using multi-source satellite data and random forest approach
Remote sensing observations and artificial intelligence algorithms have emerged as key components for crop yield estimation at various scales during the past decades. However, the utilization of multi-source satellite data and machine learning for estimating aggregated crop yield at regional level in Europe has been explored scarcely. Our study aims to bridge this research gap by focusing on the district-level spring barley yield estimation in southern Sweden from 2017 to 2022. We developed an estimation method with the Random Forest (RF) approach using four satellite-derived products along with two climate variables. These variables were used individually and in combinations as inputs for the RF approach. The results showed that vegetation indices (VIs) outperformed SIF in barley yield estimation, while combining VIs and SIF variables achieved the highest model performance for yield estimation (R2 = 0.77, RMSE = 488 kg/ha). The inclusion of climate variables generally had little added contributions to the model performance. Importantly, barley yield prediction could be achieved two months prior to harvest, using monthly VIs and SIF data from April and May. Our study demonstrated the feasibility of using freely accessible satellite data and machine learning approaches for estimating crop yield at the pan-European regional level. We expect that our proposed methodology can be extended to different crop types and regional-scale crop yield estimation in Europe, benefiting national and local authorities in making agricultural productivity decisions.
Profile and Publications
https://www.lunduniversity.lu.se/lucat/user/2eee61378e73be07ad42f5072ab9ec3d
https://portal.research.lu.se/en/persons/xueying-li
ECHOSAT: Estimating Canopy Height Over Space and Time
Forest monitoring is critical for climate change mitigation. However, existing global tree height maps provide only static snapshots and do not capture temporal forest dynamics, which are essential for accurate carbon accounting. We introduce ECHOSAT, a global and temporally consistent tree height map at 10 m resolution spanning multiple years. To this end, we resort to multi-sensor satellite data to train a specialized vision transformer model, which performs pixel-level temporal regression. A self-supervised growth loss regularizes the predictions to follow growth curves that are in line with natural tree development, including gradual height increases over time, but also abrupt declines due to forest loss events such as fires. Our experimental evaluation shows that our model improves state-of-the-art accuracies in the context of single-year predictions. We also provide the first global-scale height map that accurately quantifies tree growth and disturbances over time. We expect ECHOSAT to advance global efforts in carbon Monitoring and disturbance assessment.
Profile and Publications
https://www.wi.uni-muenster.de/department/mlde/people/jan-pauls
https://scholar.google.com/citations?user=FHghFnMAAAAJ
10:30 – 11:00 Coffee Break and Poster session
11:00 – 12:30 Short Talks
Thematic Group A [Earth Observation, Environmental Monitoring & Biodiversity] (Part 2)
Physics-Informed Low Rank Adaptation for Interpretable Earth Observation
The ultimate goal of Earth Observation (EO) is to retrieve interpretable and physically meaningful representations. Scientists often model Earth system physical processes to understand and quantify the physical causation mechanisms behind observations. However, these processes are inherently complex and often noisy. As a result, physical models inevitably introduce simplifications, leading to discrepancies between model predictions and measured observations. Existing approaches to address inverse problems for Earth observation often overlook model incompleteness or make assumptions about the unmodelled components during data preprocessing, resulting in implausible outcomes that limit their applicability.
To account for the incompleteness of physical models, we introduce data-driven auxiliary variables — originally developed for inverting idealised dynamical systems — to the inversion of static EO operators of practical significance. Inspired by the Low Rank Adaptation techniques originally proposed for fine-tuning large language models, we introduce Physics-Informed Low Rank Adaption (PhyLoRA) — a physics-informed learning framework with a low rank residual to integrate static EO operators of varying incompleteness.
We view the radiative transfer process of the Earth surface as a planet-scale renderer, and its inversion as inverse graphics with an auto-encoder architecture. Additionally, we apply our method to the inverse problems arising from geophysics to allow a comprehensive understanding of our living planet from surface to underground. We validate the effectiveness of our method on two disparate static EO operators: a complex forest radiative transfer model from optical remote sensing and a volcanic deformation model from geophysics. Extensive experiments demonstrate that our approach outperforms the state-of-the-art in inverting incomplete physical models, and reveal insights on key aspects of the inverse problem, offering a compelling pathway for interpretable Earth observation and beyond.
Profile and Publications
https://www.cst.cam.ac.uk/people/ys611
https://scholar.google.com/citations?hl=zh-CN&user=MHWgmoIAAAAJ&view_op=list_works
Do Foundation Models Disentangle Change? An Embedding-Based Benchmark Grounded in Disasters
Foundation models are increasingly used in Earth observation, yet their ability to distinguish different types of change remains uncertain. In this exploratory study, we assess whether current geospatial foundation models can separate meaningful disaster-induced changes from natural variations such as seasonality, illumination, or sensor differences. We introduce a benchmark based on multi-temporal, multi-modal satellite imagery annotated over real flood events, enabling systematic evaluation of model robustness and representation quality. Our analysis suggests that existing models often conflate distinct sources of change, highlighting open challenges for developing more interpretable and reliable AI systems for environmental and disaster monitoring.
Profile and Publications
https://www.kth.se/profile/hfang
https://scholar.google.se/citations?user=39qlz0YAAAAJ&hl
Modeling and Evaluating Uncertainty in Canopy Height Estimation using Quantile Regression
Accurate tree height estimation is vital for ecological monitoring and biomass assessment. We apply quantile regression to existing tree height estimation models to incorporate uncertainty quantification. Most of the current approaches for tree height estimation rely on point estimates, which limit their applicability in risk-sensitive scenarios. In this talk, we show that existing deep neural networks can be adapted to provide statistically calibrated uncertainty estimates via quantile regression by modifying the corresponding head of the network architecture. Furthermore, we demonstrate how our results align with known challenges in remote sensing, such as terrain complexity or vegetation heterogeneity, showing reduced model confidence in such conditions.
Profile and Publications
https://www.wi.uni-muenster.de/de/institut/mlde/personen/karsten-schroedter
https://mlde-lab.github.io/team/karsten-schroedter/
Leveraging Generative Deep Learning for Fine-Scale Precipitation Retrieval from Infrared Satellite Observations
Accurate precipitation estimation is critical for climate research, water resource management, and disaster mitigation. Satellite infrared observations provide wide coverage but have a weak correlation with rainfall. This makes fine-scale precipitation retrieval particularly challenging, especially for rare and extreme events that are poorly represented in the data. Here, we explore deterministic
and generative deep learning models to estimate precipitation at 1 km resolution over metropolitan France, leveraging 15 years of radar and geostationary infrared data. Deterministic models, including U-Net variants Ronneberger et al. [2015], provide robust mean rainfall estimates, while generative models, such as GANs Goodfellow et al. [2014] and diffusion models Ho et al. [2020], capture the full distribution, including rare and extreme precipitation events. By comparing deterministic and conditional generative models, our study addresses the challenge of learning from weakly informative data, highlighting the trade-offs between pixel-wise accuracy and the diversity of predicted precipitation patterns. This work demonstrates how modern deep learning architectures can be applied to Earth observation tasks, producing realistic and informative precipitation estimates even with limited input information.
References
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. https://arxiv.org/abs/1406.2661.
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. https://arxiv.org/abs/2006.11239.
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. https://arxiv.org/abs/1505.04597.
Profile and Publications
https://latmos.ipsl.fr/en/directory/
https://scholar.google.com/citations?user=Z-eUzeYAAAAJ&hl=en&oi=ao
Thematic Group B [Weather & Climate Modeling]
Future wildfire risk in Southern Europe under changing land use and climate: A data-driven approach using ClimEx2
Wildfires are one of the most devastating natural disasters with wide impacts across economic sectors and society in Europe. Modeling wildfire risk remains a complex challenge, particularly as predictions of fire risk often rely on weather-based indices, such as the Fire Weather Index (FWI) which exclude key factors such as human activities and other ignition sources. Machine Learning (ML) has been shown to effectively quantify wildfire risk using diverse data sources such as weather, land cover, and human factors (Oliviera et al. 2012). It is, however, unclear how wildfire risk will evolve in the future under different climate scenarios, especially as it depends on both future climate and land use changes. The aim of the study is to train a ML model on historical data and to use the novel ClimEx2 simulations (Asselin, et al. 2024) to assess the future wildfire risk, comparing simulations which only include greenhouse gas (GHG) changes to simulations which include both GHG and land use changes (LUC) in the Shared Socio-economic Pathways (SSP) 1-2.6 and 3-7.0 scenarios. The best performing model is a XGBoost (XGB) model with an AUC of 0.91 and F1 score of 0.83. Explainable AI is used to explain the model output and identify the most influential predictors. Future projections of wildfire risk show that the wildfire risk will increase by the end of the century (2071-2100) in the SSP3-7.0, with a greater increase in the simulation including both LUC and GHG changes compared to the one including GHG only. The study highlights the importance of incorporating LUC as well as GHG changes to model future wildfire risk.
Profile and Publications
https://www.linkedin.com/in/oph%C3%A9lie-meuriot-668a4b156/?originalSubdomain=uk
https://scholar.google.com/citations?user=Kkf6CDYAAAAJ&hl=en
FedRAIN-Lite: Federated Reinforcement Algorithms for Improving Idealised Numerical Weather and Climate Models
Sub-grid parameterisations in climate models are traditionally static and tuned offline, limiting adaptability to evolving states. This work introduces FedRAIN-Lite, a federated reinforcement learning (FedRL) framework that mirrors the spatial decomposition used in general circulation models (GCMs) by assigning agents to latitude bands, enabling local parameter learning with periodic global aggregation. Using a hierarchy of simplified energy-balance climate models, from a single-agent baseline (ebm-v1) to multi-agent ensemble (ebm-v2) and GCM-like (ebm-v3) setups, we benchmark three RL algorithms under different FedRL configurations. Results show that Deep Deterministic Policy Gradient (DDPG) consistently outperforms both static and single-agent baselines, with faster convergence and lower area-weighted RMSE in tropical and mid-latitude zones across both ebm-v2 and ebm-v3 setups. DDPG’s ability to transfer across hyperparameters and low computational cost make it well-suited for geographically adaptive parameter learning. This capability offers a scalable pathway towards high-complexity GCMs and provides a prototype for physically aligned, online-learning climate models that can evolve with a changing climate. Code accessible at https://github.com/p3jitnath/climate-rl-fedRL.
Profile and Publications
https://www.turing.ac.uk/people/enrichment-students/pritthijit-nath-0
https://scholar.google.com/citations?user=KdzVBhQAAAAJ
CRPS-LAM: Regional ensemble weather forecasting from matching marginals
Machine learning for weather prediction increasingly relies on ensemble methods to provide probabilistic forecasts. Diffusion-based models have shown strong performance in Limited-Area Modeling (LAM) but remain computationally expensive at sampling time. Building on the success of global weather forecasting models trained based on Continuous Ranked Probability Score (CRPS) we introduce CRPS-LAM, a probabilistic LAM forecasting model trained with a CRPS-based objective. By sampling and injecting a single latent noise vector into the model, CRPS-LAM generates ensemble members in a single forward pass, achieving sampling speeds up to 39 times faster than a diffusion-based model. We evaluate the model on the MEPS regional dataset, where CRPS-LAM matches the low errors of diffusion models.
Profile and Publications
https://liu.se/en/employee/erila85
https://scholar.google.com.co/citations?user=sg-aEUUAAAAJ&hl=vi
Fourier Neural Operators with Temporal Continuity for Ocean Emulation
Neural operators are becoming the default tools to learn solutions to governing partial differential equations (PDEs) in weather and ocean forecasting applications. Despite promising early results, challenges such as long-term prediction stability and physical consistency remain. In neural operators, it is important to preserve the continuous–discrete equivalence to bridge the gap between discrete numerical implementations and the continuous operators being approximated. This becomes particularly critical in time-dependent PDEs, where the dynamical system is non-stationary and continuity in time is an inherent property of the underlying operator. In this study, we address the challenges of stability and accuracy in high-resolution ocean prediction through the incorporation of temporal structure into operator
learning. To this end, we compare the standard Fourier Neural Operator (FNO) architecture with its modified variant, FNOtD (FNO with temporal Dimension), which has been designed to internalize the dispersion relations while learning the solution operator of ocean PDEs. We show that parameterizing integral kernels in both spatial and temporal Fourier domains significantly improves the FNO architecture in terms of efficiency and generalization under limited data. Compared to standard FNO architecture, FNOtD demonstrates substantially improved long-term emulation stability and greater consistency with the underlying physical
dynamics in our experimental setup. Our results also demonstrate competitive predictive skill relative to a state-of-the-art three-dimensional numerical ocean model, while requiring significantly lower computational cost. This study represents a step toward developing accurate and stable emulators based on the FNO framework for high-resolution ocean dynamics and paves the way for learning effectively from sparse and limited observational data.
Profile and Publications
https://taltech.ee/en/contacts/road-engineering-and-geodesy-research-group/vahidreza-jahanmard
https://scholar.google.com/citations?user=8mWA8LoAAAAJ&hl=en
12:30 – 13:30 Lunch Break and Poster session
13:30 – 14:00 Keynote Talk
AI for Tree Ecosystem Monitoring
Abstract will be shared soon
Profile and Publications
https://di.ku.dk/english/staff/?pure=en/persons/400547
https://scholar.google.com/citations?user=d-jF4zIAAAAJ&hl=en
14:00 – 14:40 Short Talks
Thematic Group C [Carbon Cycle, Sustainability & Socio-Economic Systems]
Upscaling tropical peatland CO₂ fluxes in Southeast Asia with remote sensing and knowledge-guided machine learning
Improving estimates of greenhouse gas emissions from tropical peatlands in Southeast Asia at continental scales is critical for understanding their role in the global carbon budget. These ecosystems store vast amounts of carbon but are increasingly threatened by climatic and human disturbances, leading to high carbon dioxide (CO2) emissions. Current large-scale CO2 estimates are based on emission factors or global products, and these often fail to capture the inter-annual variability and fine-scale spatial heterogeneity of these ecosystems, potentially leading to flux misestimations.
The increasing availability of Earth Observation (EO) data and EO-derived products offers new opportunities for long-term monitoring across areas, capturing both temporal and spatial variability. However, translating remote sensing information into reliable flux estimates requires robust models able to identify patterns in large EO datasets. Although machine learning (ML) models are well suited for this, they lack embedded knowledge of underlying processes and consequently often fail to generalize well, which is crucial when upscaling to larger areas. To address this challenge, we use a knowledge-guided ML framework where we integrate ecophysiology with ML, with Eddy Covariance measurements providing CO2 flux data.
We use two simple process-based representations of photosynthesis and respiration processes, while ML captures residual complexity. For respiration, we also aim to further disentangle autotrophic and heterotrophic components, using EO-based information on vegetation structure and disturbance history alongside groundwater level. Here, we present preliminary results of capturing spatial and temporal CO2 flux dynamics in tropical peatlands in Southeast Asia with knowledge-guided machine learning.
Profile and Publications
https://ign.ku.dk/english/employees/geography-1/?pure=en/persons/836721
https://scholar.google.com/citations?user=XaJ9VYUAAAAJ&hl=en&oi=sra
Nowcasting SDG indicators across 193 countries with machine learning
Achieving the Sustainable Development Goals (SDGs) by 2030 requires timely, country-level estimates across all 234 indicators. However, persistent reporting lags in official SDG statistics hinder real-time indicator assessment. As a result, governments and multilateral funders lack cross-country evidence to prioritize targeted interventions. In this paper, we propose a scalable machine learning pipeline to nowcast country-level SDG indicator values across all 234 indicators for 193 countries. Using state-of-the-art time series models with exogenous covariates, we generate probabilistic nowcasts of country-indicator values with prediction intervals, thus filling current reporting gaps. We use this dataset for three key analyses: First, we classify current country-indicator trajectories (accelerating, stalling, reversing) to identify where course corrections are needed before 2030. Second, for indicators with defined targets or universal thresholds, we quantify the current distance-to-target, i.e., whether—and by how much—current levels meet or fall short of the target; for indicators without defined targets, we substitute transparent benchmarks, i.e., peer reference levels and improvements from the 2015 baseline. Third, we estimate the average annual change required from 2026–2030 to reach the 2030 target and report acceleration gaps relative to the current pace. We release all nowcasts, prediction intervals, and code to enable replication and downstream use. Our analyses provide a cross-country dataset that supports evidence-based policy, independent monitoring, and risk-informed allocation by multilateral donors. In turn, our findings guide policy and investment decisions to accelerate progress towards the SDGs and keep the 2030 targets within reach.
Profile and Publications
https://www.som.lmu.de/ai/en/institute/contact-page/kerstin-forster-e782f1fd.html
https://scholar.google.com/citations?user=Opuqm2kAAAAJ&hl=en
AI-Enhanced Carbon Simulation Framework: Towards Real-Time Enterprise Sustainability Planning
Traditional Integrated Assessment Models often require hours to days for scenario evaluation, limiting their practical application in enterprise sustainability planning. This talk presents a comprehensive AI framework combining physics-informed neural networks, uncertainty quantification, and efficient inference to enable rapid carbon simulation and decision support. Our approach achieves 8.3× computational acceleration while maintaining regulatory accuracy requirements (±7.2%), incorporates Sparse Polynomial Chaos Expansion for 20× faster uncertainty propagation, and delivers carbon pricing forecasts with 8-12% MAPE. The framework enables sub-minute carbon pathway simulation with quantified uncertainty bounds, supporting privacy-preserving deployment for ISO 14064-3 compliance. Experimental validation on established climate datasets demonstrates competitive performance against traditional IAMs while enabling practical climate-economic decision support for enterprise applications. This work addresses key barriers to enterprise AI adoption in sustainability planning through interpretable uncertainty quantification and efficient client-side deployment.
Profile and Publications
https://www.linkedin.com/in/anupam-srivastava1007/?trk=public_profile_browsemap&originalSubdomain=ie
https://orcid.org/0009-0006-2699-8587
Integrating Hydrological Modelling and Machine Learning with Earth Observation for Tropical Wetland Carbon Fluxes
Much of the uncertainty in global wetland methane emissions modeling comes from the tropics, where wetland mapping is poor and in situ measurements are limited. This PhD research will merge process-based hydrologic models with data-driven surrogate models to improve regional estimates of carbon flux in tropical wetlands. An integrated hydrologic model of the Usangu Wetlands in Tanzania will be calibrated with satellite remote sensing data and serve as the base model for data-driven model development. Here we will present the planned model setup for the base hydrologic model and the plan for how this will be calibrated with data from the Surface Water and Ocean Topography (SWOT) satellite. First results from the base model show significant untracked losses, likely due to abstractions for irrigation. We also present the plan for the second stage of the research, in which the traditional hydrological model is replaced by a machine learning surrogate such as a CNN or encoder-decoder network.
Profile and Publications
https://ign.ku.dk/english/employees/all-employees-ign/?pure=en/persons/837055
https://orcid.org/0009-0003-2269-4883
14:40 – 15:00 Closing Discussion
15:00 End of Workshop
09:00 – 09:15 Welcome and Introduction
Prof. Dr. Joachim Denzler, Friedrich Schiller University Jena, ELLIS Unit Jena Director
ELLIS Unit Jena and GENAI-X
09:15 – 09:30 Impulse Talk
Dr. Alexander Winkler, Max Planck Institute for Biogeochemistry
Title: TBC
09:30 – 10:00 Keynote
Dr. Steffen Knoblauch, Heidelberg University
AI for Action: Geospatial Intelligence for Humanitarian Response and Climate Resilience
10:00 – 10:30 2 short talks
Thematic Group A [Earth Observation, Environmental Monitoring & Biodiversity] (Part 1)
Dr. Xueying Li, PostDoc, Lund University
Predicting spring barley yield in southern Sweden using satellite remote sensing data and random forest approach
Jan Pauls, PhD Candidate, University of Münster
ECHOSAT: Estimating Canopy Height Over Space And Time
10:30 – 11:00 Coffee Break and Poster session
11:00 – 12:30 8 short talks
Thematic Group A [Earth Observation, Environmental Monitoring & Biodiversity] (Part 2)
Yihang She, PhD Candidate, University of Cambridge, ELLIS Jena Summer School Participant
Towards Explainable Earth Observation with Model Inversion and Bias Correction
Heng Fang, PhD Candidate, KTH Royal Institute of Technology
Do Foundation Models Disentangle Change? An Embedding-Based Benchmark Grounded in Disasters
Dr. Karsten Schrödter, PhD Candidate, University of Münster
Modeling and Evaluating Uncertainty in Canopy Height Estimation using Quantile Regression
Matthieu Meignin, PhD Candidate, Sorbonne Université
Leveraging Generative Deep Learning for Fine-Scale Precipitation Retrieval from Infrared Satellite Observations
Thematic Group B [Weather & Climate Modeling]
Dr. Ophélie Meuriot, PostDoc, Denmark Technical University, ELLIS Jena Summer School Participant
Future wildfire risk in Southern Europe under changing land use and climate: A data-driven approach using ClimEx2
Pritthijit Nath, PhD Candidate, University of Cambridge, ELLIS Summer School Jena Applicant
FedRAIN-Lite: Federated Reinforcement Algorithms for Improving Idealised Numerical Weather and Climate Models
Erik Larsson, PhD Candidate, Linköping University
CRPS-LAM: Regional ensemble weather forecasting from matching marginals
Vahidreza Jahanmard, Research scientist, Tallinn University of Technology
Fourier Neural Operators with Temporal Continuity for Ocean Emulation
12:30 – 13:30 Lunch Break and Poster session
13:30 – 14:00 Keynote
Prof. Dr. Christian Igel, University of Copenhagen
Title: TBC
14:00 – 14:30 3 short talks
Thematic Group C [Carbon Cycle, Sustainability & Socio-Economic Systems]
Laura van der Poel, PhD Candidate, University of Copenhagen, ELLIS Jena Summer School Participant
Upscaling tropical peatland CO₂ fluxes in Southeast Asia with remote sensing and knowledge-guided machine learning
Kerstin Forster, PhD Candidate, Ludwig Maximilian University of Munich
Nowcasting SDG indicators across 193 countries with machine learning
Anupam Srivastava, PhD Candidate, University College Dublin
AI-Enhanced Carbon Simulation Framework: Towards Real-Time Enterprise Sustainability Planning
Sarah Elizabeth Franze, PhD Candidate, DHI and Copenhagen University
Integrating Hydrological Modelling and Machine Learning with Earth Observation for Tropical Wetland Carbon Fluxes
14:40 – 15:00 Closing Discussion
15:00 End of Workshop
Contact: conrad.philipp@uni-jena.de