In many scientific domains, the goal is not simply to classify an image or produce a visually plausible segmentation, but to estimate quantities that scientists and decision-makers can use: river width, water extent, sediment concentration, habitat change, or other physical variables. These tasks require models that are robust to distribution shift, efficient under limited supervision, and reliable enough to support downstream measurement. In this talk, I will discuss machine learning and computer vision methods for this setting, with a focus on geospatial imagery and hydrology.

Scientific domains such as remote sensing expose several fundamental limitations of current vision systems. Models must operate across sensors, geographic regions, seasons, and acquisition conditions, often with limited labeled data and multispectral inputs that differ from standard natural image benchmarks. They must also detect small or thin structures, handle temporal variation, and avoid errors that may appear minor under conventional vision metrics but lead to substantial errors in scientific measurements.

I will discuss three aspects that try to solve these problems: learning transferable representations that can adapt across tasks and datasets, developing data-efficient perception methods for multispectral and geospatial imagery, and translating model predictions into reliable physical measurements. I will highlight work on task transfer, satellite representation learning, river segmentation, and downstream estimation of river width and other hydrological variables. More broadly, this work aims to develop robust, transferable, and data-efficient machine perception systems that can turn large-scale satellite and scientific imagery into reliable measurements of the physical world.

Rangel Daroya is a PhD student in Computer Science at the University of Massachusetts Amherst, where she works in the Computer Vision Lab with Prof. Subhransu Maji. Her research focuses on computer vision and machine learning, particularly transfer and multi-task learning and applications to environmental remote sensing. Her recent work includes RiverScope, a high-resolution river segmentation dataset, and WildSAT, which learns satellite representations from wildlife observations. Rangel received the 2025 AGU Remote Sensing Student Award for RiverScope and passed her PhD Portfolio with Distinction. Her work has appeared at CVPR, ICCV, WACV, AAAI, and IEEE JSTARS, including a CVPR 2024 Highlight paper.

Reliable forecasting of the Earth system is essential for mitigating natural disasters and supporting human progress. Traditional numerical models, although powerful, are extremely computationally expensive. 

Recent advances in artificial intelligence (AI) have shown promise in improving both predictive performance and efficiency, yet their potential remains underexplored in many Earth system domains. Here we introduce Aurora, a large-scale foundation model trained on more than one million hours of diverse geophysical data. 

Aurora outperforms operational forecasts in predicting air quality, ocean waves, tropical cyclone tracks and high-resolution weather, all at orders of magnitude lower computational cost. With the ability to be fine-tuned for diverse applications at modest expense, Aurora represents a notable step towards democratizing accurate and efficient Earth system predictions. These results highlight the transformative potential of AI in environmental forecasting and pave the way for broader accessibility to high-quality climate and weather information. Link to paper.

Ana Lucic is an assistant professor in artificial intelligence at the University of Amsterdam working on interpretability, AI for science and AI safety. Previously, she was a researcher at Microsoft Research Amsterdam and at the Partnership on AI. She has a PhD in explainable machine learning from the University of Amsterdam, along with an MSc and BSc, both in mathematics, from McMaster University. Ana is a member of the European Laboratory for Learning and Intelligent Systems (ELLIS) and the Amsterdam ELLIS Unit. Her work has been published in venues such as Nature, ICML, AAAI, AISTATS, CVPR and FAccT.

This project addresses one major challenge in climate science: accurately quantifying how forests remove carbon dioxide (CO₂) from the atmosphere. Forests currently account for nearly one-third of the global land carbon sink, but their capacity to absorb CO₂ is increasingly threatened by climate change(s), land-use change(s), and environmental disturbance(s). At the same time, afforestation, reforestation, and carbon offset initiatives are expanding rapidly, despite persistent concerns about the reliability and transparency of the methods used to estimate carbon sequestration.

The INFOSCO project (Integrating Forest Structure and Carbon Observatories) proposes a new approach by shifting the focus from forest-scale carbon assessments to continuous physiological estimates at the level of individual trees. By combining ecosystem CO₂ flux observations, laser-based measurements of forest structure, and theory of plant physiological processes, it aims to estimate the contribution of each tree to overall forest carbon uptake. This will provide a more detailed and biologically grounded understanding of carbon sequestration while remaining consistent with established micrometeorological measurements and theory.

Methodologically, the research integrates data and concepts from ecosystem ecology, laser proximal sensing, plant physiology, and statistical modelling into a unified framework. The core modelling strategy uses a Bayesian hierarchical structure to quantify uncertainties at multiple levels. By means of Dirichlet regression the method ensures that tree-level estimates are consistent with ecosystem-scale micrometeorological observations.

Beyond improving carbon accounting, the project aims to bridge between detailed field-based ecological measurements and AI models scalable to biogeographic applications. By linking high-resolution observations of individual trees with ecosystem-atmosphere fluxes, it seeks to advance both our understanding of forest physiology and the development of more robust tools for monitoring and verifying nature-based climate solutions.

Jaime C. Revenga was born in Madrid, Spain, and has developed his career across France, Austria, Switzerland, and Denmark. Trained as energy and environmental engineer, his previous research has focused on developing machine learning applications in biogeography, environmental remote sensing (UAV-LiDAR), and modelling of terrestrial ecosystem-atmosphere interactions.

Jaime earned his PhD in Geosciences at the University of Copenhagen (2020-2023), developing UAV-LiDAR and ML methods for biomass and carbon stock estimation in forests and croplands.

He is currently a Postdoctoral Researcher at the Statistics and Probability Theory group, at the Dept. of Mathematical Sciences, University of Copenhagen, on the INFOSCO Project, where he models time-resolved tree-level carbon metabolism and atmosphere-biosphere carbon interactions, from terrestrial laser scanning and eddy-covariance observations.

His work now focuses on (i) applications for ecosystem physiology predictive modelling and (ii) method development in environmental imaging spectroscopy.