Zum Inhalt springen

Webinar Series AI for Earth and Sustainability Science​

Since February 2023 the AI for Earth and Sustainability Science webinar series highlights seminal and recent progress in AI-enabled modelling and understanding of the Earth system from local to global scale and AI-science based diagnosis, prediction and remedy of environmental crises.

Interdisciplinary researchers from academia, industry, UN and government agencies as well as NGOs will talk about tackling systemic real-world challenges with AI, including disaster risk reduction and preparedness, environmental degradation, climate change, societal impacts and dynamics, and the sustainable and responsible use of natural resources such as water and energy.

The webinar series is curated by the ELLIS Unit Jena and the ELLIS program “Machine Learning for Earth and Climate Sciences” in collaboration with AI for Good.

 

Curators

Gustau Camps-Valls

University of Valencia

Maria Piles

University of Valencia

Markus Reichstein

MPI for Biogeochemistry

Joachim Denzler

University of Jena

OUR SPEAKERS AT A GLANCE

MISSED A PRESENTATION?
ALL RECORDINGS ARE AVAILABLE ONLINE.

How deep learning improves weather forecasts and reduces energy demands

Mapping environmental variables continuously across space and time is a key challenge in environmental science and provides essential baseline information to support Sustainable Development Goal (SDG) 15, which seeks to protect, restore, and promote the sustainable use of terrestrial ecosystems and halt biodiversity loss. Remote sensing and machine learning algorithms have emerged as powerful tools in this domain, enabling the prediction of ecological variables in areas where direct measurements are unavailable, often at a global scale. However, recent research has highlighted significant challenges in using machine learning for large-scale spatial mapping, stemming from the characteristics of remote sensing datasets and the distribution of reference data. This talk aims to raise awareness of these challenges and to propose ideas for addressing them, with the goal of developing more robust environmental spatial datasets that can better support the goals of SDG 15.
Hanna Meyer
Feb 26, 2025

Adapting Machine Learning for Atmosphere-Biosphere Coupling in Earth System Models

Understanding the interactions within the Earth system is key to predicting change and assessing the impacts of human activities. Ecosystem responses to these changes are complex due to the diversity of organisms, and we lack fundamental laws to describe them. As the volume of ecosystem-atmosphere interaction data grows, identifying consistent response patterns becomes possible, though challenges remain. While machine learning (ML) has advanced in fields like computer vision, it needs adaptation for Earth system sciences. The mismatch in spatial scales complicates integrating observations into Earth system modeling. This presentation discusses integrating mechanistic land models with observation-informed ML methods. We focus on three critical processes: (1) ML and causality methods to detect CO2 effects on ecosystems, influencing the land carbon sink in climate projections, (2) ML techniques to model phenological changes and their feedbacks on energy, water, and carbon fluxes, and (3) a physics-constrained ML approach to infer stomatal conductance and integrate it into Earth system models. Finally, we propose a pathway for enhancing ML-driven Earth system models.
Alexander Winkler and Christian Reimers
Feb 12, 2025

Advancing earth observation and climate research through self-supervised learning: from land monitoring to cloud modelling

Earth Observation (EO) satellites produce vast amounts of data annually, yet limited labeled annotations hinder supervised machine learning. Self-supervised learning (SSL) overcomes this by learning from unlabeled data and fine-tuning with minimal labeled examples. Our research showcases SSL in two areas: land monitoring and 3D cloud modeling. For land monitoring, we pre-train models on Sentinel-1 SAR and Sentinel-2 imagery, fine-tuning for tasks like vegetation prediction and land cover mapping. SSL reduces the need for labeled data significantly. In cloud modeling, we apply SSL to geostationary imagery, fine-tuning on sparse radar profiles. This approach generates accurate 3D cloud maps, improving climate model precision. SSL thus unlocks the full potential of EO data for Earth system science.
Anna Jungbluth
Jan 29, 2025

Causality in Dynamical Systems

In many real-world applications, predicting how a system reacts under an active perturbation is critical. Achieving this requires robust causal methodology. While many causal methods and theoretical results have been developed for i.i.d. (independent and identically distributed) data, real-world data often arises from dynamical systems where temporal structure cannot be ignored. This session highlights the need for adapting causal methodology to address these dynamic contexts. Using two examples, we demonstrate how leveraging the temporal structure of dynamical systems can uncover new possibilities. Specifically, we explore how these structures enable unique assumptions that help eliminate the effects of hidden confounding. We use this insight to separate the effects of internal variability and external forcing in Earth system science and to estimate price elasticities in the electricity market.
Jonas Peters
Jan 27, 2025

How deep learning improves weather forecasts and reduces energy demands

Deep learning weather prediction (DLWP) models are about to outperform traditional numerical weather prediction (NWP) models. In this talk, we will review the evolution of deep learning for weather prediction and hypothesize what we may expect in the next years and what not. Beyond that, we will search for explanations by contrasting potentials and limitations of today‘s DLWP and NWP models. Among those, we will understand how DLWP can drastically reduce energy consumption and where machine learning helps us in understanding land-atmosphere interactions processes.
Matthias Karlbauer
Jan 15, 2025

Toward robust surrogate components of land surface states for hybrid earth system models

The land surface, including the biosphere, is an important component of earth system models (ESMs) and weather forecasting systems: It provides terrestrial boundary conditions and surface fluxes of energy, water and carbon over land inform parameters of atmospheric processes. At this interface, ESMs profit from hybridasation where the numerical surface parameterisations are substituted with differentiable emulators. This can strongly reduce runtimes of forward and backward procedures, such as ensemble generation for uncertainty quantification, or deriving adjoint models in data assimilation. Neural networks as powerful function approximators seem suitable candidates, yet they tend to violate physical contraints and consequently can cause instabilities in numerical model predictions. This talk will overview model development parts of my PhD in the environmental sciences and synthesise aspects of constraining neural networks from a land surface forecasting perspective. We explore different strategies for enforcing physical constraints at the example of an self-standing simple process-model for carbon turnover. Further, we compare the potential of different machine learning approaches for land surface emulation with ECMWF‘s land surface scheme ECLand. Finally we outline how a combining approaches targets a robust surrogate model with explainability at best.
Marieke Wesselkamp
DEC 4, 2024

Deep learning for large-scale ecosystem monitoring

Tree-based ecosystems play a critical role in carbon sequestration, biodiversity, and the production of timber and food. Understanding how these ecosystems are influenced by climate change and human management necessitates a more comprehensive global characterization of woody resources. Recent advancements in remote sensing and machine learning-based computer vision have made such detailed analysis possible. For example, applying deep learning techniques to high-resolution satellite imagery enables the large-scale mapping of individual trees. The biomass of each tree, and thus its carbon content, can be estimated from the crown size using data-derived allometric equations. These equations are non-decreasing, and their integration into machine learning models not only enhances scientific validity but also acts as a strong regularizer. In this session, we will discuss a specific application in Rwanda, where assessing tree carbon stocks supports the development of strategies to achieve the country’s sustainable development goals. The presentation will conclude with an introduction to recently released open-source image datasets that aid in training machine learning models for various environmental monitoring tasks.
Christian Igel
NOV 6, 2024

Deep Generative Models for Molecular Simulation

This session will explore advanced methods in deep generative modeling for molecular simulations, focusing on a novel approach to improve the accuracy and efficiency of normalizing flows in approximating complex target distributions, such as Boltzmann  distributions of physical systems. Traditional methods for training flows often struggle with mode-seeking behavior, use samples from the target generated beforehand by expensive MCMC methods, or use stochastic losses that have high variance. reliance on costly MCMC-generated samples, or high-variance stochastic losses. We will introduce Flow AIS Bootstrap (FAB), a new technique that integrates annealed importance sampling (AIS) to overcome these limitations. By minimizing the mass-covering α-divergence with α=2, FAB reduces importance weight variance and generates samples in regions where the flow poorly approximates the target, aiding in the discovery of new modes. The session will highlight how FAB accurately approximates multimodal targets, outperforming previous methods. To the best of our knowledge, we are the first to learn the Boltzmann distribution of the alanine dipeptide molecule using only the unnormalized target density, without access to samples generated via Molecular Dynamics (MD) simulations: FAB produces better results than training via maximum likelihood on MD samples while using 100 times fewer target evaluations. After reweighting the samples, we obtain unbiased histograms of dihedral angles that are almost identical to the ground truth.
José Miguel Hernández Lobato
Oct 2, 2024

Confronting Climate Change with Generative and Self-supervised Machine Learning

The stunning recent advances in AI chatbots rely on cutting-edge generative deep learning algorithms and architectures trained on massive amounts of text data. Generative deep learning has also shown remarkable results when trained on video data and on combinations of different data types (i.e., multi-modal). The recent advances in generative deep learning can also benefit a variety of applications for addressing climate change. For example, generative deep learning trained on climate and weather data can be a powerful tool in generating an ensemble of weather predictions and in quantifying the uncertainty of long-term projections of climate change. As opposed to text and video, the relevant training data for this domain includes weather and climate data from observations, reanalyses, and even physical simulations. As in many massive data applications, creating “labeled data” for supervised machine learning is often costly, time-consuming, or even impossible. Fortuitously, in very large-scale data domains, “self-supervised” machine learning methods are now actually outperforming supervised learning methods. In this lecture, I will survey our lab’s work developing generative and self-supervised machine learning approaches for applications addressing climate change, including detection and prediction of extreme weather events, and downscaling and temporal interpolation of spatiotemporal data. Our methods address problems such as forecasting the path and intensity of tropical cyclones, renewable energy planning, and projecting future sea-level rise.
Claire Monteleoni
SEP 11, 2024

Harnessing Machine Learning and Satellite Data for Planetary-Scale Impact

Remote sensing satellites capture petascale, multi-modal data capturing our dynamic planet across space, time, and spectrum. This rich data source holds immense potential for addressing local and planetary-scale challenges including food insecurity, poverty, climate change, and ecosystem preservation. Fully realizing this potential will require a new paradigm of machine learning approaches capable of tackling the unique character of remote sensing data. Machine learning approaches must be flexible enough to make use of the multi-modal multi-fidelity satellite data, process meter-scale observations over planetary scales, and generalize to the challenging diversity of remote sensing tasks. In this talk, I will present examples of how we are developing machine learning approaches for planetary data processing including self-supervised transformers for remote sensing data. I will also demonstrate how treating ML research and deployment as a unified approach instead of siloed steps leads to research advances that result in immediate societal impact, highlighting examples of how we are partnering directly with stakeholders to deploy our innovations in areas of critical need across the globe.
Hannah Kerner
SEP 4, 2024

Tracking permafrost landscape dynamics in a rapidly warming Arctic

The Arctic is warming at an unprecedented pace, which leads to rapid landscape changes, particularly in the terrestrial Arctic which is largely underlain by permafrost. These landscape dynamics have strong impacts on local (geohazards, infrastructure) to global (climate system) scales. The integration of Machine Learning (ML) and Artificial Intelligence (AI) techniques with remote sensing data has become the standard approach for mapping these intricate landscape transformations within Arctic permafrost. Yet, what is going on in the Arctic remains still unknown in many places. Here we specifically dive into our analysis tracking Retrogressive Thaw Slumps (RTS), which are typical landscape processes of thawing and degrading permafrost. We will show the advances and challenges of using Deep-Learning techniques for tracking these comparably small, but dynamic landscape features across the Arctic. Finally, we will dig into further AI based applications in permafrost science, tracking changes in geomorphological features, but also tracking wildlife related to the massive changes in the northern permafrost region.
Ingmar Nitze
FEB 28, 2024

Fast statistical inference with neural networks and amortisation: Golden ticket or red herring?

Neural networks can provide solutions to tasks that were inconceivable just a few years ago and have benefitted society in numerous ways. These benefits primarily stem from a property often referred to as “amortisation“: Training a neural network usually requires significant effort and resources but, once trained, the network can solve similar problems repeatedly and rapidly with virtually no additional computational cost. Hence, the substantial initial training cost of training neural networks is “amortised” over time. Amortisation can also be used to enable fast inference with parametric statistical models: Once a network is trained using observational data as input and inferential statements (e.g., point parameter estimates) as output, the network can make inference with future data in a tiny fraction of the computing time needed by conventional likelihood or Monte Carlo methods. These amortised inferential tools have several compelling advantages over classical methods: they do not require knowledge of the likelihood function, are relatively easy to implement, and facilitate inference at a substantially reduced computational cost. In this lecture I will first give a brief review of recent work that has leveraged the property of “amortisation” in statistical inference in the context of (spatial) environmental and geophysical applications. I will then evaluate the merits and drawbacks of amortised inference from a statistician’s perspective and conclude by outlining the challenges that need to be overcome for these inferential tools to gain widespread acceptance.
Andrew Zammit-Mangion
FEB 14, 2024

Physics-constrained machine learning for scientific computing

Intact ecosystems, such as forests, are equally important for animals and humans and represent the basis of life on our planet. The integrity of these ecosystems is worldwide at risk due to the impending climate catastrophe and the encroachment of humans on natural habitats. The maintenance and promotion of biodiversity is crucial for maintaining a healthy environment. One important tool for this is comprehensive monitoring of wildlife in their natural habitats. In recent years, drones have been increasingly used for this purpose, often in conjunction with artificial intelligence. This talk will discuss the use of light-field technology with drones for wildlife observation. While imposing special demands on data collection and processing, light-field technology offers the possibility of precisely localizing animals in 3D and allows visualization of animals on the forest floor. The speaker will also put special emphasis on the ongoing research project BAMBI, which focuses on wildlife monitoring in Austrian forests.
Danielle Maddix Robinson
JAN 31, 2024

Explain it to me! On the use of explainable machine learning for the agricultural and environmental sciences

Climate models are essential in policy making. Current climate models, however, are quite impractical for answering quick ‘what-if’ questions, because they are too slow. For example, ‘what-if we plant one trillion trees’ or ‘what if we increase the carbon tax’. This talk will overview a doctoral thesis defense on how deep learning could create approximations of climate models that can quickly answer such questions. We call those approximations ’emulators’. After introducing climate emulators, I will address a core challenge on how to develop emulators on physical, chaotic, multiscale data using Multiscale Neural Operators. Lastly, I will present a roadmap towards an online tool that uses deep learning emulators to create accessible simulations of climate policy impacts.
Ribana Roscher
JAN 17, 2024

Large-scale monitoring of nature with deep learning

Applying nutrients to agricultural systems is critical to maximizing crop yields while minimizing the negative environmental impacts of fertilizers, such as nitrous oxide emissions or groundwater pollution. These goals can be achieved by timely calculating crop nutrient demand, which allows for precise fertilization management. In this talk, I will discuss non-invasive AI techniques to detect nutrient deficiencies from RGB images captured by smartphones or UAVs. Since it is very difficult to acquire training data of crops with nutrient deficiencies in the field, I will discuss how generative adversarial networks (GANs) can be trained in low data regimes where only one or very few training images are available. I will show that the generated images can be used as a source for data augmentation to improve the segmentation accuracy of rare classes, one-shot semantic image segmentation, or one-shot video object segmentation. In the second part of the talk, I will discuss AI approaches for forecasting wildfires and vegetation.
Jan Dirk Wegner
JAN 10, 2024

Airborne light fields for wildlife observation

Intact ecosystems, such as forests, are equally important for animals and humans and represent the basis of life on our planet. The integrity of these ecosystems is worldwide at risk due to the impending climate catastrophe and the encroachment of humans on natural habitats. The maintenance and promotion of biodiversity is crucial for maintaining a healthy environment. One important tool for this is comprehensive monitoring of wildlife in their natural habitats. In recent years, drones have been increasingly used for this purpose, often in conjunction with artificial intelligence. This talk will discuss the use of light-field technology with drones for wildlife observation. While imposing special demands on data collection and processing, light-field technology offers the possibility of precisely localizing animals in 3D and allows visualization of animals on the forest floor. The speaker will also put special emphasis on the ongoing research project BAMBI, which focuses on wildlife monitoring in Austrian forests.
David C. Schedl
DEC 20, 2023

Building fast emulators in climate modeling

Climate models are essential in policy making. Current climate models, however, are quite impractical for answering quick ‘what-if’ questions, because they are too slow. For example, ‘what-if we plant one trillion trees’ or ‘what if we increase the carbon tax’. This talk will overview a doctoral thesis defense on how deep learning could create approximations of climate models that can quickly answer such questions. We call those approximations ’emulators’. After introducing climate emulators, I will address a core challenge on how to develop emulators on physical, chaotic, multiscale data using Multiscale Neural Operators. Lastly, I will present a roadmap towards an online tool that uses deep learning emulators to create accessible simulations of climate policy impacts.
Björn Lütjens
DEC 6, 2023

Using AI for sustainable agriculture and forecasting

Applying nutrients to agricultural systems is critical to maximizing crop yields while minimizing the negative environmental impacts of fertilizers, such as nitrous oxide emissions or groundwater pollution. These goals can be achieved by timely calculating crop nutrient demand, which allows for precise fertilization management. In this talk, I will discuss non-invasive AI techniques to detect nutrient deficiencies from RGB images captured by smartphones or UAVs. Since it is very difficult to acquire training data of crops with nutrient deficiencies in the field, I will discuss how generative adversarial networks (GANs) can be trained in low data regimes where only one or very few training images are available. I will show that the generated images can be used as a source for data augmentation to improve the segmentation accuracy of rare classes, one-shot semantic image segmentation, or one-shot video object segmentation. In the second part of the talk, I will discuss AI approaches for forecasting wildfires and vegetation.
Jürgen Gall
NOV 22, 2023

Novel approaches to model assessment and interpretation in geospatial machine learning: addressing spatial dependence and high dimensionality

As the interpretability and explainability of artificial intelligence decisions has been gaining attention, novel approaches are needed to develop diagnostic tools that account for the unique challenges of environmental data, especially the spatial dependence of measurements and the high dimensionality of feature spaces. These are addressed by novel methods presented in this contribution with examples such as the regionalization of pollutants in the environment, and remotely-sensed mapping of mountain permafrost features as an essential climate variable and potential hazard.
Alexander Brenning
NOV 8, 2023

Machine learning for Earth Observation: Emissions, mitigation and sequestration

Constellations of satellites such as the SENTINELs provide almost daily coverage of the earth using a variety of sensors: multispectral and hyperspectral images, synthetic aperture radar, LIDAR, etc. Recent advances in machine learning allow us to process these images at planet scale on a near real-time basis and to paint a very broad picture of key climate related topics such as greenhouse gas emissions, wildfire risk, flooding or carbon sequestration in forests. This talk will cover some of these examples in depth, illustrating some key open numerical challenge.
Alexandre d’Aspremont
OCT 25, 2023

Food security and AI: From data collection to early warning systems

Characterizing the socio-economic status of populations and providing reliable, up-to-date estimates of who the most vulnerable are, how many of them there are, where they reside, and why they are vulnerable, is essential for governments and humanitarian organizations to make informed and timely decisions regarding the implementation of humanitarian assistance policies and programs. Anticipatory and preventive actions have been demonstrated to yield significant benefits compared to post-disaster and post-shock aid. It is, therefore, crucial to understand the drivers of food insecurity in a context where extreme weather events play an increasingly significant role in exacerbating already critical situations. The work of WFP’s Hunger Monitoring Unit revolves around collecting reliable food security data, which is then analyzed in conjunction with climate, economic, and conflict data to support early warning systems, including the use of machine learning.
Giulia Martini
OCT 11, 2023

Using machine learning to track human development at global scale and high spatial resolution

The Human Development Index (HDI) is widely used by policymakers and academics to summarize three key dimensions of wellbeing: the population’s health, human capital, and standard of living. A more comprehensive measure of wellbeing than income or wealth alone, HDI is used to categorize countries by their level of human development, which, in turn, can determine allocations of global resources. However, the United Nations releases official global estimates of HDI annually only at the highly aggregated national level, preventing their use for many policy and aid applications. This project builds on recent advances in machine learning and satellite imagery to develop the first global estimates of HDI for second-level administrative units (e.g., municipalities/counties) and for a global ~10km × 10km grid. To accomplish this, we develop and validate a generalizable downscaling technique based on satellite imagery that allows for training and prediction with observations of arbitrary shape and size. Our results indicate that more than half of the global population was previously assigned to the incorrect HDI quintile within each country, due to aggregation bias resulting from lower resolution estimates.
Hannah Druckenmiller
SEP 27, 2023

Physics-informed machine learning to push the ocean frontier in climate

The global ocean is central to the planet’s health and modulates global levels of heat and carbon, biological productivity, and sea level. However, open questions remain about what drives the circulation which hinders our understanding and ability to monitor ongoing, rapid changes. Climate models suggest that the ocean surrounding Antarctica, a critical region, is changing. However, the limited observations in one of the Earth’s most extreme and inaccessible environments poorly constrain the physical drivers. Here, machine learning is used to construct hypotheses that lead to new theoretical understanding of the circulation and to design a monitoring framework that can assess sensitivity to climate change. Monitoring the circulation is challenging because observations are generally limited to sparse data from the surface. With our theoretical insight, we developed a new physics-informed methodology to fill this gap using available data. The monitoring method Tracking global Heating with Ocean Regimes (THOR) can ‘reason’ using geophysical fluid dynamics. It is explicitly transparent and consists of a series of neural networks that combine eXplainable AI and Bayesian confidence scores for its predictions. We reveal differences in model physics that cause model divergence and spread in projections, opening the door to further discovery and observational strategies.
Maike Sonnewald
MAY 24, 2023

Global vegetation monitoring with probabilistic deep learning

Mapping forest structure on a global scale is an important component for understanding the Earth’s carbon cycle and conserving biodiversity. Several new space missions have been developed to support these goals by measuring forest structure to estimate biomass and carbon stocks. To analyze the vast amount of remote sensing data, efficient modelling approaches are needed that are robust to the inherent noise in these data. Data-driven approaches, especially modern deep learning methods, promise great potential for interpreting and combining data from different space missions to estimate vegetation parameters with enhanced spatial and temporal resolution. This talk will present recent research results for global vegetation height mapping exploiting ESA’s Sentinel-2 optical imagery and NASA’s GEDI full waveform LIDAR. By integrating probabilistic deep learning approaches (i.e. deep ensembles), the predictive uncertainty of the models is quantified, which is crucial for downstream applications that depend on reliable estimates.
Nico Lang
MAY 10, 2023

Remote sensing enables monitoring life above and under water

The oceans are a major source of biodiversity and food. The preservation of the quality of the foodchain, as well as of life underwater is a primary concern for society. However, observing water media routinely is a complex task that remote sensing technology empowered by AI can unlock. This talk will present research avenues where the use of remote sensing can help monitoring marine ecosystem, both above – via the monitoring of floating plastic debris – and under the water surface – via 3D reconstruction and mapping of coral reefs.
Devis Tuia
MAY 10, 2023

Unraveling the Effects of Climate Extremes on Ecosystems: towards predictive capabilities

Climate change is causing an increase in the frequency and intensity of extreme events such as heatwaves and droughts, which have severe consequences on ecosystems and human well-being. Understanding the complex dynamics of these events is crucial for developing robust models and scenarios for the future. This talk will discuss the challenges and opportunities of leveraging machine learning to enhance our understanding of the impact of climate extremes on ecosystems. Using local case studies on the impacts of climate extremes on forests and other ecosystems, we will demonstrate the benefits of utilizing AI and high-resolution satellite-derived data cubes for addressing this quest. We will also present the DeepExtremes project, funded by the European Space Agency, aiming to develop explainable AI methods for global extreme event impact detection and prediction. This presentation will invite the community to address this pressing environmental issue together. By effectively combining Earth system science, Earth system data cube approaches and AI, we expect to advance the management of consequences of extreme events and contribute to the global effort to use AI for good.
Miguel Mahecha
APR 26, 2023

Mapping the urban forest across North America: A case study in Computer vision for large-scale environmental monitoring

Generalization to novel domains is a fundamental challenge for computer vision. Near-perfect accuracy on benchmarks is common, but these models do not work as expected when deployed outside of the training distribution. To build computer vision systems that solve real-world problems at global scale, we need benchmarks that fully capture real-world complexity, including geographic domain shift, long-tailed distributions, and data noise. We propose urban forest monitoring as an ideal testbed for studying and improving upon these computer vision challenges, while working towards filling a crucial environmental and societal need. Urban forests provide significant benefits to urban societies. However, planning and maintaining these forests is expensive. One particularly costly aspect of urban forest management is monitoring the existing trees in a city: e.g., tracking tree locations, species, and health. We introduce a new large-scale dataset that joins public tree censuses from 23 cities with a large collection of street level and aerial imagery, containing over 2.5M trees and 300 genera. This benchmark enables the exploration of automated urban forest monitoring across modalities and with respect to geographic distribution shifts, vital for such a system to be deployed at-scale.
Laure Zanna
APR 26, 2023

Towards physics-AI hybrid modeling in hydrology: Opportunities and challenges

Recent advances in AI provide unprecedented opportunities to predict and understand components of the hydrological cycles. Nonetheless, many critical issues remain unresolved. For instance, with the remarkable predicting power of machine learning demonstrated in the field, it is under debate whether and how hydrological theory can still have a place in hydrological models or even benefit from “black-box” machine learning methods. In this presentation, we introduce two case studies that attempt to incorporate hydrological knowledge and machine learning in a unified framework. The first study develops a novel deep learning architecture that is encoded with a conceptual hydrological model, resulting in an end-to-end hydrology-AI hybrid learning system. The simulation results show that the hybrid model has improved prediction accuracy, robust transferability, and good intelligence for inferring unobserved processes. The second study develops an explainable machine learning approach to improve our understanding of runoff mechanisms, thereby gaining a deeper insight into how floods are changing under climate change. Overall, this presentation stresses how physics-AI integration can help us u
Shijie Jiang
APR 12, 2023

Deep learning and the dynamics of physical processes

Deep learning has been studied for a few years for the modeling of complex physical processes in industrial fields such as aeronautics or energy production and in scientific fields such as environment or health.  This area of research, although still emerging, is rapidly gaining momentum and developing as an interdisciplinary field. It raises new challenges for the interaction between machine learning and physics. This talk will focus on deep learning approaches for modeling dynamic physical systems and illustrate three main challenges: incorporating prior physical knowledge into learning models, generalizing learning models to multiple environments, and learning models operating continuously in space and time thus allowing flexible extrapolation at arbitrary spatiotemporal locations. This presentation will be illustrated by applications in different domains.
Patrick Gallinari
MAR 29, 2023

Climate modeling with AI: Hype or Reality?

Climate simulations remain one of the best tools to understand and predict global and regional climate change. Uncertainties in climate predictions originate partly from the poor or lacking representation of processes, such as ocean turbulence and clouds, that are not resolved in global climate models but impact the large-scale temperature, rainfall, sea level, etc. The representation of these unresolved processes has been a bottleneck in improving climate simulations and projections. The explosion of climate data and the power of machine learning (ML) algorithms are suddenly offering new opportunities: can we deepen our understanding of these unresolved processes and simultaneously improve their representation in climate models to reduce climate projections uncertainty? This talk discusses the advantages and challenges of using machine learning for climate projections. The focus will be on recent work in which we leverage ML tools to learn representations of unresolved ocean processes – in particular, learning symbolic expression. Some of this work suggests that machine learning could open the door to discovering new physics from data and enhance climate predictions. Yet, many questions remain unanswered, making the next decade exciting and challenging for ML + climate modeling for robust and actionable climate projections. 
Laure Zanna
MAR 29, 2023

Long Short-Term Memory networks for large-scale rainfall-runoff modeling

Flooding is one of the most common and costly types of natural disaster, with disproportionate human impacts in developing countries and communities. Google has developed an AI-based river and inundation forecasting system, and is partnering with governments and water agencies around the world to provide real-time flood alerts directly to individuals, communities, and NGOs through existing Google information channels like Maps, Search, and Android Alerts. This talk will cover the background, development, and impact of this effort.
Grey Nearing
MAR 15, 2023

Long Short-Term Memory networks for large-scale rainfall-runoff modeling

This talk will neither be about the depths of deep learning nor the depths of hydrology. Rather it will be a mix of both worlds from the perspective of solving problems in one field with tools from the other. If you are interested to hear about real-world applications of deep learning models to engineering tasks and all the research behind it, this talk is for you. More specifically, the talk will revisit the outcomes of years of research in the intersection of deep learning and rainfall-runoff modeling and why a more than 30-year-old neural network architecture (Long Short-Term Memory, LSTM) seems to be a perfect fit for this application. It will also be the tale of a young researcher that tried many (too complicated?) things but always found out that simplicity wins. More than anything, This talk may inspire (young) researchers to look for interdisciplinary problems/projects and to learn about new tools/techniques from different fields.
Frederik Kratzert
MAR 1, 2023

The role of AI in achieving the Sustainable Development Goals

The emergence of Artificial Intelligence (AI) is rapidly progressing with wide impacts across many sectors. As part of AI for Good’s new series on AI for Earth and Sustainability Science, this session provides an assessment of the impact of AI on the achievement of the Sustainable Development Goals (SDGs) specifically looking at Sustainable Cities (SDG11) and Climate Change (SDG13). Using a consensus-based expert elicitation process, it is clear that AI can enable the accomplishment of 134 targets across all the goals, but it may also inhibit 59 targets. However, current research may be overlooking important aspects. The fast development of AI needs to be supported by the necessary regulatory insight and oversight for AI-based technologies to enable sustainable development. Failure to do so could result in gaps in transparency, safety, and ethical standards leading to less than optimal outcomes for the environment and related sustainability targets.
Ricardo Vinuesa
FEB 1, 2023

Our Supportive Encouraging Nurturing Facilitating Promoting Assisting Backing Advancing Sustaining Fostering Partners

Linkedin Instagram Youtube X-twitter
Linkedin Instagram Youtube X-twitter
Contact us

Dr. Conrad H. PHILIPP
European Laboratory for Learning and Intelligent Systems (ELLIS)
ELLIS Unit Jena | Project Coordinator