In 2022, a series of heat waves, droughts and wildfires marked a summer of extremes in Europe. Understanding such events, the Earth’s climate and climate change as a whole requires contributions from a great variety of scientific disciplines - and modern AI plays an increasingly important role in this. With the ELLIS program ‘Machine Learning for Earth and Climate Sciences‘, leading researchers from across Europe join forces to advance our understanding of the Earth system and to foster excellence in this field.
ELLIS Fellow Gustau Camps-Valls, a Full Professor in Electrical Engineering at the University of Valencia and Group Leader of the Image and Signal Processing (ISP) group, and ELLIS Fellow Markus Reichstein, Director of the Biogeochemical Integration Department at the Max-Planck-Institute for Biogeochemistry, are heading the program. In this interview they explain, why modern AI can contribute to climate science, and what kind of projects the members of their program are currently working on.
Understanding and counteracting climate change requires multidisciplinary cooperation on a global scale. How can modern AI research contribute to this challenge?
Markus Reichstein: “Understanding climate change and its consequences means understanding a very complex system. The reductionist approach with the idea that everything can be modelled from first principles is an important idea that has led to physics-based climate and Earth system models. But we see that many uncertainties have not been reduced much during the last 15-20 years. On the other hand, in the last decade, a plethora of Earth observations have been gathered. This offers the chance to build data-driven models and complement the reductionist modelling approaches. But the observed data must be used in a meaningful way, the key patterns detected and the most relevant information extracted. AI is mighty and efficient for this and can be used where classical models are either too slow or very uncertain, for example for biological processes. This helps diagnose droughts, floods, fires and other extreme events and their impact, as well as early warning or anticipation of disasters and facilitating evidence-based decision making.”
Can you give us some concrete examples of breakthroughs in this field?
Markus Reichstein: “A significant breakthrough has been the fully data-driven quantification of major global carbon cycle components. A spatially sparse, but temporally dense set of in-situ observations of CO2 uptake and release by ecosystems needed to be combined with satellite remote sensing and weather forecast data to give estimates of CO2 fluxes ‘everywhere, all the time’. The data-driven fields have then been used to benchmark climate models which are giving climate projections, for example in the IPCC reports. Given the non-linear relationships, this was only possible with modern machine learning approaches.”
Gustau Camps-Valls: “Yet, AI and deep learning can help address many other important challenges with great societal and environmental impact; from the automatic detection of wildfire risk to the anticipation of droughts, heatwaves and floods. Weather nowcasting and forecasting are now typically aided by statistical learning methodologies that help improve for example the nowcasting precipitation prediction task, and the detection of extreme weather patterns. Assessing, quantifying and understanding the impacts of extremes are important applications of AI: from understanding the causality of climate-induced migrations to the intricate relations of economy-climate-conflicts leading to food insecurity and societal crises.”
With your ELLIS program you are dedicated to modelling and understanding the Earth system with machine learning and process understanding. How exactly do you and the members of your program advance science in this field?
Markus Reichstein: “To make a leap forward in Earth system and climate science, it is important that domain and AI scientists work closely together. Our challenge is about modelling this complex system, where the real world is abstracted into mathematical equations. Traditionally, these have been based on physical, chemical, ecophysiological and ecological theories leading to a system of differential equations (mechanistic modelling). This is now complemented by a data-driven approach, where the specific machine learning algorithms want to be defined for each scientific question. Interestingly, this also brings together different fields of Earth system science, for example land, ocean and atmospheric science. But also various machine learning aspects play an important role, for instance XAI or robust AI. Currently, one of the key topics is to integrate mechanistic modelling with machine learning into hybrid modelling approaches, which are data-adaptive but respect domain scientific laws at the same time.”
Gustau Camps-Valls: Yes, exactly, and all this is being achieved by organizing joint webinars and intensive workshops, which generate a lot of interest in different communities, both applied and theory research. Concrete research ideas are then followed in joint EU, ESA and ERC-funded projects, as well as concrete developments. An important far-end goal that I’d like to highlight is that our developments can also reach other fields of science and engineering, where very often one has observational data and mechanistic models, assumptions and domain knowledge. AI tools that can blend them and reconcile paradigms can be tremendously useful in other arenas.”
What kind of projects are you currently working on in your program?
Markus Reichstein: “One of the flagship projects is the ERC-SyG USMILE, where almost a dozen ELLIS ML4ECS researchers are involved. This fundamental research project melts together land and atmosphere Earth system science with machine learning as outlined above. The goal is an improved understanding of how the feedbacks in the Earth system work and ultimately to achieve better predictions of climate and its impacts.”
Gustau Camps-Valls: “Other example projects are more applied: In the H2020 DEEPCUBE project, we are interested in better understanding the impact of drought on vegetation, as well as the drivers of induced migration and displacement of people. Two other projects highlight the role of AI on Earth and climate science: the H2020 XAIDA project funded by the EU Commission and the European Space Agency ESA DeepExtremes project, both dealing with the detection, forecasting and attribution of extreme events - mainly droughts - with deep learning and their impacts in Earth Observation data.”
Fostering cooperation among leading scientists across Europe is one of the objectives of your program. What kind of researchers are part of your program, and what do interested scientists have to do to join it?
Markus Reichstein: “A broad range of scientists is associated in the program, from physicists, mathematicians and biologists to environmental engineers, geo-ecologists and computer scientists. But also in terms of scale: from very local scale looking at ecosystems, cities and landscape to global scale addressing the whole Earth, and from hourly time-scale to centennial. The program would benefit from quantitative social scientists and from machine learners strong in theory, although this latter knowledge is also nicely leveraged by interacting with other ELLIS programs.”
Gustau Camps-Valls: “I think that researchers in ELLIS programs working on interpretability, causality and robust ML will be a definite asset.”
AI is a rapidly evolving research field, and its importance is increasing in nearly every area of society. If you take a look into the future: Which developments do you expect for your research area?
Gustau Camps-Valls: “AI should not only be focused on attaining accurate predictions, which is answering the ‘what’ question, but on helping domain scientists with the more challenging ‘why’ and ‘what if’. In Earth and climate sciences we are often asking these questions, and for that we have traditionally used model simulations and mechanistic models. The exciting field of hybrid AI that blends data and domain knowledge could certify or refute already established rules. Explainable AI (XAI) could help in understanding AI models, gain confidence in them, and shed light on their governing -not necessarily physical- principles. And finally, the field of causal discovery from observational data and hypothesis could help in understanding the complex and coupled system, as well as imagining actions, for example interventions of humans in the system. ELLIS is a fantastic environment for exploring all these venues because of the heterogeneity of backgrounds and because of the possibility to interact in a relaxed atmosphere.”
Markus Reichstein: “ELLIS is clearly becoming an important crystallization point for us, locally, regionally and Europe-wide.”
Learn more about the ELLIS research program “Machine Learning for Earth and Climate Sciences” and its members here.
Read about the latest program workshop here.
Find an overview of all ELLIS research programs here.
Get in touch with ELLIS here.
Follow ELLIS on