Spanish version in issue 42 of Geom@il
In the last five years or so, we have seen the incredible growth in the use of Artificial Intelligence (AI) in different aspects of our daily lives. All technological disciplines, and science in general, are not immune to this wave of applications. Thus, in 2023, the Global Geodetic Observing System (GGOS), the executive branch of the International Association of Geodesy (IAG), formed a working group or focus area (AI4G) to study and promote AI applications in geodesy. The use of AI in geodesy has not been without controversy. For many years, we have seen AI applications in digital image processing, for example. However, its use in the processing and analysis of GNSS measurements or in the calculation of a geoid model, to name just two topics, is something else entirely and some geodesists say, “That’s a different story.” To discuss this topic, we invite the researcher appointed by GGOS-IAG to coordinate this working group or focus area: Prof. Dr. Benedikt Soja.
Benedikt Soja is a distinguished Austrian professor and researcher born in the beautiful city of Vienna in 1988. In 2013, he received his Diplom-Ingenieur (MSc) in “Geodesy and Geophysics” from the Vienna University of Technology, and in 2016, he completed his PhD in “Geodesy and Geoinformation” from the same university.
During his doctoral studies (2013-2016), he was an assistant on the “Real-Time VLBI Analysis” project at the GFZ Potsdam (GeoForschungs Zentrum – GeoResearch Center) in Germany.
Subsequently, until 2019, he was a Postdoctoral Researcher in the field of Geodynamics and Space Geodesy at the renowned NASA Jet Propulsion Laboratory (JPL), USA, later holding the position of Scientist II at the same institution. He then returned to Europe and, since 2020, has been Assistant Professor of Space Geodesy at ETH Zurich (Eidgenössische Technische Hochschule Zürich – Swiss Federal Institute of Technology Zurich) and Deputy Director of the Institute of Geodesy and Photogrammetry at the same university. He has received several awards and recognitions. Since May 2023, he has been the Chair of the GGOS (Global Geodetic Observing System) Focus Area on Artificial Intelligence for Geodesy.
For the editors of Geom@il, it is a true honor to have Benedikt Soja contribute to this issue, and we are very grateful for his participation.
- During your academic training and professional life, you have worked in different areas of geoscience, for example, VLBI, Kalman filtering, atmospheric studies, among others. What were the main motivations (personal or institutional) that led you to dedicate yourself more closely to Artificial Intelligence in Geodesy?
Throughout my academic journey in geodesy, I’ve always been interested in methodological aspects. A large part of my earlier work focused on Kalman filtering and its optimization for various geodetic applications. What fascinated me was that many AI methods are also fundamentally about solving optimization problems. The “learning” process in machine learning is essentially parameter estimation – just on a much larger scale. For example, in a neural network, we estimate weights and biases, typically using optimizers like stochastic gradient descent. While classical geodetic methods often estimate hundreds or thousands of parameters, ML methods handle even millions.
I became interested in AI around 2018–2019, initially inspired by breakthroughs like AI systems outperforming human players in complex games. That was a turning point where I started actively seeking opportunities to apply such methods in the field of geodesy.
In 2020, I joined ETH Zurich as a group leader, which gave me the platform to fully commit to this path. I was fortunate to receive strong institutional support from the start. ETH launched its AI Center shortly after I arrived, backed by major tech companies based in Zurich, such as Google. It now includes over 100 professors, and I’ve had the opportunity to collaborate with many of them.
More recently, I joined the Swiss AI initiative, which is working on developing foundation models for a wide range of domains. I’m currently involved in a climate-focused foundation model, and my personal goal is to integrate geodetic data to improve weather and climate predictions.
- What do you think are the most important applications of AI in geodesy? Which are the least explored? And what areas of geodesy are difficult to apply AI to? (If you think there are any)
AI has already shown strong results in spatial and temporal prediction of geodetic data. For example, in tropospheric and ionospheric modeling, machine learning has often outperformed traditional approaches. It has also proven useful for GNSS time series modeling, including denoising, anomaly detection, and identifying geophysical events. GNSS reflectometry applications like wind speed and soil moisture estimation have also greatly benefited. AI is also increasingly used in satellite gravimetry and Earth system modeling, especially for downscaling, reconstruction, and signal separation, as well as InSAR processing. Another promising area is Earth orientation parameter (EOP) prediction, as demonstrated in the Second EOP Prediction Comparison Campaign.
On the other hand, I initially thought that AI would have a lot of potential for automation in geodetic data processing, but it turned out quite challenging and is therefore still underdeveloped. Other areas, like terrestrial reference frame determination, use (by definition) simple models and do not gain as much from complex ML models, meaning that AI might only impact some of the preprocessing steps.
- What message would you have for colleagues geodesists who believe that AI cannot be used confidently in geodesy for various reasons, including, for example, its inability to predict or model patterns or physical phenomena inherent to these applications?
Confidence is indeed a key issue. In geodesy, providing realistic uncertainty estimates is essential for interpreting results. While often ignored in AI and ML applications, uncertainty quantification is actually very feasible using probabilistic machine learning. In fact, with careful modeling and calibration, we can often obtain more realistic uncertainties than from basic least-squares methods, which typically rely on assumptions like normality and uncorrelated observations.
It’s true that ML models are only as good as the data they are trained on. Predicting rare or extreme events is difficult if such events aren’t well represented in the training set. This is where physics-informed machine learning becomes valuable – it allows us to incorporate known physical laws into the ML framework, combining the strengths of both approaches. These hybrid methods have great potential and can help bridge the gap between traditional geodetic methods and modern AI techniques.
- In the SIRGAS (Geodetic Reference System of the Americas) community, hundreds of GNSS stations are processed daily, time series are developed, gravity and geoid models are calculated, and atmospheric studies are conducted. In fact, some universities are already conducting research with AI. What do you believe is the full and real potential for AI applications in SIRGAS, and how can the working groups (government institutes and universities) interact (through agreements or arrangements?) with the group you coordinate at GGOS?
It’s impressive what the SIRGAS community has built – both in terms of infrastructure and the quality of geodetic products. This large amount of high-quality, regularly processed data is a great resource for AI, since good training data is essential.
There’s already promising AI research focused on GNSS processing, such as enhancing station position time series or mitigating multipath effects. These efforts could contribute to enhancing the SIRGAS terrestrial reference frame. In principle, AI can help integrate diverse datasets more efficiently and potentially automate parts of the processing chain.
Improving geodetic products, such as those from SIRGAS, is the main goal of the GGOS Focus Area on AI for Geodesy (AI4G). This creates great potential for collaboration and knowledge exchange. Ideally, members of the community would join the AI4G Joint Study Groups to help facilitate this process.
- What expectations do you and the GGOS AI focus group have for the upcoming Symposium J02: Artificial Intelligence and Machine Learning in Geodesy, to be held during the IAG Scientific Assembly in Italy next September?
It’s a fantastic opportunity to have nearly 40 contributions from geodesists working on AI and ML topics all in one symposium, covering a wide range of areas. I’m not aware of any other conference that has brought together so many experts working on AI in geodesy.
Our main goal is to bring the community together to highlight the state of the art, enable collaborations, and foster engaging discussions and knowledge exchange around AI applications in geodesy.
- Any additional messages or reflections on the topic of AI in Geodesy?
To me, curiosity is a key driver in pushing AI forward in geodesy. Seeing what AI has achieved in other neighboring fields is a great source of inspiration. While geodesy has already seen important breakthroughs with AI, we are still at the beginning of this journey. I believe the next decade will be transformative. Education will play a central role – now is the time that geodesy curricula should begin to include AI concepts. While some of that might sound a bit daunting, it’s most important to stay open to experimenting with new concepts, enjoy the process, and have fun.
Professor Soja, thank you very much for allowing us to share with our readers these important answers on this fascinating topic, which will surely be of great value to everyone.
