The Role of AI in Modern Geodesy: Insights from GGOS Focus Area AI4G
Soja, Benedikt; Kaselimi, Maria; Asgarimehr, Milad; Modiri, Sadegh; Sharifi, Mohammad Ali; Belda, Santiago; Liu, Lei; Omidalizarandi, Mohammad; Śliwińska-Bronowicz, Justyna
- Publisher
- Zenodo
- Language
- en
Abstract
Artificial Intelligence (AI) is revolutionizing Earth observation by enabling large-scale data analysis, accurately detecting patterns in the data, and improving predictions. Similarly, AI and machine learning (ML) have led to significant advancements in geodesy. Studies have successfully used AI/ML to enhance geodetic products, such as refining gravity field model resolution, improving Earth orientation parameter predictions, and detecting anomalies in station coordinate time series. With the growing adoption of deep learning, larger datasets are increasingly used for more accurate, high-resolution modeling and prediction. The applications of such approaches are diverse, ranging from climate monitoring to natural hazard detection. The AI for Geodesy (AI4G) Focus Area within the Global Geodetic Observing System (GGOS) has been established to promote and facilitate the use of AI for improving geodetic products. Thorough quality assessment and uncertainty quantification of AI results are key aspects to ensure their acceptance within the geodetic community. Through specialized study groups, AI4G tackles specific challenges in the fields of GNSS remote sensing, gravity field and mass change, Earth orientation parameter prediction, and geodetic deformation monitoring. In this contribution, we highlight recent developments within the AI4G initiative, including groundbreaking research by its members. We will discuss emerging topics such as explainability, foundation models, and physics-informed learning. Finally, we will provide an outlook on the future activities of AI4G and explore the role of AI in shaping modern geodesy.
Full text
Partner/Sponsor: • A multimodal self-supervised model is proposed for GNSS-R Earth surface monitoring, Generalist Earth surface Monitor (GEM) •Operational ML pipelines for GNSS-R soil moisture competitive with SMAP for near-real-time use • Physics-informed DL fuses geometry and environmental features to improve robustness • AI calibration of multi-constellation winds •Unsupervised ML tracks inland water bodies •Edge ML demos for real-time land–water classification The Role of AI in Modern Geodesy: Insights from GGOS Focus Area AI4G Benedikt Soja(1), Maria Kaselimi(2), Milad Asgarimehr(3), Sadegh Modiri(4), Mohammad Ali Sharifi(5), Santiago Belda(6), Lei Liu(7), Mohammad Omidalizarandi(8), Justyna Śliwińska-Bronowicz(9) (1) Institute of Geodesy and Photogrammetry, ETH Zurich (2) National Technical University of Athens (NTUA) (3) GFZ Helmholtz Centre for Geosciences Summary GGOS Focus Area AI for Geodesy (AI4G) •Higher accuracy and resolutions (spatial/temporal) •Improved real-time and prediction quality •Trustworthy & interpretable Enabled by new developments in geodetic data: •Huge increase in data volume from GNSS, InSAR, etc. •Useful auxiliary data: weather, environmental models, etc. 4 Joint Study Groups •100+ Members •50+ institutions •10+ countries Recent activities: •Conference session organization: EGU25, AGU25, IAG25 Symposium J02 with 3 oral sessions and 38 abstracts •COST Action proposal to be submitted in November 2025 Above:GNSS-R generalist Earth surface monitor (GEM) Right: Predicted environmental variables by GEM AI for Gravity Field and Mass Change Review paper on uncertainties in deep learning •Focus on GRACE/GRACE-FO Gou et al.: Uncertainties of Satellite-based Essential Climate Variables from Deep Learning. https://doi.org/10.48550/arXiv.2412.17506 AI for GNSS Remote Sensing (4) Federal Agency for Cartography and Geodesy (BKG) (5) University of Tehran (6) University of Alicante (7) University of Colorado at Boulder (8) Leibniz University Hannover (9) Space Research Centre of Polish Academy of Sciences Objective: develop & evaluate improved geodetic products based on AI and machine learning Webinar series EOP I&I (Innovation & Insight) New EOP Prediction Comparison Campaign for ML with standardized rules for input data: EOP PML AI for Earth Orientation Parameter Prediction •Machine learning based-classification of InSAR time series •Deep learning-based approaches to fill gaps in InSAR time series •Crack detection in civil structures and historical buildings using deep learning •Supervised and unsupervised DL for change point detection in InSAR time series •DL in standard least-squares theory of linear models AI for Geodetic Deformation Monitoring AI4G member countries & cities Nr. Speaker & Affiliation Title / Focus Link (GGOS‘s Youtube) 1Junyang Gou, IGP, ETH Zürich, Switzerland Recurrent neural networks and their applications for EOP prediction — sequence models (RNN/LSTM/GRU) for shortto midterm EOP forecasts 2 Mostafa Kiani Shahvandi, IGP ETH Zürich, Switzerland (Currently at University of Vienna, Austria) Explaining the Causes of Polar Motion by Physics-Informed Neural Networks — embedding physical constraints in NN training to improve interpretability 3 Justyna ŚliwińskaBronowicz, CBK PAN, Poland Advancing EOP Prediction Using ML – Insights from the new EOP PML sub-campaign — scope, rules, and first experiences 4Yuanwei Wu, NTSC, Chinese Academy of Sciences, China Report on our activities of the 1st CEOPPCC — organization and progress of the China EOP Prediction Comparison Campaign •AI drives rapid developments across several geodetic fields •Applications of AI are continuously expanding •Interested in AI4G? Contact us to become involved! PMx PMy UT1– UTC LOD dX dY Total Number of predictions 62 62 62 24 221 221 652 Number of IDs 3 3 3 1 16 16 17 Below: Number of predictions submitted to the EOP PML (42 weeks) Number of IDs registered in EOP PML: 21 Number of active IDs in EOP PML (at least one submission): 17 Above: Preliminary results of 10-day-ahead MAE for dX predictions, evaluated against the IERS 20 C04 solution. The results are very promising and indicate a significant improvement in CPO prediction accuracy during EOP PML sub-campaign. Closure of water balance equation in terms of equivalent water height (EWH).Left:Uncertainties from Bayesian Neural Networks.Right:Prediction errors (RMSE) Left: Synthetic data generation from real InSAR time series using Generative Adversarial Networks (GANs) (Dhanasekaran 2025) Above: A PS point and its neighboring PS time series within a 15-meter radius Middle: Detected CPs in InSAR time series (Shahryarinia et al. 2025)
