The advances in geodetic theory made an increased emphasis on mathematical methods necessary (Heiskanen and Moritz, 1956). For several decades a rigorous mathematical framework has been developed - Gauss/Markov BLUE, MLE, DIA - in the context of parameter estimation and statistical testing, thus paving the way for a better understanding of Earth's shape, orientation in space, and gravity field. With the introduction of machine learning, the focus has been shifted from a model-driven to a data-driven approach, also thanks to the large amount of data made available through several different terrestrial and space techniques (e.g. GNSS, InSAR, VLBI, SLR, Altimetry, Gravimetry, etc.).
In this short course we first provide a broad overview of geodetic theory, addressing different mathematical problems and well-established solutions adopted in Geodesy. Therefore, we highlight gaps in the current theoretical framework and introduce machine/deep learning paradigms as potential alternative to classical solutions. In this way, we further discuss key relationships between statistical learning and ML/DL methods, in particular focusing on fundamental issues in the adoption of AI techniques as "black box" solutions. Hence, we provide a clear understanding of the major pitfalls, especially concerning the quantification of uncertainty and confidence levels for ML/DL solutions.
Ultimately, we highlight the key role in science of 'explainability' and 'reproducibility', both often overlooked when adopting AI techniques in Geodesy. Target audience is Geodesy and Earth-science practitioners who deploy or evaluate ML in their research works. The suggested format is 60 minutes (e.g. lunch slot) with 30′ for a mini lecture on theoretical fundamentals, 20′ live demo with relevant geodetic examples, and 10′ for Q&A.
Prerequisites: basic linear algebra; no prior ML/DL knowledge is required.