May 13 2026 14:00 CEST
Live on Microsoft Teams.
On May 13 at 14:00 CEST, the ESA Φ-Lab Collaborative Innovation Network will host a new Φ-talk. Details are below.
Meet the speaker
Lorenzo Giuliano Papale is a researcher at the University of Rome Tor Vergata, where he obtained a PhD in Geoinformation. His research focuses on the integration of physical models and machine learning for Earth Observation, with particular expertise in SAR data analysis, soil moisture estimation, and physics-based AI approaches. He has collaborated with the European Space Agency (ESA) and was a visiting researcher at NASA’s Jet Propulsion Laboratory (JPL) in California, working on hybrid and physics-constrained learning methods. He is also involved in R&D activities with GEO-K, the first spin-off of the University of Rome Tor Vergata, and contributes to training and capacity building initiatives in the field.
Talk abstract
Artificial Intelligence has rapidly transformed the analysis of Earth Observation (EO) data, particularly in the context of Synthetic Aperture Radar (SAR), where complex signals are used to retrieve environmental parameters. However, purely data-driven approaches, especially Deep Learning models, often raise concerns related to interpretability and physical consistency. While highly flexible, these models can behave as black boxes and may fail to capture the underlying mechanisms governing the radar signal interactions.
In this context, this contribution explores the integration of AI with physical models as a way to balance flexibility and physical interpretability. On one side, physical models provide a rigorous description of electromagnetic scattering processes, but are often difficult to invert and require strong assumptions. On the other, AI models can learn complex mappings from data, but their behavior may be hard to interpret and difficult to generalize beyond the training domain.
The first part of the work investigates a hybrid framework based on the Tor Vergata electromagnetic model, used to generate physically consistent simulations. These are exploited to train and analyze soil moisture retrieval machine learning models, providing insights into feature relevance and model behavior in a physics-driven setting.
The second part focuses on research activities carried out at the NASA Jet Propulsion Laboratory – California Institute of Technology), where physics-informed machine learning approaches are used to explicitly embed physical constraints into neural networks. By guiding the learning process with radiative transfer principles, these methods enable models that are not only data-driven but also physically consistent and more interpretable.
Overall, this work highlights how combining AI and physical modeling can support a more transparent and physically grounded analysis of EO data.
Register here!