Enhancing Satellite Deformation Measurements using Deep Learning
PI: Christelle Wauthier, Department of Geosciences, EMS & Associate Director, ICDS
Other team members: Romit Maulik, IST, and Steve Greybush, Department of Meteorology and Atmospheric Science, EMS
The goal of this research at the nexus of deep learning and geosciences is to identify with confidence and accuracy surface displacements from satellite geodetic data (Interferometric Synthetic Aperture Radar – InSAR – datasets) that can be cause of concern in active volcanic areas. InSAR is indeed commonly used to study areas affected by natural hazards such as near active volcanoes due to its capability to provide sub-centimetric deformation measurements under favorable conditions without the need for ground-based instruments. Ground deformation is one of the most important measurements to forecast volcanic unrest that can impact local and global populations, economies, as well as disrupt significantly air traffic. A significant limitation of InSAR processing is the use of a simplistic or incomplete atmospheric correction that can lead to misinterpretations or difficulties to identify deformation signals of concern, especially if they are subtle or slow (~mm to cm/year). We thus propose to develop a deep learning approach which considers realistic atmospheric conditions and produces high quality noise-free 2D surface displacements over time. The combination of AI deep learning approaches with multi-temporal InSAR (InSAR time series of surface displacements) is very promising for monitoring volcanic activity by detecting minor ground deformations that might occur before eruptions, which could enhance early-warning systems. However, this integration is still in its infancy across multiple fields including volcanic unrest. There is thus a crucial need for the proposed deep learning methodology that considers actual atmospheric characteristics and can be applied to a broad suite of areas affected by various atmospheric and deformation conditions.
List of specific areas of computational and/or data science expertise or skills: AI, deep learning, CNN.
List of specific objectives for work supported by this call: generating high quality training datasets and test a CNN workflow that can be applied to various volcanic areas in different parts of the World affected by various atmospheric conditions. The method should be successful at isolating the InSAR deformation signals that are cause of concern from atmospheric and other noise signals. The results will then support grant proposals to NSF, NASA, and other relevant opportunities (for example to the NSF CAIG call in February 2026). We will submit at the minimum three papers showcasing the CNN using realistic atmospheric models in the training stage for the three following active and high-risk volcanic areas: Big Island, HI, Mount Baker, Washington, and Medicine Lake, California. Connection of the project to ICDS’s mission: we will develop and apply AI methods to natural hazards processes and forecasting.