3D Target Characterisation with the Physics Informed SAR Machine Learning

The project exploits the new opportunity offered by recent high-resolution SAR systems, enabled by parabolic reflectors and electronically steered phased arrays, to extract not only the surface map of the scene but also information in the third dimension across a wide range of spectral bands, from P/L- to X-band. However, volumetric investigation of objects remains challenging due to the limited number of look angles. The project addresses this shortcoming by using prior knowledge of the physics of EM wave propagation to regularise the problem.  

Access to large spaceborne SAR datasets allows us to learn a fast surrogate model of 3D wave–target interaction, based on neural operators, incorporate it into SAR imaging algorithms, and exploit all the information recorded by the instrument. The resulting 3D target information will then be fed into a machine intelligence system for localisation, characterisation, and feature extraction. While these downstream tasks often need ground-truth information, the project aims to use self-supervised learning methods and incorporate PINNs to reduce the need for labelled data as much as possible. This approach has been shown to be successful in SAR change detection using pretrained models and polarimetric SAR processing. The project builds on these results and uses manually labelled data mainly for validation.