Milillo's Lab

Milillo's Lab

SAR methods and machine learning

Getting more out of the same radar data than the acquisition was designed for.

The problem

A radar acquisition contains more information than the standard processing chain extracts. Sub-aperture phase carries vibration. Optical and radar data carry complementary information that neither alone resolves. And the volume of imagery now available exceeds what can be interpreted by hand.

Synthetic aperture (one pass) t₁t₂ t₃t₄ sub-apertures: four looks at the same target, seconds apart Vibrating target phase history → velocity, to 0.01 m/s
Splitting one acquisition into sub-apertures gives several looks at the same target seconds apart, so the phase history carries motion the full image averages away.

How we measure it

We work on the extraction side rather than the acquisition side: micro-motion and sub-aperture phase analysis for vibration, self-supervised optical-SAR fusion for classification, and vision-language models for damage assessment that can explain their own output.

What we have found

Vibration from a single pass

Micro-motion analysis recovers target velocities as low as 0.01 m/s from one high-resolution X-band image, validated in Trento and Glasgow.

Interpretable damage

Vision-language models applied to multi-hazard damage assessment with commercial satellite data.

Automated grounding lines

Deep learning and phase-gradient methods for grounding line delineation, replacing manual digitisation.

Papers on this theme 38

Showing the 12 most relevant of 38. See all 92 publications.

← All research themes