pyunwrap: Learned InSAR Phase Unwrapping
A network that predicts the integer 2π cycle count at every pixel, so it can never contradict the measured phase. 99.1% of pixels exact on easy test scenes, 96.6% on moderate.
Overview
Radar satellites see through cloud and darkness, and interferometric SAR (InSAR) turns the phase difference between acquisitions into ground displacement. I build and run complete Sentinel-1 chains, from SLC search to SBAS velocity maps, for land subsidence, infrastructure stability and earthquake and volcanic deformation.
What you get
- Deformation velocity maps and displacement time series for your area of interest
- Reproducible processing pipelines you can rerun as new acquisitions arrive
- Quality reporting: coherence, network connectivity, baselines and where results should not be trusted
- Results delivered as Cloud Optimized GeoTIFF, GeoParquet time series or web-map tiles
- Feasibility assessments before you commit to a full processing campaign
Who this helps
Cities & utilities
Tracking urban subsidence from groundwater extraction, construction or reclamation.
Infrastructure owners
Screening roads, dams, embankments and mine sites for slow, hidden movement.
Hazard & research teams
Measuring deformation after earthquakes and eruptions with a reproducible chain.
Tropical-region programmes
Where persistent cloud cover makes optical monitoring unreliable.
How I approach it
-
01
Define the motion that matters
Expected rate, spatial extent and the decision it informs. These set the stack length, acquisition geometry and the choice between SBAS and persistent-scatterer methods.
-
02
Check feasibility before processing
Preflight checks on AOI coverage, network connectivity and perpendicular and temporal baselines catch a doomed stack before compute is spent.
-
03
Process with the physics corrected
Orbit-based coregistration with per-burst Enhanced Spectral Diversity, flat-earth and topographic phase removal, ERA5 tropospheric and IONEX ionospheric correction.
-
04
Unwrap with the right tool for each region
SNAPHU by default; pyunwrap's learned unwrapper where low coherence defeats it. Each region is routed to the method that benchmarks stronger, rather than claiming one method always wins.
-
05
Invert and select scatterers
SBAS inversion to velocities and time series; graph-neural-network persistent-scatterer selection (ps-gnn) for irregular urban targets.
-
06
Validate and state the limits
Compare with published references, GNSS or levelling where available, and report network fragmentation and coherence gaps alongside the map.
Evidence from my work
- 67Sentinel-1 SLC scenes processed for Mexico City
- 2,211 → 78Candidate pairs screened to valid interferograms
- −35.5 cm/yrRecovered, vs a published −39.1 cm/yr peak
- 3Deformation events run through the same chain
Mexico City subsidence, the 2021 Piton de la Fournaise eruption and the 2016 Amatrice earthquake were processed with the same pygeofetch InSAR chain.
Selected work
View case study
PS-GNN: Graph Neural Networks for Persistent Scatterer Selection
Persistent-scatterer selection as node classification on a spatial graph. F1 0.82 vs 0.69 for StaMPS-style selection on hard test scenes, and 0.97 recall on borderline scatterers.
Mexico City Land Subsidence
67 SLC scenes; 2,211 candidate pairs reduced to 78 valid interferograms; −35.5 to −2.9 cm/yr, consistent with the peer-reviewed −39.1 cm/yr peak.
Urban Subsidence Watcher
Sentinel-1 from the ASF Vertex API through an SBAS chain, with a Tiny U-Net on an 8-channel stack exported to ONNX for edge deployment.
What sets my work apart
- 1
I report the network, not just the map
Connectivity, coherence and fragmentation are part of every deliverable, so you know how far to trust each pixel.
- 2
Hybrid, not dogmatic
Classical and learned unwrapping are combined region by region, based on benchmarks rather than preference.
- 3
Built on my own open chain
The full Sentinel-1 workflow lives in pygeofetch, CI-tested and archived with a DOI, so results can be reproduced.