Slow-moving landslide velocity mapping with InSAR time series

Persistent Scatterer and Small Baseline Subset InSAR time series detect millimetre-to-centimetre annual surface creep on slow-moving landslides, distinguishing background deformation from the episodic acceleration that often precedes catastrophic failure.

Sensors

Sentinel-1 (C-band SAR, ESA/Copernicus): 5.6 cm wavelength, 5 x 20 m Interferometric Wide Swath resolution, 12-day exact repeat from a single satellite (6-day with both A and B units when both are operational). The standard dataset for PS-InSAR and SBAS processing. Consistent acquisition geometry and free archive back to 2014 make velocity time series practical. Dense vegetation can cause temporal decorrelation and limit coherence.
ALOS-2 PALSAR-2 (L-band SAR, JAXA): 23.6 cm wavelength penetrates forest canopy and maintains coherence where Sentinel-1 C-band loses it. 3 x 3 m to 10 x 10 m resolution depending on mode; 14-day revisit. The longer wavelength also increases sensitivity to large displacements without phase wrapping, but archive depth outside Japan is thinner than Sentinel-1.
COSMO-SkyMed (X-band SAR, ASI): 3.1 cm wavelength, down to 1 m resolution in Spotlight mode. High spatial resolution resolves small individual landslide bodies and structural features. Shorter wavelength means faster decorrelation over vegetation and a lower displacement ambiguity threshold (roughly 1.5 cm per half-wavelength), so it is best suited to slow, coherent rock slides rather than soil-dominated flows.
NISAR (L- and S-band SAR, NASA/ISRO, expected 2025): Dual-frequency design will provide L-band (23.6 cm) and S-band (9.3 cm) acquisitions with a 12-day global repeat. L-band will extend coherent coverage into tropical forest terrain where both C- and X-band struggle. The archive will begin at launch; no historical data exists yet.

Why slow is the dangerous kind

A catastrophic landslide that kills hundreds often begins as a slow creep that has been measurable for years. The 2014 Oso landslide in Washington State, the 2017 Xinmo rockslide in Sichuan, and numerous Alpine failures all show precursory acceleration in the months before collapse. The pattern is consistent enough that monitoring velocity, and especially changes in velocity, is now considered the most operationally useful early-warning signal available.

The difficulty is that 'slow' means rates of a few millimetres to a few centimetres per year. Ground-based instruments (GPS, extensometers, tiltmeters) can resolve that, but they cover single points and require installation on terrain that is, by definition, unstable. InSAR time-series methods can measure the same displacements across entire slopes, at spatial densities of thousands of measurement points per square kilometre, from orbit.

What the phase difference actually measures, and what it does not

InSAR works by comparing the phase of radar returns from the same target across two or more acquisitions separated in time. Any change in the distance between the satellite and the target shifts the phase. At C-band (Sentinel-1, 5.6 cm wavelength), one full phase cycle corresponds to a line-of-sight displacement of 2.8 cm. Processing stacks of tens or hundreds of acquisitions through PS-InSAR or SBAS algorithms reduces noise to the millimetre level and produces a velocity map with a time series at each coherent point.

The critical constraint: InSAR measures only the component of displacement along the satellite's line of sight. A slope moving horizontally toward the sensor looks identical, in phase terms, to a slope moving vertically downward at a different rate. To recover the actual three-dimensional displacement vector, analysts must combine ascending-pass and descending-pass geometries. Even then, the north-south component is poorly constrained by polar-orbiting SAR because the satellite's velocity vector is nearly north-south and sensitivity in that direction is very low. For most landslide geometries, east-west and vertical components can be decomposed reasonably well; north-south motion cannot.

Atmospheric artefacts are the other main source of error. Tropospheric water vapour delays the radar signal and mimics displacement signals at the centimetre scale. Correcting for this requires either external weather-model data (ERA5 is commonly used) or statistical approaches that exploit the spatial and temporal structure of atmospheric noise. Residual atmospheric error is typically 1 to 3 mm per year in well-corrected products; larger over mountainous terrain where orographic effects are strong.

PS-InSAR versus SBAS: choosing the right algorithm for the terrain

Persistent Scatterer InSAR identifies individual pixels that maintain stable radar reflectivity across all acquisitions in a stack, typically man-made structures, bare rock outcrops, and coarse debris. It delivers very precise displacement histories at those points but can be sparse over vegetated or soil-covered slopes. Small Baseline Subset (SBAS) processing uses pairs of acquisitions with short spatial and temporal baselines, then combines them to maximise spatial coverage at the cost of some precision per point. In practice, landslide studies often run both and merge the results.

For tropical or heavily forested terrain, neither C-band method works well. L-band ALOS-2 data, processed through the same PS or SBAS frameworks, maintains coherence through forest canopy because the longer wavelength is less sensitive to leaf movement between acquisitions. The trade-off is that ALOS-2 has a narrower tasking schedule outside its priority regions, and the archive is less dense than Sentinel-1's.

Reading the velocity map: creep, acceleration, and the threshold question

A typical slow-moving landslide velocity map shows a spatial gradient: fastest displacement near the main scarp and along lateral margins, slower movement in the accumulation zone at the toe. Rates below about 10 mm per year are often within atmospheric-noise margins and should be treated cautiously. Rates of 10 to 100 mm per year are reliably detectable with Sentinel-1 PS-InSAR given a stack of at least 20 to 30 acquisitions. Rates above roughly 100 mm per year begin to exceed the phase-unwrapping limit for 12-day C-band pairs and require either shorter-revisit commercial data or L-band to avoid aliasing.

The operationally important signal is not the absolute velocity but its change over time. A landslide creeping at 15 mm per year for three years that abruptly accelerates to 60 mm per year in a single season is a very different risk profile from one moving steadily at 60 mm per year. Time-series analysis makes this distinction visible. Thresholds for alert generation are site-specific and should be calibrated against historical failure records where they exist; there is no universal safe velocity.

Seasonal signals complicate interpretation. Many landslides accelerate during wet seasons or snowmelt periods and decelerate in summer. Distinguishing seasonal modulation from true long-term acceleration requires at least two to three full annual cycles of data, which is one reason the Sentinel-1 archive depth since 2014 is so operationally valuable.

Practical limits and what ground truth still has to do

InSAR time series cannot detect landslides that lack coherent scatterers, which excludes most active debris flows and many soil-dominated slides in wet tropical environments. It cannot measure displacements faster than about one quarter of the radar wavelength per revisit interval without phase aliasing (roughly 7 mm per 12 days for Sentinel-1 C-band). It produces no information about subsurface failure geometry, pore-water pressure, or the mechanical state of the basal shear zone. Those require boreholes and inclinometers.

What satellite InSAR does well is wide-area screening. A single Sentinel-1 Interferometric Wide Swath frame covers 250 x 250 km. Processing a multi-year stack over a mountain range can identify dozens or hundreds of previously unmapped slow-moving bodies simultaneously, prioritising where ground investigation is most urgently needed. That is the practical workflow: satellite screening first, ground instrumentation second, targeted at the highest-velocity or fastest-accelerating sites.

Satellize runs PS-InSAR and SBAS processing on open Sentinel-1 archives and can add commercial COSMO-SkyMed or ALOS-2 tasking on client licence for sites where open-data coherence is insufficient.

Typical figures

Spatial resolution (Sentinel-1 IW mode)	5 x 20 m ground range (multilooked to ~14 x 14 m in standard processing)
Spatial resolution (ALOS-2 PALSAR-2 Fine mode)	3 x 3 m to 10 x 10 m depending on beam
Revisit interval	12 days (Sentinel-1 single satellite); 6 days (dual satellite when both operational); 14 days (ALOS-2)
Minimum detectable velocity	~1 mm per year with large PS stacks in low-noise conditions; practical floor ~3 to 5 mm per year in mountainous terrain after atmospheric correction
Maximum measurable displacement per interval (C-band, 12-day)	~28 mm (one phase cycle at 5.6 cm wavelength); faster motion requires phase unwrapping or L-band
Radar frequency / wavelength	C-band 5.6 cm (Sentinel-1); L-band 23.6 cm (ALOS-2, NISAR); X-band 3.1 cm (COSMO-SkyMed)
Sentinel-1 archive depth	From 2014 (Europe priority); 2015 to 2017 onward for most global regions
Swath width (Sentinel-1 IW)	250 km
Minimum stack size for reliable PS-InSAR	20 to 30 acquisitions (approximately 1 to 1.5 years of Sentinel-1 data)
Deliverable formats	GeoTIFF velocity maps, CSV/Shapefile time-series point clouds, KMZ overlays, PDF technical report

Analytics Satellize can run

Mean annual velocity map	PS-InSAR or SBAS processing of multi-year Sentinel-1 stack; atmospheric correction via ERA5 tropospheric delay model	GeoTIFF raster of line-of-sight velocity in mm per year, with uncertainty estimate per pixel; Shapefile of PS/SBAS point cloud
3D displacement decomposition	Joint inversion of ascending and descending line-of-sight velocity fields to recover east-west and vertical components	Paired GeoTIFFs for east-west and vertical velocity; technical note quantifying north-south component uncertainty
Displacement time series at user-defined points	PS-InSAR or SBAS cumulative displacement extraction at GPS benchmark locations or user-specified coordinates	CSV time series with acquisition dates and cumulative displacement; plotted PDF for each monitoring point
Acceleration detection and change-point analysis	Statistical change-point detection (e.g. CUSUM or Bayesian methods) applied to velocity time series to flag episodic acceleration events	Alert shapefile with acceleration onset dates, magnitude of velocity change, and confidence level per landslide body
Seasonal deformation model	Harmonic regression on time series to separate annual and semi-annual periodic signals from secular velocity trend	Decomposed time-series plots; GeoTIFF of secular trend after seasonal removal
Wide-area slow-landslide inventory	Automated clustering of high-velocity PS/SBAS points into spatially coherent displacement bodies; cross-check against slope aspect and geology	Polygon shapefile of detected landslide bodies with velocity statistics, ranked by peak velocity and acceleration rate
L-band coherence gap fill	ALOS-2 PALSAR-2 SBAS processing for vegetated sites where Sentinel-1 coherence is insufficient; merged with C-band results where both are available	Supplementary velocity GeoTIFF for forested zones; merged multi-sensor velocity mosaic

Sources

Who does the work

We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.