Post-failure landslide runout and deposit volume mapping

DEM differencing between pre- and post-event elevation models quantifies landslide deposit volume and runout geometry within days of failure, directly informing dam-break and downstream flood risk. This page covers the methodology, co-registration error budgets, and published results from the 2017 Maoxian and 2018 Baige events.

Sensors

TanDEM-X bistatic DEM: X-band interferometric pair acquired simultaneously from two spacecraft; global 12 m DEM product (TanDEM-X WorldDEM) with relative vertical accuracy better than 2 m (1 sigma) over moderate slopes. Tasked post-event acquisitions can be delivered within days. Penetrates cloud but surface scattering in dense vegetation introduces a canopy-height bias of 2–10 m.
Copernicus DEM GLO-30: TanDEM-X-derived global 30 m DEM, freely available, used as the pre-event baseline. Absolute vertical accuracy is approximately 4 m RMSE globally, degrading to 10–15 m in steep, shadowed terrain. Acquisition dates vary by tile (mostly 2011–2015), so it predates most recent events.
Pléiades 1A/1B stereo: Tri-stereo or along-track stereo at 0.5 m panchromatic resolution; photogrammetric DEMs achievable at 1–2 m post-spacing with 0.3–0.5 m vertical accuracy in open terrain after ground control. Revisit over a given point is typically 1–4 days. Cloud is the primary failure mode; a single overcast pass wastes the tasking window.
SPOT 6/7 stereo: 1.5 m panchromatic stereo pairs; DEM post-spacing typically 4–5 m with vertical accuracy of 1–3 m after co-registration. Wider swath (60 km) than Pléiades makes it better suited to long runout events extending tens of kilometres down valley. Archive extends to 2012.
Sentinel-1 SAR (change detection support): C-band SAR at 5–20 m resolution, 6-day repeat (12-day single satellite). Not used for DEM generation here, but coherence loss between pre- and post-event images delineates the disturbed area within hours of the first post-event pass, guiding where to task optical stereo. Free and open.

What a difference map actually tells you

Subtract a pre-event DEM from a post-event DEM and you get a map of elevation change: negative values where material was evacuated from the source scar, positive values where debris was deposited. Integrate the positive anomaly over area and you have deposit volume; integrate the negative anomaly and you have source volume. The ratio of the two, always greater than one because landslide material bulks as it disaggregates, is the bulking factor, typically 1.1 to 1.3 for rock avalanches.

For the 2018 Baige landslide on the Jinsha River, published studies using pre-event TanDEM-X WorldDEM differenced against post-event Pléiades DEMs estimated a deposit volume of approximately 25 to 30 million cubic metres, enough to impound a lake that eventually overtopped and triggered a downstream flood affecting communities more than 600 km away. For the 2017 Maoxian (Xinmo) rock avalanche in Sichuan, DEM differencing placed the deposit volume at roughly 8 to 10 million cubic metres, with a runout of about 2.8 km. These are not round numbers plucked from press releases; they come from peer-reviewed analyses in journals such as Remote Sensing and Landslides, and different studies disagree at the 10–15 percent level, which is itself a useful calibration of method uncertainty.

Co-registration: the error that swamps small signals

DEM differencing is simple in concept and treacherous in practice. A horizontal misalignment of one pixel between pre- and post-event DEMs produces a spurious elevation-change signal proportional to the local slope gradient. On a 30-degree slope with a 30 m DEM, a single-pixel (30 m) horizontal shift creates apparent vertical errors of roughly 17 m, which is larger than the deposit thickness of many small landslides.

The standard correction, described by Nuth and Kääb (2011) and now widely adopted, regresses elevation difference against terrain aspect and slope angle over stable, unvegetated terrain outside the landslide boundary. Iterating this regression to convergence typically reduces horizontal misalignment to sub-pixel levels and brings residual vertical bias below 0.5 m on flat stable ground. Even after co-registration, steep shadowed slopes and areas of dense forest remain problematic; honest volume estimates should carry an uncertainty band of at least 10–20 percent for well-constrained cases and up to 40 percent where terrain is extreme or stereo coverage is partial.

Vegetation is a separate problem. A pre-event DEM derived from X-band radar sees the canopy surface, not the ground. A post-event optical stereo DEM, generated after the forest has been stripped by the landslide, sees bare rock. The difference map will show apparent material loss in forested source zones that partly reflects canopy removal rather than rock displacement. Correcting for this requires either a lidar-derived canopy height model or a forest-height estimate from GEDI or ICESat-2, neither of which is always available at the right moment.

Choosing the right DEM pair for the event

The pre-event baseline is usually fixed by what exists. Copernicus GLO-30 is the default for global coverage, free, and consistent. Where the event is recent enough, a pre-event TanDEM-X WorldDEM tile at 12 m is preferable because it matches the X-band geometry of a post-event SAR acquisition. National mapping agencies sometimes hold lidar surveys at 1 m resolution, which, when available, transform the analysis entirely.

Post-event DEM choice depends on cloud, budget and required accuracy. Pléiades tri-stereo at 0.5 m is the gold standard for deposit mapping but requires a cloud-free window and tasking lead time. SPOT 6/7 stereo at 1.5 m covers larger areas per pass, which matters when the runout extends many kilometres. For events in persistently cloudy regions (the eastern Tibetan Plateau, the Colombian Andes, parts of Papua New Guinea), waiting for optical stereo can take weeks. In those cases, a post-event TanDEM-X bistatic acquisition, taskable within 24–48 hours, may be the only timely option despite its coarser resolution.

Mixing sensor types across the pre/post pair introduces systematic biases that co-registration alone cannot fully remove. Documenting the sensor provenance of each DEM, its acquisition date, and the co-registration residual is not bureaucratic housekeeping; it is the information a downstream engineer needs to decide whether the volume estimate is good enough to size a spillway or evacuate a town.

From volume to downstream hazard

A deposit volume number is not itself a hazard assessment. Its value lies in what it feeds: hydraulic dam-break models, debris-flow runout models such as DAN3D or r.avaflow, and sediment-budget calculations for river reaches downstream. The deposit geometry, specifically the width and height of the natural dam, controls the time to overtopping and the peak discharge of any outburst flood.

At Baige in 2018, the dam crest was estimated at roughly 61 m above the pre-landslide river bed from DEM differencing, a figure that fed directly into flood routing models used by Chinese authorities to manage controlled releases and pre-emptive evacuations downstream. The satellite-derived geometry was available within 48 hours of the second (larger) failure, before any ground survey team could safely access the site.

Runout length and lateral spread, both readable from the deposit outline in the difference map, also constrain empirical mobility indices such as the Fahrböschung angle (the ratio of vertical drop to horizontal runout). For large rock avalanches, published Fahrböschung angles cluster between 0.1 and 0.3; an event plotting below 0.1 suggests unusual mobility, possibly from entrainment of saturated substrate or air lubrication, and warrants particular attention to secondary runout risk.

Practical limits and what they mean for decision-makers

No method discussed here works through cloud. Optical stereo fails entirely; even SAR-derived DEMs degrade in heavy rain because of atmospheric path-delay noise. For events in tropical or monsoon climates, the first usable post-event optical window may arrive weeks after failure, by which time a natural dam may already have overtopped.

Volume estimates below roughly 100,000 cubic metres are unreliable with 30 m pre-event DEMs; the signal is too close to the co-registration noise floor. Pléiades-derived post-event DEMs push this floor down to perhaps 10,000–50,000 cubic metres, depending on terrain roughness. Small failures on active slopes, the kind that precede catastrophic collapse, are better monitored by InSAR precursor methods covered in the sibling page on pre-failure landslide precursor motion detection.

Satellize structures post-event DEM differencing workflows as a rapid-response analytic, with a standard output package including co-registered DEM pair, elevation-change raster, deposit outline polygon, and a volume estimate with documented uncertainty.

Typical figures

Pre-event DEM resolution (Copernicus GLO-30)	30 m post-spacing; relative vertical accuracy ~4 m RMSE globally, degrading in steep terrain
Pre-event DEM resolution (TanDEM-X WorldDEM)	12 m post-spacing; relative vertical accuracy <2 m (1 sigma) over moderate slopes
Post-event DEM resolution (Pléiades stereo)	1–2 m post-spacing achievable; vertical accuracy 0.3–0.5 m in open terrain with ground control
Post-event DEM resolution (SPOT 6/7 stereo)	4–5 m post-spacing; vertical accuracy 1–3 m after co-registration
Tasking latency (Pléiades / SPOT)	First acquisition typically within 24–72 hours of request, subject to cloud and orbital geometry
Minimum detectable volume (30 m baseline DEM)	Approximately 100,000 m³; smaller events approach co-registration noise floor
Minimum detectable volume (Pléiades post-event DEM)	Approximately 10,000–50,000 m³ depending on terrain roughness and co-registration quality
Volume uncertainty (well-constrained case)	10–20%; up to 40% in extreme terrain or partial stereo coverage
Copernicus GLO-30 archive depth	Acquisition epoch mostly 2011–2015; freely available via Copernicus Data Space
Sentinel-1 change detection support	5–20 m SAR; 6-day repeat; coherence loss map available within hours of first post-event pass

Analytics Satellize can run

Co-registered DEM pair	Nuth and Kääb (2011) iterative aspect/slope co-registration over stable terrain	Two aligned GeoTIFF DEMs with documented residual bias, delivered as GIS-ready rasters
Elevation change raster (DoD)	Pixel-wise subtraction of co-registered pre- and post-event DEMs with uncertainty propagation	Signed elevation-difference GeoTIFF with per-pixel uncertainty layer
Deposit and source volume estimate	Integration of positive (deposit) and negative (source) DoD anomalies within manually or automatically delineated landslide boundary	Tabular volume report with bulking factor, uncertainty range, and boundary polygon (GeoJSON/Shapefile)
Runout geometry profile	Centreline extraction from deposit outline; Fahrböschung angle calculation from DEM-derived vertical drop and horizontal runout	Longitudinal profile plot and mobility index table in PDF and CSV
Natural dam geometry characterisation	Dam crest elevation, width, and freeboard derived from post-event DEM relative to pre-event river thalweg	Dam geometry summary table for input to hydraulic dam-break models (HEC-RAS, BREACH compatible format)
Disturbed-area delineation from SAR coherence	Sentinel-1 interferometric coherence differencing between pre- and post-event image pairs	Binary disturbed-area mask (GeoTIFF) available within 6–12 hours of first post-event SAR pass, used to prioritise optical tasking
Vegetation bias correction	Canopy height subtraction using GEDI Level 2 footprint data or ICESat-2 ATL08 land-surface product over forested source zones	Corrected DoD raster with canopy-bias adjustment note in metadata

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.