Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at the equator with both satellites. SWIR bands (1610 nm and 2190 nm) help distinguish bare soil from shadow. Free and openly archived from 2015, giving pre-event baselines for almost any location on Earth.
- Planet SuperDove: 3 m resolution, 8 spectral bands including red-edge and two NIR channels. Daily revisit over most landmasses makes it the practical choice for rapid post-event optical coverage when cloud permits. No public archive; commercial tasking or standing licence required.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral. Used for individual scar characterisation: measuring crown width, estimating deposit volume from stereo-derived DSMs, and confirming ambiguous detections flagged by coarser sensors. Tasked commercially; latency from order to delivery is typically same-day to 48 hours.
- ALOS-2 PALSAR-2: L-band SAR (1.27 GHz) operated by JAXA. L-band penetrates moderate forest canopy, so coherence loss between pre- and post-event acquisitions reveals mass movement even under cloud or partial vegetation cover. Stripmap mode delivers 3 m resolution; ScanSAR covers wider swaths at 10–100 m. Revisit is 14 days in standard mode, shorter when JAXA activates emergency observation protocols.
What a fresh scar looks like to a sensor
A landslide strips vegetation and exposes bare mineral soil or rock. That transition is spectrally loud. Vegetated surfaces have high reflectance in the near-infrared and low reflectance in the red, producing a high NDVI. Bare soil inverts that relationship: NIR drops, red rises, NDVI falls sharply. In shortwave infrared bands, freshly disturbed soil also shows elevated reflectance compared with intact forest, partly because the surface moisture regime changes and partly because the soil composition differs from the organic-rich topsoil it replaced.
The practical consequence is that a simple differenced NDVI image, computed between a cloud-free pre-event acquisition and the first usable post-event image, lights up landslide scars quite reliably. The challenge is separating them from agricultural clearance, fire scars and cloud shadows, all of which produce similar NDVI drops. Adding SWIR bands and contextual slope data from a digital elevation model removes most false positives. This is standard practice in the published literature and the basis of operational rapid-mapping workflows used by agencies such as Copernicus EMS.
SAR coherence: seeing through the cloud that caused the problem
The same rainfall that triggers a landslide typically blankets the affected area in cloud for days. Optical sensors are useless until the sky clears. SAR sensors transmit their own microwave signal and record the backscatter regardless of cloud or darkness, which is why SAR coherence analysis is often the first usable data product after a major event.
Coherence measures how similar the radar phase is between two acquisitions over the same patch of ground. Stable surfaces, buildings, bare rock, maintain high coherence. Any surface that has physically moved or changed its dielectric properties between acquisitions loses coherence. A landslide does both: it moves material and replaces vegetation with wet debris. The resulting coherence-loss map shows affected areas as dark patches against a brighter stable background. At L-band, PALSAR-2 penetrates forest canopy well enough to detect movement that C-band Sentinel-1 might miss under dense vegetation, though Sentinel-1's 6-day revisit and free data access make it the default first look.
Coherence loss is not specific to landslides. Flooding, agricultural tillage and strong wind events all reduce coherence. Slope context and the spatial pattern of the loss, elongated downslope, fan-shaped at the toe, help analysts distinguish landslides from other causes. Honest caveat: in areas of very high natural coherence variability, such as tropical forest with frequent wind or rain disturbance, the method produces more ambiguity and requires optical confirmation.
Minimum mappable area and the resolution floor
Sensor resolution sets a hard lower bound on what can be detected. A single Sentinel-2 pixel covers 100 square metres. Mapping a scar reliably requires at least a handful of contiguous affected pixels, putting the practical minimum detectable scar at roughly 1,000 to 2,500 square metres with Sentinel-2. Planet SuperDove at 3 m resolution lowers that floor to around 100 to 250 square metres. Pléiades Neo at 30 cm can resolve individual trees displaced by a shallow translational slide, but it covers a much smaller footprint per acquisition and costs correspondingly more.
This matters for event inventories. A rainfall event in steep terrain may trigger hundreds of small shallow slides alongside a handful of large deep-seated failures. Sentinel-2 captures the large ones well and misses the small ones entirely. A complete inventory, the kind used to calibrate susceptibility models, requires either very high-resolution coverage or a statistically defensible acknowledgement that the inventory is truncated below a given size threshold. Omitting that caveat produces susceptibility models that underestimate hazard at the small-scar end of the distribution.
Building an event inventory that is actually useful
A post-event inventory is a polygon dataset where each polygon represents one mapped landslide, attributed with location, estimated area, type (fall, slide, flow), confidence level and the detection method used. That sounds straightforward. In practice, debris from multiple slides often merges at valley bottoms, individual scarps are partially obscured by the debris they generated, and cloud gaps mean different parts of the affected area are imaged on different dates at different stages of the event sequence.
Good inventory practice documents those limitations explicitly. Each polygon should carry a date-of-detection field and a confidence score. Polygons mapped from 10 m Sentinel-2 data should not be mixed without qualification with polygons mapped from 30 cm Pléiades data, because the size distributions are incomparable. The inventory should also distinguish the depletion zone (the scar, where material was removed) from the accumulation zone (the deposit, where it arrived), because susceptibility models and runout models need them separately.
Inventories built this way feed directly into frequency-area analysis, which is the standard statistical method for characterising a landslide population and estimating the probability of future events of a given size. The GLAD forest-change dataset at the University of Maryland provides a useful complementary layer for distinguishing landslide-driven vegetation loss from anthropogenic clearing in forested terrain.
Latency, archive depth and what they mean for response
In the first 24 to 72 hours after a triggering event, the most valuable satellite product is often not a finished inventory but a rapid damage proxy: a coarse coherence-loss or NDVI-difference map that tells emergency managers where to look. JAXA activates ALOS-2 emergency observations under the Sentinel Asia framework. The International Charter on Space and Major Disasters can task multiple commercial and government sensors within hours of a request from an authorised user. Copernicus EMS produces rapid-mapping products, typically within 24 to 48 hours of activation, using Sentinel and commercial data.
For susceptibility modelling, the archive matters as much as the current event. Sentinel-2 imagery goes back to 2015. Landsat provides usable 30 m optical data from the early 1970s. A multi-decadal optical archive allows analysts to identify scars from historical events that were never field-mapped, substantially enriching the training dataset for any susceptibility model. Satellize runs change-detection pipelines on open constellations and can add commercial tasking on client licence, similar to the approach used in its Tonga crop-estimation work, where archive depth and current imagery are combined to extract signals that neither alone would reveal.
Honest limits of the method
Cloud is the dominant operational constraint for optical methods. A prolonged wet season can mean weeks without a usable optical acquisition over the affected area. SAR fills part of that gap but introduces its own ambiguities. Neither method reliably detects landslides beneath intact dense forest canopy where the surface expression is minimal.
Stereo-derived volume estimates from Pléiades or WorldView carry uncertainties of 10 to 30 percent depending on DEM quality and the complexity of the deposit geometry. Coherence-loss mapping cannot distinguish a landslide from a flood in a narrow valley without additional context. And any inventory is only as good as the image dates available: a slide that occurred, stabilised and was partially revegetated before the first post-event image may not be detected at all. These are not reasons to avoid satellite-based mapping. They are reasons to report it honestly and to combine it with field validation wherever access permits.
Typical figures
| Optical spatial resolution (survey-grade) | 10 m (Sentinel-2), 3 m (Planet SuperDove) |
| Optical spatial resolution (scar characterisation) | 30 cm pan / 1.2 m multispectral (Pléiades Neo) |
| SAR spatial resolution | 3 m stripmap (ALOS-2 PALSAR-2); 5 × 20 m IW mode (Sentinel-1) |
| Revisit (optical, open) | 5 days at equator (Sentinel-2, both satellites combined) |
| Revisit (optical, commercial) | Daily (Planet SuperDove); tasked same-day to 48 h (Pléiades Neo) |
| Revisit (SAR) | 14 days standard (ALOS-2); 6 days (Sentinel-1); shorter under emergency activation |
| Minimum mappable scar area | ~1,000–2,500 m² (Sentinel-2); ~100–250 m² (SuperDove); smaller with Pléiades Neo |
| Key spectral bands for detection | Red, NIR (NDVI differencing); SWIR 1610 nm and 2190 nm (bare soil discrimination) |
| SAR frequency | L-band 1.27 GHz (ALOS-2); C-band 5.4 GHz (Sentinel-1) |
| Archive depth (open optical) | Sentinel-2 from 2015; Landsat usable from early 1970s at 30 m |
Analytics Satellize can run
| Rapid damage proxy map | Differenced NDVI or SAR coherence loss computed between nearest pre-event and first post-event acquisition | GeoTIFF and vector overlay, delivered within 24–48 h of usable imagery becoming available |
| Landslide event inventory | Semi-automated optical change detection with manual QA; polygons attributed by type, area, confidence and detection method | GeoPackage or Shapefile with full attribute table; PDF summary report with caveats on size truncation and cloud gaps |
| Depletion / accumulation zone separation | Spectral segmentation combined with slope-aspect analysis from a DEM; depletion zones show higher SWIR reflectance, accumulation zones show heterogeneous texture | Separate polygon layers for scar and deposit, compatible with standard susceptibility-modelling inputs |
| Deposit volume estimate | Stereo DSM differencing from Pléiades Neo or equivalent VHR pair; pre-event DEM subtracted from post-event DSM over accumulation zone | Volume estimate in cubic metres with stated uncertainty range (typically ±10–30%); raster difference map |
| Historical scar inventory from archive | Multi-temporal NDVI stack analysis over Sentinel-2 and Landsat archive to identify past vegetation-loss events on steep slopes | Polygon dataset of historical scars with estimated event dates, for use as training data in susceptibility modelling |
| Coherence-loss map (cloud-penetrating) | Interferometric coherence computed from ALOS-2 or Sentinel-1 pre/post-event SLC pairs; thresholded and filtered by slope mask | Raster coherence-difference layer and vectorised affected-area polygons, with notes on ambiguous detections |
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.