Earthquake-triggered landslide inventory from change detection
Co-seismic landslides can outnumber building collapses as a cause of earthquake fatalities, yet they are invisible to ground teams for days. Satellite change detection cuts that gap to hours, producing georeferenced inventories that guide search-and-rescue routing and infrastructure triage.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at the equator (2-3 days with both satellites). The contrast between bare soil reflectance and vegetated surface in the red-edge and NIR bands (B8, B8A) drives most pixel-based change indices. Free and open; archive from 2015. Cloud cover is the dominant operational constraint.
- PlanetScope: 3 m resolution, near-daily revisit from a constellation of roughly 200 smallsats. The finer spatial resolution resolves individual landslide scarps and deposits that Sentinel-2 conflates into mixed pixels. Useful for inventorying smaller failures (down to roughly 900 m² at 3 m) in the days immediately after an event, subject to commercial tasking and cloud.
- SPOT-7: 1.5 m panchromatic, 6 m multispectral. On-demand tasking allows rapid acquisition over a defined epicentral zone. Particularly valuable for distinguishing landslide boundaries from adjacent disturbed surfaces at scales relevant to road-network assessment. Archive depth and tasking latency depend on operator scheduling.
- Sentinel-1 SAR (C-band, IW mode): 20 m ground range resolution, 6-day repeat (12-day with a single satellite). SAR coherence between pre- and post-event image pairs drops sharply over surfaces that have moved or been stripped of vegetation. Coherence loss is cloud-independent, making it the primary fallback when optical acquisitions are blocked for days after a seismic event. Backscatter intensity change is a secondary indicator.
What a bare scar looks like to a sensor
When a landslide strips vegetation from a hillslope, it exposes mineral soil and rock. That surface has a markedly different spectral signature from the forest or scrub it replaced: higher reflectance in the shortwave infrared (around 1.6 µm and 2.2 µm), lower near-infrared reflectance, and a collapsed normalised difference vegetation index (NDVI). The magnitude of the spectral shift depends on the density of the pre-event canopy and the soil type, but in humid tropical and subtropical terrain, the contrast is large enough to be detectable at Sentinel-2's 10 m pixel size for scars above roughly 1,000 m².
SAR tells a related but distinct story. C-band radar coherence measures how similar the phase of backscattered energy is between two passes. A hillslope that has been stripped and displaced becomes incoherent: the coherence value drops toward zero. This signal is independent of solar illumination and cloud cover, which is why SAR coherence maps are often the first usable product after a major earthquake in a monsoon-affected region. The limitation is that coherence also drops over dense vegetation agitated by wind, over flooded areas, and over any surface that has changed between the two acquisition dates for reasons unrelated to landsliding.
Three detection approaches and where each breaks down
Pixel-based change detection computes a difference or ratio of spectral indices (NDVI, NDWI, or a principal component) between a pre-event composite and the first cloud-free post-event image. It is fast and reproducible, but it cannot distinguish a landslide scar from a freshly harvested field, a fire scar, or a construction cut. In agricultural landscapes near the epicentre, false-positive rates can be high enough to require manual filtering before the inventory is operationally useful.
Object-based image analysis (OBIA) segments the post-event image into spectrally and geometrically coherent regions, then classifies each region using shape, texture, and spectral attributes. Elongated, high-slope-angle objects with bare-soil spectral properties score as landslides. This reduces salt-and-pepper noise and improves the separation from rectangular agricultural clearings, but it is sensitive to segmentation parameters and requires a digital elevation model of adequate quality.
Deep learning classifiers trained on labelled landslide polygon datasets (such as the Cooperative Open Online Landslide Repository, COOLR, maintained by NASA) can generalise across terrain types better than threshold-based methods, but they require sufficient training examples from geologically similar settings. A model trained on the Himalayas may underperform in the Andes if soil colour and vegetation structure differ substantially. All three approaches share one hard constraint: without a cloud-free post-event optical acquisition, they cannot run. SAR coherence is the only cloud-independent fallback, but it produces a proxy for surface change rather than a confirmed landslide boundary.
Latency: the gap between shaking and knowing
For a major earthquake in a mountainous region, the realistic timeline runs as follows. Sentinel-1 acquires on a fixed 6-day repeat, so a coherence pair may be available within 0 to 6 days depending on where in the orbit cycle the event falls. Sentinel-2 and PlanetScope optical acquisitions depend on cloud clearance, which in tropical highlands can mean waiting 5 to 15 days or longer during the monsoon season.
The International Charter on Space and Major Disasters can task commercial optical satellites within 24 to 48 hours of activation, but cloud is not negotiable. Practical experience from events such as the 2015 Gorkha earthquake in Nepal showed that the first usable Sentinel-2 optical inventory took several days to weeks to compile because of persistent cloud, while SAR-based coherence maps were available within one repeat cycle. Planners should treat the inventory as a living product: an initial SAR-based probability map updated by optical confirmation as cloud clears.
Minimum detectable size and the completeness problem
Landslide inventories compiled from satellite imagery are systematically incomplete at the small end of the size distribution. At Sentinel-2's 10 m resolution, scars smaller than roughly 1,000 to 2,500 m² are either missed entirely or merged with adjacent failures. PlanetScope at 3 m pushes the detection floor down to approximately 200 to 900 m², depending on contrast and terrain shadow. SPOT-7 at 1.5 m panchromatic can resolve individual scarps at the upper end of that range, but coverage is limited to tasked strips.
This incompleteness matters for two reasons. First, small failures can block roads and cut off communities even if they do not appear in a coarse inventory. Second, landslide frequency-area distributions follow a power law: the majority of failures by count are small, even though the largest failures dominate total displaced volume. A satellite inventory should always be presented with an explicit completeness threshold so that downstream users (road engineers, search-and-rescue coordinators) know what the map cannot see.
Confusion sources that inflate or deflate the count
Several surface types produce spectral or backscatter changes that mimic landslides. Agricultural clearings and harvested fields show reduced NDVI and exposed soil. River channel migration exposes fresh gravel bars with high shortwave reflectance. Wildfire scars, if a fire preceded or accompanied the earthquake, produce nearly identical spectral signatures. Shadow from steep terrain can suppress reflectance in post-event imagery taken at a different solar angle from the pre-event image, generating spurious change signals.
Mitigation strategies include masking agricultural land using pre-existing land-cover maps, applying slope thresholds from a DEM (landslides occur on slopes above roughly 15 to 20 degrees in most settings), and using multi-date SAR coherence to separate persistent change from transient effects. None of these filters is perfect. An honest inventory report states its false-positive and false-negative rates, or at minimum acknowledges that they have not been independently validated.
Turning an inventory into an operational product
A polygon layer of suspected landslide scars is a starting point, not a finished product. For it to be useful to a civil protection agency or a humanitarian logistics coordinator, it needs to be intersected with road network data, population distribution, and river channel geometry to identify which failures are blocking access routes, which communities are isolated, and which scars are positioned to deliver debris into drainage systems during the next rainfall event.
Satellize runs this intersection workflow on open constellations and adds commercial tasking where a client holds the relevant licence, producing prioritised access-route reports and updatable GIS layers rather than static maps. The analytic approach is the same one applied in the Tonga crop-estimation programme: open data processed to a decision-ready output, with explicit documentation of what the imagery can and cannot resolve. The next rainfall after a co-seismic inventory is often as dangerous as the earthquake itself; the inventory's value is in identifying where that risk is concentrated before the rain arrives.
Typical figures
| Optical spatial resolution | 10 m (Sentinel-2), 3 m (PlanetScope), 1.5 m pan / 6 m MS (SPOT-7) |
| SAR spatial resolution | 20 m ground range, IW mode (Sentinel-1); coherence computed at 40–80 m effective resolution after multilooking |
| Revisit (optical) | 2–5 days (Sentinel-2, both satellites); near-daily (PlanetScope); on-demand (SPOT-7) |
| Revisit (SAR) | 6 days per satellite (Sentinel-1 A/B combined: 6-day repeat per track) |
| Minimum detectable landslide scar | ~1,000–2,500 m² at Sentinel-2 resolution; ~200–900 m² at PlanetScope 3 m; completeness threshold must be stated in any delivered inventory |
| Spectral bands used | Red, NIR, SWIR (Sentinel-2 B4, B8, B11, B12); Red, NIR (PlanetScope); C-band VV/VH backscatter and coherence (Sentinel-1) |
| Latency after event | SAR coherence: 0–6 days depending on orbit phase; optical: cloud-dependent, typically 3–15+ days in tropical highlands |
| Archive depth | Sentinel-2 from 2015; Sentinel-1 from 2014; PlanetScope from ~2016; SPOT-7 from 2014 |
| Coverage per acquisition | Sentinel-2: 290 km swath; Sentinel-1 IW: 250 km swath; PlanetScope: global daily mosaic; SPOT-7: 60 km swath |
| Delivery formats | GeoTIFF change rasters, GeoJSON / Shapefile polygon inventories, PDF summary report with completeness statement |
Analytics Satellize can run
| SAR coherence loss map | Interferometric coherence differencing between pre- and post-event Sentinel-1 IW SLC pairs; coherence computed per pixel over a 5×1 or 3×3 multilook window | GeoTIFF probability layer delivered within one SAR repeat cycle of the event, cloud-independent |
| Optical change-detection scar layer | Pixel-based NDVI and SWIR difference between pre-event cloud-free composite and first usable post-event Sentinel-2 or PlanetScope scene; threshold determined by Otsu or ROC analysis | GeoJSON polygon inventory with per-polygon area, centroid coordinates, and confidence score |
| OBIA-refined landslide inventory | Object-based segmentation (multiresolution algorithm) applied to post-event imagery, classified by slope, shape elongation, and spectral attributes; DEM-derived slope mask applied to suppress flat-terrain false positives | Shapefile with attribute table including area, slope, aspect, and distance to nearest road or river |
| Road-network accessibility assessment overlay | Spatial intersection of landslide polygon inventory with OpenStreetMap or national road-network dataset; buffered scar footprints flagged as probable blockages | Prioritised blockage report (PDF + GIS layer) listing affected road segments ranked by downstream population isolation |
| Updatable multi-date inventory | Sequential optical change detection as cloud-free acquisitions accumulate post-event; new confirmed scars merged and previously flagged false positives removed using ancillary land-cover mask | Versioned GeoJSON feed updated on each new usable acquisition, with changelog documenting additions and removals |
| Secondary hazard exposure layer | Landslide scar centroids intersected with upstream catchment delineation and 24-hour rainfall forecast grids; scars in steep catchments above settlements flagged as debris-flow initiation zones | Alert-ready GeoJSON layer with ranked secondary-hazard zones for civil protection briefing |
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.