Short answer

A landslide inventory is the spatial record of where landslides have actually happened — the single most important dataset in any landslide study, because it is the ground truth that every susceptibility and hazard model is built on and validated against. In GIS it is a vector layer (points or polygons) carrying attributes for type, age, source, and confidence. The hard part is not drawing it; it is understanding what the inventory misses, because an inventory's completeness bias propagates directly into any model trained on it. Get the inventory and its limitations right first, then the modeling becomes defensible.

What goes in a usable inventory

A folder of dots on a map is a picture, not an inventory. A usable inventory is a structured attribute table. A practical minimum schema:

  • id — unique, stable identifier
  • geom — point or, better, polygon footprint
  • type — using a standard classification (Varnes / Hungr): rotational slide, translational slide, debris flow, rock fall, topple, lateral spread, complex
  • date — absolute date if known, otherwise a relative age class (historic, recent, relict) and a temporal confidence
  • source — field survey, aerial photo interpretation, satellite/InSAR, historical archive, reported event
  • mapping_confidence — definite / probable / questionable
  • area_m2 — for polygons
  • zone — source/scarp vs. transport vs. accumulation (for detailed inventories)
  • trigger — earthquake, rainfall, snowmelt, anthropogenic, unknown (when attributable)

The two attributes people skip — date and confidence — are exactly the two that turn a display layer into an analytical one. Without date you cannot build an event inventory or a frequency, and without confidence you cannot filter for training a model.

Point vs polygon, and why the zone matters

A point inventory stores one centroid or scarp location per failure. It is fast to compile from reports and small-scale imagery, and it is fine for regional density mapping. Its weakness is that it discards size, shape, and the distinction between where a slope failed and where the material ended up.

A polygon inventory maps the failure footprint. The best practice, especially for any statistical modeling, is to separate the source (depletion/scarp) area from the accumulation (deposition) area. This is not pedantry: susceptibility analysis asks "where will a slope fail?", so the predictor relationships (steep slope, particular lithology, high wetness index) must be sampled at the scarp, not in the runout where slopes are gentle and the material is debris. Training on whole-footprint or accumulation polygons quietly biases the model toward the wrong terrain.

For runout and consequence work the accumulation zone is what you need — so the cleanest inventories keep both and label them.

Completeness bias: the issue that decides everything

Every inventory is a sample, and the sample is rarely random. The dominant biases:

  • Temporal/visibility bias. Recent failures with fresh scarps are easy to map; older ones revegetate and disappear from imagery. An aerial-photo inventory under-represents relict slides.
  • Accessibility/consequence bias. Failures near roads, settlements, and infrastructure get reported; failures in remote terrain do not. A model trained on this learns "landslides happen near roads," which is an artifact, not physics.
  • Resolution bias. Small failures vanish below the resolution of the imagery used; an inventory from 30 m imagery systematically misses small debris flows visible at 1 m.
  • Operator/method bias. Photo interpreters, field teams, and automated InSAR/optical change detection produce systematically different inventories of the same area.

The practical consequence: before you use an inventory as ground truth, you must characterize its completeness. A common diagnostic is the landslide frequency-area distribution — plotting the number of landslides against their area on log-log axes. Well-sampled inventories typically follow a power law (inverse-gamma) above a "rollover" size; the rollover indicates the size below which the inventory becomes incomplete. If your distribution rolls over at a large size, your inventory is missing the small failures, and you should restrict modeling and validation to the size range where it is complete.

A worked screening workflow in GIS

Suppose you have a polygon inventory and want a first-pass susceptibility screen over a study area, in QGIS / GDAL with a projected CRS (UTM, meters) for honest slope and area.

  1. Normalize inputs. Reproject the inventory, DEM, and geology to one projected CRS. Confirm the DEM resolution is adequate (10-30 m for regional screening; finer for site work).
  2. Derive predictors from the DEM.
    • Slope: gdaldem slope dem.tif slope.tif (degrees). Most shallow failures cluster in characteristic slope bands by material; very gentle and near-vertical slopes are often less susceptible to a given failure type.
    • Aspect (proxy for moisture, insolation, and storm exposure): gdaldem aspect dem.tif aspect.tif.
    • A wetness/curvature term: compute the SAGA Topographic Wetness Index or plan/profile curvature via the QGIS Processing toolbox.
  3. Add thematic predictors. Geology/lithology (rasterized), distance to faults, distance to streams (gdal_proximity.py), land cover.
  4. Extract predictor values at scarp locations. Sample each predictor at the source-zone polygons (or a random set of points within them) to learn the conditions where failure occurred. Sample an equal set of non-landslide points for contrast.
  5. Weight the predictors. For a transparent screen use a heuristic weighting or weights-of-evidence; for a statistical model use logistic regression. Avoid presenting any of these as a calibrated probability without validation.
  6. Combine predictors into a relative susceptibility index and classify into a few classes (e.g., low/moderate/high), not a falsely precise continuous percentage.
  7. Validate by withholding part of the inventory (or using temporally independent events) and computing a success/prediction-rate curve and its area under the curve (AUC). Report it.

Common pitfalls and why they happen

  • Treating the inventory as complete. It almost never is; the frequency-area rollover usually proves it. The fix is to characterize and bound the analysis to the complete size range.
  • Mixing source and accumulation zones. Because polygons are drawn as one shape by default, models get trained on runout terrain. Separate the zones.
  • Mixing inventories of different vintages and methods without flagging it. A combined point+polygon, field+InSAR layer with no source/date columns cannot be filtered or audited.
  • Reporting susceptibility as probability. Susceptibility is relative spatial likelihood; hazard adds a time/frequency dimension, and risk adds exposure and vulnerability. Calling a susceptibility class a probability overstates certainty.
  • Ignoring the trigger. A rainfall-triggered inventory and a coseismic inventory have different spatial logic; pooling them blurs both.

QA and validation

  • Confirm every record has type, date/age, source, and confidence; reject or flag records that do not.
  • Plot the frequency-area distribution and report the completeness rollover.
  • Check topology: no overlapping duplicate polygons from merged datasets, valid geometries (ST_IsValid or QGIS Fix Geometries).
  • Validate the model with held-out landslides and a success-rate/AUC curve; an AUC near 0.5 is no better than random.
  • State scale, DEM resolution, inventory completeness, and review boundaries on the final product.

Bathyl perspective

We treat the inventory, not the model, as the deliverable that earns or loses trust. The most useful thing we can hand a planner is a clearly attributed inventory with an honest completeness statement, because a susceptibility map is only ever as good as the failures it was trained on. Screening maps then say what they can decide — and explicitly where a specialist geotechnical assessment must take over.

Related reading

Sources