Why does DEM resolution change my slope and landslide results?

Slope is computed from elevation differences between neighbouring cells, so a coarser DEM averages over a larger footprint and systematically reduces calculated slope. A 30 m DEM can understate steep slopes by several degrees versus a 1 m lidar DEM, shifting which cells cross a landslide-susceptibility threshold.

How does vertical datum error affect flood screening?

Flood models compare ground elevation to a water surface, both in metres. If the DEM is in an ellipsoidal height (e.g. raw GNSS) and the flood level is in an orthometric datum like NAVD88 or EGM2008, the offset can be tens of metres and silently inundate or spare the wrong areas. Confirm both are in the same vertical datum.

What data-quality checks matter most for terrain risk models?

Confirm horizontal CRS and vertical datum, check DEM resolution against the decision scale, inspect for voids and stripes, verify source date against current ground conditions, and compare derived layers to a known control. Document map scale because a 1:250,000 source cannot support site-level decisions.

How do I document uncertainty in a screening map?

State the DEM source, resolution, vertical datum, and date; mark the screening as susceptibility, not assessed risk; and label which boundaries are regulatory versus modelled. Add a confidence note and an explicit threshold for when specialist review is required.

Data Quality for Terrain Risk Models

Short answer

The accuracy of a terrain risk model is capped by the quality of its inputs, and the dominant inputs are the DEM and the hazard layers laid over it. Resolution, vertical datum, source date, and map scale each propagate directly into slope, flood depth, and landslide susceptibility outputs — usually silently. A screening map built on a 30 m global DEM with an unchecked vertical datum can place a hazard threshold tens of metres from reality while looking perfectly plausible on screen.

This article explains how each error source propagates, gives concrete QA checks with tools and thresholds, and shows how to document the residual uncertainty so the map supports decisions without overstating them.

Hazard, susceptibility, and risk are not the same

Loose terminology causes overconfident maps, so fix the vocabulary first:

Hazard is a potentially damaging process with a location, magnitude, and frequency (a 1% annual-chance flood, a debris-flow source zone).
Susceptibility is the spatial likelihood that a process could occur, from terrain factors — slope, lithology, drainage. It carries no return period.
Risk combines hazard with exposure (what is there) and vulnerability (how badly it is harmed) and consequence.

Most GIS terrain "risk" layers are actually susceptibility screening. Calling a susceptibility surface a risk assessment is the single most common overreach, because it implies engineering-grade consequence analysis that the data does not support.

How input quality propagates into outputs

DEM resolution drives slope and curvature

Slope is the gradient between neighbouring cells, so a coarser grid averages elevation over a larger footprint and smooths away local steepness. The same hillside can read as 18 degrees on a 30 m SRTM DEM and 27 degrees on a 1 m lidar DEM. If a landslide model flags cells above 25 degrees, the DEM choice alone decides whether that slope is in or out. Match resolution to the decision: regional screening tolerates 30 m; a site or single corridor needs lidar at 1–2 m.

Vertical datum errors silently mislocate water

Flood and inundation screening compares ground elevation to a water-surface elevation. Both must share a vertical datum. A DEM delivered in ellipsoidal height (raw GNSS / WGS84 ellipsoid) differs from an orthometric height (NAVD88, EGM2008/EGM96 geoid) by the geoid undulation — commonly 20 to 60 m and varying spatially. Subtract a flood stage in one datum from terrain in another and the model floods or spares entirely the wrong cells. Always confirm the vertical datum in the DEM metadata, not just the horizontal CRS.

Source scale and date

A geological or hazard polygon digitised at 1:250,000 has a positional tolerance of hundreds of metres; using it for a site decision implies a precision the source never had. Likewise, a landslide inventory or land-cover layer from 2009 may not reflect current conditions after wildfire, construction, or new failures. Both scale and date belong in the metadata and in the confidence statement.

Voids, stripes, and resampling artefacts

Global DEMs contain voids (SRTM in steep terrain), banding (ASTER GDEM stripes), and vegetation/building offsets in surface models. A DSM includes canopy and rooftops; a DTM is bare earth. Running slope on a DSM over forest produces spurious steep cells at canopy edges that look like cliffs.

A QA workflow before the model runs

Confirm CRS and vertical datum. Read both from metadata. Reproject to a metric projected CRS for analysis (a UTM zone, or national grid such as EPSG:2154 / EPSG:27700) so slope is computed in metres, not degrees.
Match resolution to the decision. Document the DEM resolution and state explicitly what scale of decision it can support.
Inspect the raw DEM. Hillshade it (gdaldem hillshade) and look for voids, stripes, and vegetation artefacts before deriving anything.
Normalize all layers to one CRS and grid before overlay; misaligned hazard polygons over the DEM create false exposure.
Separate the categories. Keep observed events, susceptibility models, regulatory zones, and derived screening scores in distinct layers, clearly labelled.
Set an escalation threshold. Define in advance which outputs trigger specialist hydraulic, geotechnical, or regulatory review.

Deriving slope correctly

# reproject DEM to a metric CRS, then compute slope in degrees
gdalwarp -t_srs EPSG:2154 -r bilinear dem.tif dem_2154.tif
gdaldem slope dem_2154.tif slope_deg.tif -compute_edges

Computing slope on a geographic CRS (degrees of lon/lat) without correction is a frequent error — the horizontal and vertical units differ, so the gradient is meaningless unless the tool applies a scale factor. The equivalent in QGIS is the Slope algorithm under Raster terrain analysis; in ArcGIS Pro it is the Slope Spatial Analyst tool. All assume linear (metric) ground units.

Authoritative reference layers

Anchor screening to official products where they exist, and label them as regulatory rather than modelled:

FEMA flood maps / NFHL define regulatory flood hazard zones in the United States under the National Flood Insurance Program — these are authoritative, not your own model output.
USGS landslide hazard products provide susceptibility and inventory layers and document their own uncertainty.
NOAA sea-level-rise viewers demonstrate how agencies communicate confidence ranges rather than single lines.

Mixing these official zones with your own susceptibility score in one symbology, without labels, is how a screening overlay gets mistaken for a regulatory determination.

Worked example: corridor landslide screening

Acquire 1 m lidar DTM for the corridor; confirm horizontal CRS (EPSG:2154) and vertical datum (NGF-IGN69).
gdaldem slope and gdaldem aspect from the reprojected DTM.
Add lithology (note its 1:50,000 source scale) and a landslide inventory (note its date).
Combine slope > 25 degrees, susceptible lithology, and proximity to past failures into a susceptibility class — explicitly a screening surface, not a risk map.
Compare flagged cells to mapped historical landslides as a control; if known failures fall outside flagged areas, revisit the thresholds.
Label every layer with source, scale, date, datum, and a confidence note; flag corridor segments for geotechnical review.

Common pitfalls and why they happen

Calling a screening overlay a risk assessment. It skips exposure and consequence, implying analysis the data never did. The cause is conflating susceptibility with risk.
Mixing official and modelled layers unlabelled. A viewer cannot tell a FEMA zone from your buffer, so the modelled layer borrows unearned authority.
Slope on the wrong DEM or CRS. Coarse resolution understates steepness; a geographic CRS makes gradient meaningless. Both produce thresholds that misclassify cells.
Ignoring vertical datum in flood work. An ellipsoidal-vs-orthometric mismatch inundates the wrong ground while the map looks fine.
Trusting a DSM as bare earth. Canopy and buildings inject false steep cells; use a DTM for terrain analysis.

Validation

Compare derived slope/aspect against a higher-resolution DEM over a sample area to gauge resolution bias.
Validate susceptibility against an independent inventory of known events; coincidence is the simplest sanity check.
Confirm vertical datum agreement between DEM and any water-level layer before flood screening.
Record source, scale, date, CRS, and datum for every layer so a reviewer can reproduce and challenge the model.

Bathyl perspective

We frame terrain risk maps as decision support with stated limits, never as a substitute for hazard assessment. The value is early clarity about where to look and a clean, documented path to specialist review when safety, permitting, or insurance is on the line. A screening map that hides its DEM resolution and datum is not faster — it just fails later and less visibly.