Landslide Science
Methods
Landslide Science
Methods
This page documents how data is structured, displayed, and analyzed in the Alaska Landslide Inventory viewer. It covers the classification system, how map features are positioned, how areas and volumes are derived, the automated rules that maintain consistency across columns, and how the two timeline charts work.
Rule names throughout this page link to a detail page showing the actual function source (Python or SQL) and a live diff between the rule's output and the current database. Editors can additionally apply rule changes from there; the page is otherwise read-only.
The inventory lives in a PostGIS database (tethys_db) with two
primary tables: landslides (one row per landslide, all
attributes) and landslide_polygons (one row per polygon, with
a landslide_id foreign key). Every landslide has at least one
polygon; catastrophic landslides typically have two or more (separate
source and deposit geometries).
Source geometry is stored in WGS 84 (EPSG:4326) for compatibility with
web standards. Area, volume, and centroid computations are performed in
Alaska Albers Equal Area (EPSG:3338) via PostGIS
ST_Transform — an equal-area projection avoids the
4–6× area distortion that Web Mercator introduces at Alaskan
latitudes.
The complete inventory can be downloaded as GeoJSON + QGIS .qml style files from the ⬇ Download link in the header. The download includes a manifest, the normalized landslides and polygons GeoJSONs, a denormalized flat polygons file with parent landslide attributes merged in, and style files matching this site's map symbology.
Each landslide carries a unique, human-readable unique_name.
Because Alaska placenames recur across distinct features, names follow a
compact disambiguation standard — see
Naming conventions.
The map can display third-party landslide inventories as separate overlays. They are mirrored verbatim and never merged into this inventory: each is stored in its own table with the columns its publisher shipped, and clicking a feature shows exactly the fields they populated, under their own name and citation. Nothing in them is reclassified into our scheme, because doing so would require deciding, record by record, things the source never stated. Counts and statistics on this site describe our records only.
A record of ours may be promoted from one of theirs: an editor creates our record starting from their observation, maps its geometry here, and fills our fields by hand. A promoted record stores which inventory and which of their identifiers it came from, and credits the original observer rather than the editor who promoted it. It enters the same review queue as any other new record.
Subsets come in two kinds. Tags are hand-curated
groupings; the alaska-2025 publication subset is a tag whose
membership was frozen when its archival snapshot was published.
Regions are defined by a polygon, and membership is
computed, not assigned: a landslide belongs to a region when the centroid
of its primary polygon (see §3) falls inside the region
polygon. Membership recomputes automatically whenever geometry changes.
kim without being in alaska.
The map can overlay Sentinel-1 InSAR line-of-sight velocity (ascending and descending tracks) from the OPERA DISP-S1 product suite (© NASA/JPL), as mosaicked and served by the Alaska Satellite Facility's displacement portal; this site re-serves those tiles through a caching proxy and renders them with ASF's color ramp. Display caveats: the tiles are 8-bit quantized (≈0.24 mm/yr steps) and clipped at ±30 mm/yr, so fast-moving landslides saturate — treat the overlay as exploratory visualization, not measurement.
The Analysis tab's OPERA velocity — asc × desc panel is a
2-D histogram of the same value tiles across the pixels
currently on screen: ascending velocity on x, descending on y (log-scaled
density), computed client-side from the tile pyramid for pixels covered by
both tracks. Because the two tracks look at the ground from opposite
sides, motion signatures separate by direction: density stretched along
y = −x indicates opposite line-of-sight signs
— east/west aspect-driven downslope motion — while
y = x indicates same-sign motion in both looks
(subsidence-like / vertical). The same quantization and clipping caveats
apply; counts are pixels, uncorrected for the modest area-per-pixel
variation across a view's latitude span.
The Bedrock geology overlay is the USGS Geologic Map of Alaska (Wilson, Hults, Mull and Karl, 2015, Scientific Investigations Map 3340, 1:1,584,000, public domain), drawn from Macrostrat's tile service (CC BY 4.0) and coloured by age on the international chronostratigraphic scale rather than with the publication's unit colours. Clicking the map queries Macrostrat for the unit under the cursor and shows its name, age range and lithology. This is the same statewide map that supplies rock strength to the DGGS susceptibility model above, so the two layers read against each other. At 1:1.58 M the unit boundaries are kilometre-scale: the layer says what rock a slope is in, not where the contact is. Statewide it is the only seamless map; DGGS quadrangle mapping is finer where it exists.
Three glacier-flow overlays derive from the NASA MEaSUREs
ITS_LIVE
v2 summary mosaic (120 m, Alaska region, CC0):
climatological speed (2014–2022 window, log color
scale — glacier speeds span five orders of magnitude),
seasonal amplitude (amplitude of the fitted annual
velocity cycle — where flow is strongly seasonal), and the
speed trend (dv/dt: red = accelerating, blue =
decelerating; near-steady ice renders transparent so only genuine change
shows). Tiles are pre-rendered from the project's source rasters; colors
are baked, so the ramps live in the repository
(tools/itslive_color_*.txt). These layers are most relevant
to slow landslides carrying the Glacier contact feature.
Velocity data generated using auto-RIFT (Gardner et al., 2018) and
provided by the NASA MEaSUREs ITS_LIVE project (Gardner et al., 2025,
doi:10.5067/JQ6337239C96).
A companion glacier elevation-change overlay shows dh/dt for 2000–2019 from Hugonnet et al. (2021) (100 m, ASTER/ satellite-stereo geodetic rates; dataset doi:10.6096/13, CC BY 4.0): red = thinning, blue = thickening, with near-balance ice (|dh/dt| < 0.25 m/yr, below typical per-pixel noise) rendered transparent. A smoothed checkbox on the overlay swaps in an edge-preserving (bilateral-filtered) build with a color scale sensitive down to ±0.04 m/yr: the filter averages same-regime noise while leaving sharp steps (glacier margins, drawdown fronts) intact, which makes slight signals — notably accumulation-zone thickening — visible that the standard scale's noise floor hides. Unlike the ITS_LIVE velocity layers this dataset is a fixed 2000–2019 analysis — no updating successor at comparable resolution exists as of 2026 — so treat it as a two-decade average, not current conditions. Paired with the velocity overlays it frames the glacier-dynamics context of Glacier contact landslides: thinning rate, flow speed, its seasonality, and its trend.
Elevation, slope, aspect and northness (cos aspect × sin slope, −1 a steep south face to +1 a steep north face) are drawn from one 60 m terrain grid over Alaska and the Yukon margin, built from USGS 3DEP 1/3 arc-second elevation as served by Mapterhorn and projected to Alaska Albers; they are the same grids the permafrost downscale below uses, so an analysis that crosses them with a permafrost layer is comparing like with like. Slope and aspect are the standard Horn gradient on the 60 m grid; in the aspect layer flat cells carry aspect 0 and so read as north-facing, which is why northness (zero on flat ground) is the better layer for slope-direction questions. In the analysis panel's aspect axis flat cells are left out instead: read as 0 they put 2.5 % of the land in the first bin, a threefold spike at north. A second, smaller artefact remains on ground under 1° of slope, where the gradient's two components are often equal and aspect piles up at 45°, 135°, 225° and 315° (three to four times the neighbouring degrees; invisible above 1°); it is a property of the derived grid, so read the aspect axis of nearly flat ground with that in mind. Sea is masked at the grid's own 60 m. Two of these layers behave differently from an ordinary map: slope is recomputed at every zoom level from elevation averaged to that level's ground resolution (about 9 km at the statewide view, 60 m when zoomed in), so zoomed out only ground that is steep at that scale is coloured, and the colour standard is the lidar viewer's — nearly clear below 32°, then yellow at 40°, red at 45° and purple at 50°, the band around the angle of repose where the information is; and contours are drawn in the browser from the same 3DEP terrain tiles rather than stored, at an interval that follows the zoom — 1000 m at the state scale down to 10 m at site scale — with the heavier line labelled.
Three published permafrost products are shown as overlays, each as its authors issued it, plus a terrain downscale of the two coarse ones built here. They measure different quantities and should not be read as three estimates of one thing:
Terrain downscale (60 m). A 1 km cell in Alaskan mountains spans several hundred metres of relief and every aspect, which is the scale at which a landslide sits. The two coarse products are therefore re-expressed on a 60 m grid (USGS 3DEP via Mapterhorn) using the elevation dependence they already contain, anchored on the published value so that each 1 km cell's own number is preserved and only the within-cell terrain variation is added. The local lapse rate is a kernel-weighted regression: at every 1 km cell the product is regressed on elevation over its neighbourhood with Gaussian weights (15 km), so the lapse field is continuous and as smooth as the kernel — contrasts wider than that (interior against coast) survive, cell-to-cell noise does not — and where a neighbourhood has too little relief to determine a lapse (the North Slope, the deltas) a 50 km kernel takes over.
For Gruber, the published PZI is a normal cumulative curve of MAAT (Φ((T₀ − MAAT)/σ); it reproduces his values to 0.002 rms), but his product floors values below 0.1 at 0.01 and cuts to zero at about 0 °C, which would put a step into any elevation profile. The downscale moves MAAT by the fitted lapse (median −5.2 °C/km, 10th–90th percentile −5.9 to −3.9; free-air is −6.5) and converts it through the same curve rescaled to reach exactly zero at 0 °C, refitted to his cells at or above 0.1 (T₀ = −4.58 °C, σ = 2.67, residual s.d. 0.004), so the 60 m index is continuous from 0 to 1. The downscaled MAAT itself is offered as a layer. For Obu, MAGT is regressed on elevation and northness (cos aspect × sin slope), because land-surface temperature does respond to aspect; the fitted lapse is weak and variable (median −1.4 °C/km — snow insulation is doing real work in that model) and the aspect term is negative, north-facing colder, in 93 % of cells (median −0.41 °C per unit northness). Probability is recomputed as Φ(−MAGT / s.d.) with each cell's published spread. Obu has no value over glaciers and some near-glacier ground; the kernel fit is estimated through those gaps from the valid cells around them and the 60 m layers are written there too — read them against the Glacier covered area 2020 outline (Roberts-Pierel and others 2022, NSIDC G10040, Landsat-derived), which is what the Glaciers category carries for exactly this purpose: on ice the value is the model extended, not the model.
The one independent check is Pastick: over a million 60 m cells where
every layer is valid, the downscaled Obu probability correlates with
Pastick's at r = 0.85 against 0.85 for the 1 km original,
0.66 against 0.66 where the 1 km cell spans 200–500 m of
relief and 0.65 against 0.64 above 500 m; Gruber's PZI changes by less
than 0.01 at every relief band. So by this test the downscale is close to
neutral — and because Pastick is blank on barren ground, the test
cannot see the alpine slopes where the redistribution is largest. A
patch-extraction tool in the repository (tools/permafrost/patch.py)
draws, for any 10 km circle, the published 1 km cells against the
60 m cloud and the fitted lapse and curve, which is how the fits are
judged site by site — and the same check is on the map: with two
permafrost products on the analysis panel's axes (below), its
Sample tool reads any circle of the map and the
Permafrost detail button draws the published 1 km cells and
the 60 m downscale against elevation, with the fitted lapse and
curve, and lets you download the numbers.
The Analysis tab's Cross two overlays panel draws the joint distribution of any two mapped quantities over all terrain, with every landslide placed at its own values — the question where does this ground sit in the space of two things we can map, and where do the landslides sit in that same space? The axes are the two sides of the map's wiper: set up a wiper with two overlays and they become x (left) and y (right); pick axes in the panel and the wiper is made to show them, so the map always shows what the plot is crossing. Eighteen quantities can be axes today: elevation, slope, aspect and northness from the 60 m terrain; the Gruber, Obu and Pastick permafrost products and our downscales; the three susceptibility models; summer radar coherence; and ITS_LIVE ice speed.
How the terrain density is built. Every quantity is point-sampled from its native raster at the centre of each cell of a 1 km equal-area lattice (EPSG:3338, 5.9 million cells, 2.3 million on land). For a 1 km product that is exactly the published value; for a 60 m one it is a systematic 1-in-278 sample of the field. It is deliberately not a mean over the cell: a landslide centroid reads the 60 m slope it sits on, and a density of kilometre-mean slopes would be a smoother distribution than the one the landslides are drawn from, so the comparison would mislead. The density of any pair is computed on request from those samples (land cells with a value for both) and drawn log-scaled in blue as the terrain plot. Beside it, on the same axes and over a faint copy of the same terrain, the landslides plot draws each landslide as a point, or per cell as a count or a rate (landslides per 1000 km² of terrain with those values), the latter being the reading that says anything: the terrain density is what turns "landslides are here" into "landslides are here more than terrain is". A grid slider coarsens both plots together in factor-of-two steps, each step merging 2×2 of the cells below it: a merged cell's count is the sum of its cells' counts and its terrain the sum of their areas, so its rate is the area-weighted mean of the rates below it, empty cells included — never a mean of rates. The rate's colour scale is fixed (0.1 to 1000 landslides per 1000 km², the statewide mean being about 0.6), so the same colour means the same rate at every grid step and on every pair of axes, and coarsening shows rates falling as the landslides are spread over more terrain; the count scale is relative to the plot's fullest cell. A cell holding landslides but no sampled terrain has no rate and is shown in the off-scale colour rather than dropped. The DGGS susceptibility classes are ordinal, so that axis is spaced by class position, not class number; a flat cell has no aspect and is left out of that axis.
Sample and brush. The Sample tool shows a circle of the chosen diameter (up to 100 km) under the cursor as it moves over the map; each click reads the ground inside it onto the terrain plot as an orange cloud — from 60 m value archives where both axes have them, otherwise the lattice cells — with the median and 10–90 % range of each axis, and the landslides inside the circle are drawn as points on the landslide plot (over the grid, when a grid mode is on) and marked on the map; turning the tool off removes the sample. While the panel is open each side of the wiper shows exactly one overlay, and only one that can be an axis — choosing another replaces it and the plot follows. A click on either plot brushes a small box around it, a drag any box; the landslides inside are marked on the map, and where both axes have 60 m value archives (elevation, slope, aspect, northness and the four permafrost downscales) the box also paints the map in the sample's orange: every 60 m cell whose two values fall inside it is coloured, on both sides of the wiper, from zoom 8 in — the answer to "where is the ground that is both steep and above 1000 m?", read in the browser from the same value tiles the sample uses. With the two USGS susceptibility models on the axes the box also sets the lw and n10 filter sliders. The panel's state (axes, grid step, box, paint on or off, sample circle) travels in the page URL, so a view can be shared.
What the downscale is and is not. It adds no new
observations: it redistributes each 1 km value across the terrain
inside the cell according to a fitted local lapse, so the published value is
preserved and valley floors and ridgelines are separated. It cannot see
cold-air pooling, snow redistribution, soil or vegetation differences at
the 60 m scale, and it inherits every limitation of the parent model.
The 1 km layers are the published science; the 60 m layers are
this project's inference from them and are labelled as such in the layer
list. Attribution: Gruber (2012), Obu et al. (2019), Pastick
et al. (2015), Roberts-Pierel et al. (2022); downscaled layers
© landslidescience.org, built on Mapterhorn / USGS 3DEP terrain. Build
code is in tools/permafrost/ of the site repository.
Each record carries a derived country attribute: the primary
polygon's centroid is resolved against the
Marine
Regions “EEZ + land union” polygons (Flanders Marine
Institute, CC-BY) — each sovereign's land plus its waters to
200 nautical miles, so a centroid in a fjord or nearshore water
still resolves. This is the classification floor beneath the region layer:
a site outside every survey region still carries its nationality.
Records can carry field photographs, uploaded by
inventory editors and stored alongside the database (a
photos table, linked many-to-many through
landslide_photos — one photograph can document several
landslides, e.g. a single aerial frame spanning adjacent failures). The
original file is preserved byte-for-byte; the site serves resized
derivatives. Capture time, camera position, and view direction are read
from the image's EXIF metadata when present, and remain editor-correctable
afterwards. Note that a photo's location is the camera position,
not the landslide's — an aerial photograph may plot kilometers from
the feature it documents. When a landslide's panel is open, its geotagged
photos draw as white dots at those camera positions.
Display order is curated: editors star a small featured tier (shown first, in chosen order); remaining photos follow in capture-time order. Each photo carries its own photographer credit and license. Photos are supplementary documentation and are not currently part of the GeoJSON download.
Two observations were added in 2026, now values of the Landslide features field: Tsunamigenic (catastrophic records — the failure generated a displacement wave) and Glacier contact (slow records — the moving mass is in contact with a glacier). Both are editor-set, shown in the detail panel's Detection/evidence list when set; coverage is being populated progressively, so absence of the flag does not yet imply absence of the phenomenon.
Every landslide has a type (slow or catastrophic) and a class that encodes both the type and the strength of evidence for ongoing creep (for active landslides) or its age (for prehistoric events). Classes are organized in the sidebar by activity and recency.
| Class | Color | Meaning |
|---|---|---|
| Slow Obvious creep | Red | Clear, ongoing surface deformation visible in satellite imagery |
| Slow Patchy obvious creep | Red | Obvious creep in parts of the landslide, not uniformly |
| Slow Subtle creep | Yellow | Creep detectable by InSAR or careful image comparison, not visually obvious |
| Slow Geomorph creep | Green | Geomorphic indicators of past or very slow creep (e.g., lobate morphology) |
| Small slow landslide | Green | Below the main size threshold but included for completeness |
| Class | Color | Meaning |
|---|---|---|
| Catastrophic Obvious creep | Blue + red halo | Catastrophic failure from a source area showing obvious precursory creep |
| Catastrophic Patchy obvious creep | Blue + red halo | Catastrophic failure with patchy ongoing creep at the source |
| Catastrophic Subtle creep | Blue + yellow halo | Catastrophic failure with subtle creep at the source (InSAR-detectable) |
| Catastrophic Geomorph creep | Blue + green halo | Catastrophic failure with geomorphic creep indicators at the source |
| Catastrophic Cryptic | Blue | Post-2012 catastrophic event where pre-event imagery and InSAR were reviewed but no precursory creep was found — a deliberate null result, distinct from the older Modern / Holocene classes (which typically had no pre-event data to review at all). |
| Class | Color | Meaning |
|---|---|---|
| Catastrophic Modern | Pale blue | Pre-2012 "modern" events (roughly post-1850) |
| Catastrophic Holocene | Pale blue | Prehistoric catastrophic events identified from geomorphology |
| Small catastrophic landslide | Pale blue | Below the main size threshold |
Catastrophic landslides with precursory creep are drawn with a colored halo (stroke) matching their creep evidence level. The “Catastrophic Cryptic” class (no halo) marks post-2012 events whose pre-event data were reviewed and no precursory creep was found.
The classification of any individual landslide is itself derived by an automated rule from a few input columns — see Automated rules below.
Each polygon in the database has a role:
is_primary flagEvery landslide that has been digitized to completion carries exactly one primary polygon, used for dot placement on the map:
is_primary=true. All slow landslides in the inventory have
a primary body polygon.
is_primary=true. Multi-source events have one source
flagged primary, the others unflagged. Deposit polygons are
never flagged primary.
This convention is enforced automatically on every data entry path (in-browser drawing, geometry edits, and file imports): when polygons are added, removed, or re-roled, the primary flag is re-normalized — exactly one primary of the type's primary role, or none when no polygon of that role exists — so the centroid, region membership, and country attribution stay consistent with the geometry.
Each landslide's dot is drawn at the centroid of a single representative polygon, computed in EPSG:3338 then reprojected to WGS 84:
is_primary=true). If no source polygon
exists, the centroid is computed from the deposit polygon as a fallback.
The centroid coordinates are stored as four derived columns on the
landslides table: centroid_albers_x /
centroid_albers_y (meters in EPSG:3338) and
centroid_lat / centroid_lon (decimal degrees in
WGS 84). They are populated by automated rules
(centroid_albers_x
centroid_albers_y
centroid_lat
centroid_lon)
and re-applied whenever polygon geometries change.
Polygons themselves are drawn when zoomed in past zoom level 7.
Per-polygon areas (stored in landslide_polygons.area) are
computed by PostGIS as ST_Area(ST_Transform(geom, 3338)) and
rounded to whole square meters. The Alaska Albers projection guarantees
equal-area accuracy across the inventory's spatial extent.
polygon_area rule
Polygon areas are summed by role into three columns on the
landslides table:
area_body — sum of body-polygon areas. NULL for
catastrophic landslides.
area_body rule
area_source — sum of source-polygon areas. NULL for
slow landslides. Handles multi-source catastrophic landslides naturally.
area_source rule
area_deposit — sum of deposit-polygon areas. NULL for
slow landslides.
area_deposit rule
The boolean size_inclusion column marks whether a landslide
meets the inventory's size criterion:
area_body ≥ 50,000 m²area_source ≥ 50,000 m²
OR area_deposit ≥ 500,000 m²
Landslides below the threshold are tagged "Small slow landslide" or
"Small catastrophic landslide" by the
landslide_class rule,
and the size_inclusion flag is itself populated by the
size_inclusion rule.
Per-polygon volume estimates (landslide_polygons.polygon_volume)
branch on the parent landslide's type and the polygon's role:
| Condition | Formula | Rationale |
|---|---|---|
| Creeping permafrost mass among the landslide features | 20 × area |
Uniform 20 m thickness assumed for permafrost masses, regardless of role |
| Slow, role = body | 0.1 × area1.5 |
Empirical scaling of landslide volume with planform area |
| Catastrophic, role = source | 0.1 × area1.5 |
Same scaling applied to the source area |
| Catastrophic, role = deposit | 2 × area |
Minimum 2 m deposit thickness assumed |
Per-polygon volumes are summed by role into volume_body
volume_body,
volume_source
volume_source,
and volume_deposit
volume_deposit
on the landslides table.
volume_estimated is the landslide-level rule output:
volume_bodymax(volume_source, volume_deposit)
— the two role estimates can disagree because of deposit run-out,
entrainment, or removed material; the larger is the more defensible
estimate of failure size.
Some landslides have a manually-determined volume_site_specific
(from a field study, DEM differencing, or published paper). When present,
it supersedes the formula-based estimate.
volume_preferred is the resulting canonical value used on the
map and in downstream displays:
volume_site_specific is set: use it directlyvolume_estimated
The volume_method column carries either a one-off description
of a site-specific approach (paper citation, DEM-differencing note, etc.)
or the canonical string “Automated estimate” for rows using the
rule-driven value.
volume_method rule
The “Site-specific volume” filter checkbox restricts the map to
landslides with a manually-entered volume_site_specific.
Area and volume sliders use a logarithmic scale:
Landslides with no area or volume data always pass these filters — the filter only excludes records where the value is known and below the threshold.
Many columns are derived from other columns by deterministic rules. Each
rule is a pure function (Python or SQL) implemented in
inventory/derived.py and exposed at
/inventory/manage/rules/ for editors.
The displayed rule source is read at request time via
inspect.getsource() — the displayed version and the
executed version cannot drift apart.
For each rule, the audit page shows the current count of agreements and disagreements between stored values and computed values, with a row-by-row diff including a Δ % column for numeric rules. Editors can apply rule output to the whole inventory in one transaction; the bulk apply writes an audit log entry per affected landslide and invalidates the appropriate caches.
Rules are organized into dependency chains. Within a chain, earlier rules must be applied before later rules read meaningful values.
| Chain | Steps |
|---|---|
| Areas |
polygon_area
rule
→
area_body
rule
/
area_source
rule
/
area_deposit
rule
→
size_inclusion
rule
|
| Centroids |
Independent —
centroid_albers_x,
centroid_albers_y,
centroid_lat,
centroid_lon
(x
y
lat
lon)
(all read polygon geometry directly)
|
| Volumes |
polygon_volume
rule
→
volume_body
rule
/
volume_source
rule
/
volume_deposit
rule
→
volume_estimated
rule
→
volume_preferred
rule
→
volume_method
rule
|
| Creep evidence → class |
insar_creep
rule
→
creep_behavior
rule
→
landslide_class
rule
|
insar_creep
rule
is a tri-state OR-merge of the per-study booleans
(insar_schaefer, insar_kim,
insar_opera, insar_other): True if any positive
reading; False if at least one study returned negative and none positive;
NULL if no study has assessed the site. This preserves the distinction
between "we checked, no creep" and "we haven't checked yet".
creep_behavior
rule
picks the label observed for this landslide from the per-source
flags. The two Planet flags work as a pair, not a simple severity
ladder: planet_labs_creep says
“creep is apparent in Planet imagery,” and
planet_labs_patchy_creep is a refinement saying
“the creep is patchy, not uniform across the feature.”
Conceptually obvious creep is the broader assertion (the whole
feature is moving), but the patchy flag, when set, overrides
that to mean “actually it’s only patchy” — so in
the compositional rule the patchy flag wins. A third flag,
its_live_creep, records motion that shows in the
ITS_LIVE
feature-tracking velocity fields. It is a second, independent route to
Obvious creep rather than a label of its own, and the
patchy flag overrides it just as it overrides the plain Planet flag:
planet_labs_patchy_creep → Patchy obvious creepplanet_labs_creep → Obvious creepits_live_creep → Obvious creepinsar_creep → Subtle creepother_subtle_creep (any non-empty value) → Subtle creep
— one or more observations from a short vocabulary (fine ground
cracking, post-AHAP change, recurrent failures, lidar change, post-Maxar
change, post-LIA change, an OPERA coherence gap, in-situ
instrumentation, …), comma-separated when several applygeomorph_creep → Geomorph creepcreep_evaluated is TRUE →
Cryptic — a deliberate null finding (the
record was reviewed and no precursor was identified)
The creep_evaluated boolean is human-set in the data-entry
UI and is not itself a rule output. It distinguishes “reviewed
and nothing found” (TRUE) from “not reviewed yet”
(FALSE), which matters in particular for older catastrophic events
where there generally was no pre-event data to review at all.
landslide_class
rule
composes the class from landslide_type,
size_inclusion, creep_behavior, and event year:
Records that the rule can't classify (e.g., a slow landslide with no positive creep evidence in any source) are flagged in the legend under Other → Incomplete classification. Logged-in editors can click the ⚙ gear in the map detail popup to open the record for editing.
Temporal precision varies enormously across the inventory. The system encodes each landslide's age as a year number using this priority:
landslide_class contains "Holocene" →
year = −1 (prehistoric)
landslide_class contains "Modern" → year = 0
(post-settlement but pre-2000)
year_text is a four-digit year → that year (e.g., 2013)
date_min exists → year extracted from it
The Min age slider maps positions to this encoding: position 0 = all, 1 = Holocene+, 2 = Modern+, 3–16 = 2012–2025. This provides a coarse temporal filter that works despite the mix of precise dates and order-of-magnitude age estimates.
Catastrophic landslides may have:
date_min /
date_max) — a bracket: the event occurred sometime
between these dates. The span can be hours (for well-observed events) to
months or years.
The seasonal histogram shows when in the year catastrophic events occur. Only landslides with seismic datetimes or valid date brackets are included.
The year is divided into 37 bins: 36 bins of 10 days each (days 1–10, 11–20, …, 351–360) plus one shorter bin for the remaining 5–6 days. The final bin's bar height is normalized to a 10-day equivalent so it is visually comparable.
The "Max uncertainty" slider (default: 30 days) excludes range events with spans longer than the slider value. Setting it to 0 shows only seismically detected events.
Two checkboxes change the y-axis:
volume_preferred rather than a count of 1. The largest
single event in each bin is highlighted as a darker stripe at the top of
that bar.
The timeline histogram shows when in history landslides occurred. All landslides with resolvable temporal data are included (not just time-bracketed events).
The chart uses variable-width bins reflecting increasing data completeness in recent periods:
| Period | Bin width | Visual width |
|---|---|---|
| Holocene (year = −1) | ~10,000 years | 27 units |
| Modern pre-2000 (year = 0) | ~150 years | 45 units |
| Annual 2000–2011 | 1 year each | 9 units each |
| Monthly 2012–present | 1 month each | 1 unit each |
The 2012+ zone is highlighted with a blue background, indicating the period of most complete satellite coverage and systematic inventory.
Raw counts are converted to events per year:
The Holocene and Modern divisors are approximate — order-of-magnitude normalizations for visual comparison, not precise rate estimates.
All filtering is performed client-side on the full dataset. The map and both charts respond to the same set of filters.
Type (Slow / Catastrophic) and class checkboxes are combined with AND
logic. The class checkbox list always renders every known class; classes
with zero matching records appear dimmed and disabled so future data
gaps are visible rather than silently hidden. Records with NULL or empty
landslide_class are caught by the Other →
Incomplete classification checkbox, which would otherwise filter
them out entirely.
Each landslide feature in the vocabulary (grouped by the type it applies
to) has a three-way control: must have, no
preference, or exclude. Settings are AND'd, so
“molards, not tsunamigenic” is one filter. A record with no
features recorded counts as having none. The list grows with the
vocabulary: a feature added on the edit form is a filter on the next load.
In the URL it travels by name — ft=molards,!tsunamigenic
— and links written before 2026-10-06 with the old bitmask still
read as “must have.”
The remaining include-only checkboxes restrict the display to records having an attribute that is not a feature; checked boxes are AND'd:
stream_damming is a non-empty stringvolume_site_specific is setdate_min and date_max are setseismic_datetime is setThese are editor-set observations, so an unchecked box or an unlisted feature means “not recorded” rather than “observed absent” — “must have” narrows to positively-noted records, not to the complete population that has the attribute, and “exclude” keeps every record on which nobody noted it.
Landslide features (since 2026-10-06). Eight former
boolean columns — molards,
exclusively_supraglacial, super_elevated_deposits,
tsunamigenic, precursory_headscarp,
post_2012_activity_increase,
creeping_permafrost_mass, glacier_contact
— are now recorded in one multi-valued column,
features, as a comma-separated list of values from a short
vocabulary that editors can extend (each value applies to slow
landslides, catastrophic ones, or both). The boolean columns were dropped
on 2026-10-07 after a record-by-record check against a pre-transition
copy of the table; exports made before then carry them, exports made
since carry features instead, and importing an older export
ignores the old columns with a warning. Snapshots published before the
change keep their own columns.
These apply a minimum threshold using ≥ comparison. Landslides with missing data for the filtered attribute pass through unconditionally.
Both charts always restrict to the current map viewport. The sidebar counts can optionally be restricted to the viewport via the "Limit counts to map view" checkbox.
All filter settings, slider positions, and panel visibility are encoded in the URL query string — filtered views can be shared as links that reproduce the exact same view.
volume_estimated uses
0.1 × area1.5 for slow body and
catastrophic source polygons. This is an order-of-magnitude
approximation; real volumes can deviate substantially. When a site has
been independently studied,
volume_site_specific supersedes the formula, and those
records are filterable via the "Site-specific volume" checkbox.