Landslide Detection (Sentinel-1)
From radar physics and statistics to the two apps that turn a Sentinel-1 pass into a map of landslides: explained step by step, with interactive demos, and tested on two real disasters.
Find the slopes that moved, even under cloud
After a hurricane or an earthquake, hillsides fail, roads are cut and villages can be buried. Emergency teams need to know where, fast. The trouble is that the same storm that triggers landslides also fills the sky with cloud, so ordinary satellite photos are often useless. Radar is not.
The Sentinel-1 satellites send their own radio pulses down to the ground and record what bounces back. This works at night and through cloud and rain. When a slope slides, the surface and the vegetation on it change, and so does the echo. The WASDI Landslide Detection app looks for exactly that change, and this page explains, from first principles, how it does it.
WHYWhy detect landslides from space?
Landslides are among the deadliest secondary hazards of heavy rain, hurricanes and earthquakes. They often strike steep, remote terrain that is slow and dangerous to survey on foot right after the triggering event. Three facts shape the whole design of this app.
01 · Timing
They cluster exactly when conditions are worstMost landslides are triggered by heavy, prolonged rain that saturates hillside soil until it fails under its own weight. The days right after a major storm are both the most likely time for new slides and the hardest time to reach them.
02 · Visibility
Optical satellites go blind at the worst momentThe storm that triggers the slides also blankets the region in thick cloud for days. A camera-like (optical) satellite cannot see through cloud, precisely when responders most need fresh imagery.
03 · Radar
Radar brings its own light and sees through cloudSynthetic Aperture Radar (SAR) satellites carry their own source of illumination and see through cloud, smoke and darkness. Sentinel-1 revisits the same ground every few days, whatever the weather.
The aim, concretely. Given one or more Sentinel-1 passes from before and after a suspected event, over an area of interest, automatically produce a map of exactly which pixels show the radar signature of a fresh landslide, without waiting for cloud-free optical imagery and without a person comparing before/after images by eye.
APPSTwo apps that work as a team
What you run in WASDI is usually the automatic app. It does the searching for you: you give it an area and a date, it finds the right Sentinel-1 images. Behind it sits the manual app, which contains all of the detection mathematics and works on one list of satellite tracks that it is handed.
They are not two algorithms. They are two doors into the same one. If you already know exactly which radar scenes you want to compare, you can call the manual app directly. If a storm has just hit and nobody has chosen orbit numbers and filenames, the automatic app finds them for you.
A satellite track (an "orbit", or more exactly a relative orbit) is one repeating path over the Earth. Looking at the same place from two different tracks gives two different views of the same slope. The next tabs show why that helps.
The full step-by-step chain of both apps, as a flow chart, is on the Pipeline tab.
MAPHow to read this page
Each tab adds one layer, in the order the app itself uses them. Every technical term is introduced in plain words before it is used.
Tab 2
Radar physicsWhat radar measures, why a landslide changes it, what speckle is and how the app filters it.
Tab 3
StatisticsFrom a simple log ratio to a per-pixel Z-score, a multi-image baseline, and the fusion of two polarizations.
Tab 4
Score to mapAdaptive thresholds, hysteresis, the slope filter, morphological closing, shape filters and multi-orbit fusion.
Tab 5
PipelineThe flow chart of how the automatic app triggers the manual app, and what each does, step by step.
Key terms at a glance
| Term | In plain words |
|---|---|
| Backscatter | The share of the radar pulse that bounces back to the satellite. Its strength depends on how rough, wet and steep the surface is. |
| Speckle | The grainy, salt-and-pepper texture present in every radar image, caused by waves interfering with each other. It is removed before any comparison. |
| Decibel (dB) | A logarithmic scale, so that a doubling always looks like the same step. Backscatter changes are compared in dB. |
| Baseline | The "normal" for each pixel, built from several radar images taken before the event: a typical value and how much it naturally wobbles. |
| Z-score | How many "typical wobbles" the post-event value sits away from that pixel's own normal. |
| Polarization (VH, VV) | Sentinel-1 records two kinds of echo at once. A real change tends to show in both; noise rarely does. |
| Percentile threshold | A cut-off set from the scene's own scores (for example the top 0.5%), so it adapts to calm and stormy scenes alike. |
| Hysteresis | A two-level test: a weak condition to be considered, a strong condition (somewhere in the same region) to be accepted. |
| DEM | Digital Elevation Model, a terrain height map. It gives the slope of every pixel. |
| Orbit / track | One repeating satellite path. Different tracks view the same slope from different directions. |
What radar measures, and why a landslide changes it
A camera records sunlight reflected by the surface. A radar sends its own microwave pulse and records how much of it comes back. That returned energy is called backscatter, and it depends heavily on how rough or smooth the surface is.
1.1Backscatter: the echo from the ground
The middle slope is covered by forest. Its leaves and branches scatter plenty of the pulse back to the satellite, so it appears bright in a radar image.
Think of throwing a ball at a wall. Against a rough, uneven wall it bounces back in many directions and some of it returns to you. Against a smooth, tilted mirror it ricochets away. A radar pulse behaves the same way. The symbol for backscatter is σ0: the fraction of the transmitted pulse that a surface sends back per unit area, calibrated so that images from different dates can be compared.
Backscatter depends on three things above all: the roughness of the surface compared with the radar wavelength, its moisture (wet ground reflects more), and its slope relative to the beam. Sentinel-1 operates in the C-band, a wavelength of about 5.6 cm, so leaves, twigs, soil clods and stones are all the right size to scatter it.
Calm water behaves like a mirror: it reflects the pulse away from the satellite, so it appears almost black. Forest is a jumble of scatterers: it appears bright.
1.2Why a landslide changes the echo
A landslide strips vegetation and soil from the source area (the steep upper section where the slide starts) and leaves behind smoother bare ground or rock. That smoother surface reflects the pulse more like a mirror, mostly away from the satellite. The result is a drop in backscatter at the same location between a "before" and an "after" pass. That drop is the core signal this app looks for.
It happens within seconds, whereas natural changes (crops growing, soil drying, seasons) happen slowly and tend to change the echo gently. A sudden, strong, localized drop on steep ground is unusual, and that is what the statistics on the next tab quantify.
An important, specific detail. General landslide literature often mentions backscatter increasing in the rough debris pile at the bottom of a slope. This app deliberately targets the source area or scarp instead, where the typical response is a backscatter decrease. That is why the terrain filter is called a source slope filter, and why every threshold in the app is built to detect drops, not rises.
1.3Radar looks sideways
A camera on a satellite looks straight down. A radar like Sentinel-1 looks obliquely, to the side, because it measures distance by timing the echo. That has three consequences on steep terrain, all of which matter for landslides:
Foreshortening
Slopes facing the satellite are squashedA slope tilted toward the beam occupies fewer pixels than it should and is compressed in the image, with a strong, bright return.
Layover
Steep faces can fold overOn very steep slopes facing the satellite, the top can be recorded before the bottom, so the image folds over itself and the information is mixed up.
Shadow
Slopes facing away can be hiddenA slope tilted away from the beam, steeper than the beam angle, is never illuminated at all. No echo means nothing can be detected there from that direction.
The same hillside therefore looks different from different satellite tracks. A slope that is hidden in shadow from one track is clearly visible from another. This is the reason the app can combine several tracks (see multi-orbit fusion) and why a single track can miss a slide that its neighbour sees.
1.4Speckle: why every radar image is grainy
Even over a perfectly uniform field or calm lake, a raw radar image looks grainy. This salt-and-pepper texture is called speckle. It is not ordinary sensor noise. It comes from how radar imaging works.
Each ground resolution cell contains many small scatterers (blades of grass, soil grains, rock facets). Each reflects the radar wave back over a slightly different path length. At the sensor, the many small echoes add up wave by wave: sometimes reinforcing, sometimes cancelling out. A patch of ground whose physical properties are perfectly uniform can therefore still produce wildly different brightness values from one pixel to the next.
Speckle is multiplicative: it scales with the signal rather than adding a fixed amount, and it is spatially random. Comparing two dates pixel by pixel is unreliable until it is suppressed, because a speckle spike can look just like a change.
The ground is identical every time. Only the exact positions of the scatterers inside the cell change, yet the pixel flips between bright and dark. Press the button a few times to see it.
Speckle lab
The panels below simulate a hillside with one sharp-edged dark patch (a stand-in for a scarp) on uniform ground. Panel two is what the radar actually delivers. Try the two common smoothing filters and the window size. Watch two numbers: how much the noise in the flat area is reduced, and how much the filter widens a sharp edge. A good filter reduces the first without widening the second.
1.5Pre-processing: from raw scene to analysis-ready image
Before any statistics, every Sentinel-1 scene goes through the same fixed recipe. Both apps use it, so a pre-event scene and a post-event scene are always treated identically. The calibration and despeckling are run through an ESA SNAP workflow, and the order of the steps matters.
- CalibrateConvert the raw digital numbers of the GRD (ground range detected) product into physical backscatter, so that dates are comparable.
- Dual-polKeep both channels, VH and VV, as linear backscatter. They respond differently to the surface (explained on the Statistics tab).
- Fill gapsGaps with no data (scene edges, shadow) are filled from the nearest valid neighbour before filtering.
- DespeckleA 5×5 median filter is applied in the linear power domain, before conversion to decibels.
- To decibelsConvert to dB so that a doubling is the same step everywhere. Only now does the image feed the statistics.
Why the order matters
- Filtering across a raw edge into a patch of zero-valued no-data would drag a false dark halo into the result, so gaps are filled first.
- Speckle is multiplicative and defined on the power scale. Filtering values that are already logarithmic would distort the result, so despeckling is done before the dB conversion.
Why a median, not an average
- The median knocks down speckle's random spikes while a real sharp edge, like a scarp boundary, is not blurred by the filter itself, which an average always does (see the lab above).
- It is robust: one extreme pixel in a window of 25 cannot move the result.
Turning "backscatter dropped" into one reliable number per pixel
Radar images are naturally noisy. Even two images of the exact same untouched hillside, taken days apart, differ pixel by pixel. A method that just compared one "before" image with one "after" image would be swamped by that variation. So the app first builds a small statistical model of normal for every pixel, and only then asks how unusual the post-event image is.
2.1The simplest idea: a log ratio
The classic building block of radar change detection is to divide the post-event backscatter by the pre-event backscatter and express the result in decibels:
Decibels squeeze the huge natural range of radar brightness into a scale where a doubling is always the same step, whatever the starting brightness. It is the same reason sound loudness is quoted in decibels rather than raw pressure. As a rule of thumb, 3 dB is a factor of two and 10 dB is a factor of ten.
On its own, a single before/after pair is too noisy to trust: speckle and ordinary day-to-day variation add their own random ups and downs. So the app does not stop here.
See the change first
A simulated mountain scene with five hidden landslides. Drag the handle and look for where the echo went dark between the two images. Some dark patches are real slides. Some are not.
2.2A real baseline, not just one photo
Instead of one pre-event image, the app collects several from the same satellite track: 2 at minimum, 3 to 4 or more recommended, and 4 by default. For every pixel it computes two numbers from those images:
- the median: a stable "typical" value, resistant to any single noisy image;
- the standard deviation (the spread): how much that particular pixel normally wobbles from one image to the next.
The second number is what makes the method fair. A pixel on a calm, steady surface wobbles very little, so a small change there is remarkable. A pixel on a restless, windblown or wet surface wobbles a lot, so the same change means little. The post-event value is then judged against the pixel's own habits, not against a single global rule.
The pre-event images are chosen from the 100 days before the event by default. Taking them from the same relative orbit means the satellite looks at the ground at the same angle each time, so differences are not caused by changing viewing geometry.
2.3The Z-score: how unusual is the post-event value?
Combine the two ideas and you get one number per pixel and per polarization:
Put in plain words: Z counts how many "typical wobbles" the post-event value sits away from this pixel's own normal range. It is not an absolute decibel change, but a change sized relative to how much that specific pixel naturally jumps around. A Z of −1 is unremarkable background noise almost everywhere. A Z of −6 is not, whatever the raw decibel numbers behind it.
| Term | What it actually is |
|---|---|
| σ0 | Radar jargon for backscatter: the fraction of the transmitted pulse a surface sends back per unit area. The superscript "0" marks it as the normalized, calibrated quantity rather than a raw sensor count. |
| σ0post,dB | This exact pixel's backscatter in the post-event image, in decibels, after despeckling. One number, for one polarization, at one location. |
| median(σ0pre,dB) | This same pixel's typical backscatter across every pre-event image supplied for this orbit. The baseline from 2.2, computed once per pixel. |
| std(σ0pre,dB) | This same pixel's own natural spread across those pre-event images: how noisy or stable that specific spot normally is, not a scene-wide average. |
SIGMA_FLOOR_DB | A configured minimum for the denominator, 0.35 dB by default. Without it, a pixel that happened to look identical in every pre-event image would have a near-zero spread, and dividing by that would turn an ordinary, tiny change into an artificially huge Z. |
Try the Z-score
Four pre-event measurements (dark dots) set the typical value and the spread. Drag the sliders to change how much the pixel normally wobbles and how far backscatter drops after the event.
2.4More than one post-event image? The aggregation options
A track can be given several post-event images, for instance to bridge a gap in the archive or to keep watching a slope for days after the event. The setting POST_AGGREGATION decides how those several decibel differences, at the same pixel, collapse into the single number the Z-score uses. It is applied before the Z-score formula: the result becomes the post-event value in that formula, per polarization, per pixel.
| Mode | How it is computed | What it trades off |
|---|---|---|
closest (default) | Only the single post-event image closest to the event is used. Any others supplied are logged and ignored. | Simple and predictable, and matches the original single-image behaviour exactly, but it throws away any extra post-event images. |
median | The per-pixel median of the difference across every supplied post-event image. | Extra speckle suppression by averaging over more looks, at the cost of blurring a real signal that fades or partly recovers between acquisitions (debris settling, vegetation regrowing). |
max_anomaly | Per pixel, whichever supplied image shows the single strongest (most negative) backscatter drop. | Maximizes recall for a signal that fades or recovers quickly, but is more exposed to a one-off noisy dip in a single image being mistaken for a sustained change. |
In the Jamaica tests, max_anomaly recovered a real, partial detection at a site that closest had missed entirely, while median reshaped detections at the other sites without turning them into noise. Details are on the Real events tab.
2.5Using both polarizations: dual-pol fusion
Polarization is the orientation in which the radar wave oscillates. Sentinel-1 transmits in one orientation and records the echo in two: VV (received in the same orientation, dominated by scattering from the surface itself) and VH (received in the crossed orientation, which is strongly influenced by vegetation and rough, volume-like scattering). They respond differently to the same ground, so a real change tends to show in both, while a random noise spike rarely does.
The app computes a Z-score for each polarization, keeps only the negative deviations (drops, the signature from Tab 2), and combines them into one score D. It works in two steps.
Step A. If backscatter went up at a pixel (Z positive), Z− is forced to exactly 0, so that polarization contributes nothing. Only real drops survive, in each polarization separately. Step B. This is the same arithmetic as the length of the hypotenuse of a right triangle: square both, add, take the root. Treating the two scores as perpendicular axes of one small vector has a useful effect: a strong drop in either polarization alone still gives a sizeable D, while a drop present in both at once gives a larger D than either alone, without one channel being averaged away by the other.
| ZVH | ZVV | Z−VH | Z−VV | D | Reading |
|---|---|---|---|---|---|
| −4.0 | +0.5 | −4.0 | 0 | 4.00 | Drop in one polarization only: still flagged |
| −3.0 | −3.0 | −3.0 | −3.0 | 4.24 | A smaller drop in both scores higher than the larger one in one channel |
| +2.0 | +1.0 | 0 | 0 | 0.00 | Backscatter rose: no evidence of a landslide here |
Why not simply add the two? A plain sum would let one enormous drop in a single polarization (possibly a processing artefact) dominate the total exactly as a genuine two-channel detection would, and it would pass any noise in the second channel straight through. The squared-and-rooted form makes single-channel and dual-channel evidence combine in a way that rewards agreement between the channels instead of just tallying them.
Explore the fusion
Where this leaves us. Every pixel now carries a single anomaly score D: zero where nothing dropped, and larger the more surprising and the more polarization-consistent the drop. The next tab turns that score map into a clean map of landslides.
From a noisy anomaly score to a clean, plausible map
The fused score D still needs several more ideas before it becomes a trustworthy landslide mask: where to draw the line, how to keep fuzzy true edges but not noise, whether the terrain even makes physical sense, how to tidy the shapes, and what to do when more than one satellite track is available.
- ThresholdCore and extent cut-offs from the scene's own percentiles.
- LinkHysteresis keeps extent regions that touch a core pixel.
- TidySlope gate, closing, size and shape filters, cluster mode.
- FuseCombine satellite tracks: either, both, or a support count.
3.1Adaptive thresholds: let the scene set the bar
First, the score map is gently smoothed with a moving window (SCORE_SMOOTHING_WINDOW, 11 pixels by default) to suppress leftover speckle in the score itself. Then the app needs a cut-off. A fixed number would be too strict on a quiet scene and too loose on a noisy one, so the app reads two percentiles from that scene's own distribution of D:
- a strict core threshold,
CORE_PERCENTILE(default 99.5); - a looser extent threshold,
EXTENT_PERCENTILE(default 96.5).
Concretely: sort every D value in this orbit's area of interest and take the value at the 99.5th percentile as DCORE. By construction, only the most extreme 0.5% of pixels in this run clear it. The same scene next month, with different weather and background noise, gets a different number for DCORE but the same meaning: still the top 0.5%. That is what makes the threshold self-calibrating.
3.2Hysteresis: a strict test, a loose test, and connectivity
Hysteresis thresholding is not new to this app. It is the same core idea behind Canny edge detection in ordinary image processing. The general pattern is always: a pixel needs a weak condition to be considered at all, but is only accepted if it is connected to something that clears a strong condition.
It solves one specific problem. A single threshold is always a compromise. Set it strict and you cleanly reject noise, but you also chop the fuzzy, genuine edges off real features. Set it loose and you keep those edges but let a lot of noise through with them. Two thresholds, linked by connectivity, get both at once.
How the linking step works
- Build two masks from the same score map: the core mask (D ≥ DCORE, and on steep enough terrain, see 3.3) and the relaxed mask (D ≥ DEXTENT).
- Label every group of touching pixels in the relaxed mask. A pixel counts as touching its neighbour even diagonally (8-connectivity).
- Any relaxed region that contains at least one core pixel is kept in full, boundary and all. Any relaxed region with none is discarded entirely, however large.
A pixel is never individually promoted or rejected. Whole connected regions are.
Build a mask yourself
This widget runs the second half of the chain on a simulated scene with five real landslides (cyan outlines are the truth) plus a few false-change patches on flat ground. Step through the stages, then move the sliders. Note: the toy scene is far more crowded with change than a real area of interest, so its sensible percentiles are lower than the defaults of 99.5 and 96.5.
3.3Terrain awareness: the slope filter
Landslides in this class of terrain do not start on flat ground. So a terrain model (a DEM, here the Copernicus 30 m model) is used to compute the slope of every pixel in degrees. A pixel can only count as a high-confidence core detection if its slope reaches MIN_SOURCE_SLOPE, 8° by default.
This check has nothing to do with the radar data, so it removes a completely different class of false alarm: an unrelated change in backscatter on a flat field, such as flooding, ploughing or harvest. Those change the echo just as strongly, but a flat field cannot slide.
An optional upper limit, MAX_SOURCE_SLOPE, can also exclude near-vertical bare-rock faces where radar geometry creates its own false changes.
A trade-off to know about. A fixed 8° gate is tuned for mountain source areas. In the Jamaica tests, a real slide above Montego Bay sat on only 2 to 3° slopes and was below the gate by design, not by mistake. Lowering the gate for urban or gentle terrain is an option, at the price of more false alarms.
3.4Morphological closing: bridging small gaps
The mask that survives hysteresis is rarely one solid blob. Speckle and small terrain effects can punch pinholes in it, or split what should be one continuous scarp into several almost-touching fragments. Morphological closing is a standard operation from mathematical morphology, the toolbox behind most binary image clean-up. It runs two steps back to back:
- Dilation: grow every masked region outward by a fixed number of pixels in every direction.
- Erosion: shrink the result back inward by the same amount.
A small gap between two fragments is bridged by the dilation (both sides grow until they touch). The erosion then shrinks the merged shape back, but the bridge, no longer a gap, stays filled. A large empty area far from any masked pixel is never affected, since dilation only grows what is already there. Net effect: small gaps and pinholes disappear while the size and position of real regions barely change.
The size is set by MORPH_CLOSING_KERNEL (default 11 pixels): large enough to bridge realistic gaps, small enough not to weld together two unrelated regions.
3.5Shape filters: size and elongation
After closing, the mask is split into individual connected regions again, and each is checked against two simple physical expectations about what a landslide looks like in a radar image.
MIN_CLUSTER_SIZE_PIXELS · 100
Not too smallA region of only a handful of pixels is far more likely to be a speckle leftover than a slide. Anything below the minimum area is rejected. The flip side: genuinely small slides can be lost (see Limits).
MIN_ELONGATION · 2.0
Not too roundElongation is the ratio between a region's long axis and its short axis. A landslide scar and its track are typically long and narrow along the slope. A compact, round blob is more likely a patch of changed farmland or noise. Regions below 2:1 are rejected.
3.6Cluster modes: which regions make it into the final mask
What happens to the regions that survive the size and elongation filter is controlled by CLUSTER_MODE. This is a separate decision from thresholding or closing: it is about how much to trust regions that lie far from the strongest detection in the scene. The strongest region(s) act as anchors.
| Mode | How it decides what to keep | Best suited to |
|---|---|---|
multi (default) | Every surviving region whose mean confidence is within MULTI_CLUSTER_RELATIVE_STRENGTH (default 0.6, that is 60%) of the single strongest region becomes its own anchor. Any other region is kept only if it falls inside one of the anchors' proximity zones. | A storm that triggered several separate, independent slides in one area: the normal real-world case. |
single | Exactly one anchor, the single highest-confidence region in the scene. Others are kept only if they touch that anchor's proximity zone. | An area known to contain one dominant landslide, or reproducing an older run. |
off | No proximity check at all. Every region that passed the size and shape filter is kept as it is. | Maximum recall when false alarms are acceptable, or to see what the size and shape filter alone leaves behind. |
The proximity zone is a fixed dilation of each anchor's own shape (30 iterations in the implementation). It is large enough that a real secondary lobe of the same landslide, a short distance from the main body, still counts as connected, while an unrelated speckle cluster on the far side of the area does not.
3.7Multi-orbit fusion: two viewpoints are better than one
As Tab 2 explained, radar looks at the ground obliquely. A steep slope facing one satellite track can be squashed or thrown into radar shadow, while the same slope is clearly visible to a track arriving from another direction. Neither is wrong: they are simply looking from different angles.
So the pipeline above is run independently for each track, giving one clean mask per track. Those masks are then combined.
The support-count map
With two or more tracks, the app does not stop at a union and an intersection. It also keeps the arithmetic between them. For every pixel it simply counts how many tracks flagged it:
- Union (“either”): every pixel with support ≥ 1. More complete, more false alarms.
- Intersection (“both agree”): every pixel with support equal to the number of tracks. Cleaner, may miss slides only one track can see.
- Everything in between is a graded confidence. A pixel two of three tracks agree on is not the same as one only a single track saw.
The count is written out as its own output file, so a user can set their own confidence bar afterwards (for example, "trust anything two independent passes agree on") without rerunning the detection.
Compare tracks on the simulated scene
In the toy scene, each track has its own blind spots and its own false alarms. Real slides appear in both. Switch between the two tracks, "Either" and "Both agree", and toggle the filters.
In a real test (Wayanad, two tracks): requiring both tracks to agree cut the detected area from 6,249 to 2,131 pixels while giving up only 86 of 942 true-positive pixels. Precision against the optical reference rose from 0.15 to 0.40. The Real events tab has the full numbers.
3.8Hazard prior: an optional hint from a global hazard map
Some terrain is simply more likely to slide than other terrain. The World Bank / GFDRR Global Landslide Hazard Map ranks every land pixel on Earth from 1 (very low) to 4 (high) landslide hazard, combining rainfall and earthquake triggers. When USE_HAZARD_PRIOR is switched on, the app uses that rank as a hint, never as a verdict.
The hint works on the strict (core) test only. Each pixel's rank multiplies the core threshold at that pixel, so terrain the global map rates as unlikely has to show a stronger anomaly before it can anchor a detection. The relaxed, extent threshold is not touched.
The map is switched off by default. When it is on, the app looks for the file LS_TH_COG.tif in the workspace, reuses it if it is there, and otherwise downloads it once from the World Bank Data Catalog.
Soft, and in one direction only. The multipliers are never below 1.00, so the hint can only raise the bar, never lower it, and never to the point where nothing could pass. A strong enough radar signal is still detected in a “very low” hazard area: the radar evidence always has the final say.
| Rank | Hazard | Multiplier on the core threshold | Effect |
|---|---|---|---|
| 1 | Very low | 1.50 | Needs a meaningfully stronger anomaly to count |
| 2 | Low | 1.20 | Needs a somewhat stronger anomaly |
| 3 | Medium | 1.00 | No change |
| 4 | High | 1.00 | No change: not given a discount |
Why “high” hazard is not rewarded
An earlier version also lowered the bar in high-hazard terrain. On real mountain terrain that backfired. The map is coarse (about 1 km) and “high” covers most of a mountainous area, so lowering the bar over half the scene let noise through wherever a single track had a marginal pixel. The “both agree” mask filtered that noise back out, but the “either” mask did not. Since the point of the hint is to cut false alarms, not to recover missed slides, only the low-hazard ranks adjust anything now.
- Pixels outside ranks 1 to 4, for example beyond the map's coverage (60°S to 72°N), are treated as neutral (multiplier 1.00).
- The scene's core and extent thresholds are computed exactly as before, so runs with and without the hint stay comparable. Only the per-pixel comparison against the core threshold changes.
- The multipliers can be changed with the advanced settings
hazard_prior_rank1_factortohazard_prior_rank4_factor(defaults 1.5, 1.2, 1.0, 1.0). - The downloaded file is about 55 MB. It is kept between runs unless
DELETEis true (see Tab 8).
Try the hazard hint
Pick the hazard rank at a pixel and drag its anomaly score. The scene's normal core threshold is fixed at 3.00 here as an example (every real scene sets its own). Watch how the score needed in low-hazard terrain moves, while a score that is strong enough still passes.
How the automatic app triggers the manual app
Everything on the last three tabs (the baseline, the Z-score, the thresholds, the terrain filter, the multi-orbit fusion) lives in exactly one place: the manual app. The automatic app never touches that mathematics. Its entire job is working out which Sentinel-1 images to feed the manual app, and then handing off in a single call.
4.1The whole chain at a glance
Read the chart from top to bottom. The user supplies two things, an area and an event date. The automatic app turns them into an exact list of satellite tracks with named before and after images. The manual app turns that list into landslide maps.
- Make a terrain modelIf none is supplied, a 30 m DEM is generated for the area.
- Search Sentinel-1Query a window around the event date (100 days back, 100 forward by default).
- Group by orbitWork out each scene's relative orbit and group the scenes by track.
- Filter and rankKeep tracks that cover enough of the area, then rank them.
- Select imagesPick the pre-event baseline images and the post-event image(s) for each track.
- Import and pre-processFetch missing scenes and run the standard dual-pol pre-processing.
- Pre-processCalibrate, fill gaps, despeckle, convert to dB (Tab 2).
- BaselinePer-pixel median and spread from the pre-event images (Tab 3).
- Z-score and fusionPost-event deviation, drops only, VH and VV combined into D (Tab 3).
- Threshold and slopeAdaptive core and extent thresholds, hysteresis, slope gate (Tab 4).
- Clean upClosing, size and elongation filters, cluster mode (Tab 4).
- Fuse tracks and writeUnion, intersection and support count, then write the maps (Tab 4).
4.2The automatic app, step by step
Its input is just an area and an event date. Its job is to search Sentinel-1, work out the relative orbit of each scene, check footprint coverage, and rank and select orbits and images.
Automatic app
Steps 1 to 3: discover- Auto-generate a DEMIf no terrain model is supplied, one is generated for the area (Copernicus 30 m) with the DEM extractor app. It will be needed for the slope filter.
- Search Sentinel-1A search runs over a time window around the event date:
DAYS_BACKandDAYS_FORWARD, 100 days each by default. The window must be wide enough to hold several pre-event scenes from the same track. - Group by relative orbitScenes are grouped by satellite track. The relative orbit is not written in the file name, so it is derived from the absolute orbit number (see the box below).
Automatic app
Steps 4 to 6: select and hand off- Filter and rank tracksA track is kept only if its scene footprints cover at least
MIN_ORBIT_COVERAGE(10%) of the area. Tracks are then ranked: highest average coverage first, then the smallest gap between the event and the first post-event scene, then the most pre-event images. - Select the imagesFor each track, the best
NUMBER_OF_PRE_IMAGES(4 by default) before the event build the baseline, andNUMBER_OF_POST_IMAGES(1 by default) after it are tested against it. - Import, pre-process and delegateAny scene not yet in the workspace is imported and given the same dual-polarization pre-processing the manual app uses. Then everything is passed to the manual app in a single call.
How the tracks are chosen: if you name them yourself (ORBITS), exactly those are used. If you give a number (NUMBER_OF_ORBITS), that many top-ranked tracks are used. If you give neither, every track with a usable before and after set is used. A track needs at least 2 pre-event images and at least 1 post-event image, otherwise it is skipped.
Relative orbit: which of the 175 repeating ground tracks?
Sentinel-1 keeps two different counters, and it helps to think of them as a lap counter and a lane number.
- Absolute orbit is the lap counter. It starts at 1 at launch and only ever goes up, by one for every lap around the Earth (roughly 14 to 15 laps a day).
- Relative orbit is the lane number, from 1 to 175. The satellite flies over 175 fixed ground tracks one after another. After 175 laps, about 12 days, it starts the same sequence again and flies over the very same ground.
So two scenes of the same place, 12 days apart, have lap numbers exactly 175 apart but the same lane number. Only scenes with the same lane number see the ground from the same angle, which is what the baseline needs.
The catch: a scene's file name contains the lap counter but not the lane number. The app has to work the lane number out.
Why this matters. The statistics assume that the only thing changing between images is the ground. If a pre-event scene came from a different lane, the viewing angle would differ and the backscatter would differ for reasons that have nothing to do with a landslide. Grouping by relative orbit is what makes the baseline valid.
How the app works it out, in three steps
- Read the lap counter from the file name. It is the seventh piece, between underscores (highlighted below).
- Subtract the satellite's offset. Each satellite started its lane numbering at a different lap: 73 for Sentinel-1A, 27 for 1B and 42 for 1D.
- Divide by 175, keep the remainder, and add 1. The remainder tells you how far into the 175-lane cycle you are. Adding 1 makes lanes count from 1 instead of 0.
S1AsatelliteIWimaging modeGRDHproduct type1SDVdual-pol20240725T004900start time20240725T004925stop time054910absolute orbit06B02Cdata-take ID8680unique ID
In short: S1A: (abs − 73) mod 175 + 1 S1B: (abs − 27) mod 175 + 1 S1D: (abs − 42) mod 175 + 1. Here “mod 175” means the remainder after dividing by 175.
Check it on the Wayanad baseline
These are the five Sentinel-1A scenes the app used for track 63. Each lap number is exactly 175 more than the one before, 12 days later, and every one lands on the same lane.
| Scene date | Absolute orbit (lap) | Minus 73 | Remainder after dividing by 175 | Relative orbit |
|---|---|---|---|---|
| 19 Jun 2024 | 54,385 | 54,312 | 62 | 63 |
| 1 Jul 2024 | 54,560 | 54,487 | 62 | 63 |
| 13 Jul 2024 | 54,735 | 54,662 | 62 | 63 |
| 25 Jul 2024 | 54,910 | 54,837 | 62 | 63 |
| 6 Aug 2024 (post-event) | 55,085 | 55,012 | 62 | 63 |
Sentinel-1C has two offsets: 172 for scenes before 24 June 2026 and 99 from that date on, because its numbering changed. The app applies whichever matches the scene date.
4.3The manual app, step by step
Its input is an exact list of tracks, each with named before and after images. It runs the following chain once per track and then combines the tracks. It is exactly the same code path whichever app called it.
Manual app
Steps 1 to 3: build the anomaly score- Pre-processEach GRD image is calibrated and despeckled into dual-pol (VH + VV) linear backscatter through an ESA SNAP workflow, with gaps filled and a 5×5 median filter applied before conversion to dB.
- Build the baselinePer-pixel median and standard deviation across every pre-event image, per polarization (floored at
SIGMA_FLOOR_DB). - Z-score and dual-pol fusionCombine the post-event image(s) by
POST_AGGREGATION, compute the Z-score, keep only backscatter drops, and fuse VH and VV into the single anomaly score D, then smooth it.
Manual app
Steps 4 to 6: turn the score into maps- Threshold and slope gateAdaptive core and extent percentiles, hysteresis linking, and the requirement that core pixels reach
MIN_SOURCE_SLOPE. - Clean upMorphological closing, then rejection of regions below
MIN_CLUSTER_SIZE_PIXELSorMIN_ELONGATION, and theCLUSTER_MODEproximity rule. - Fuse tracks and write the outputIf several tracks were supplied, combine them (union, intersection and support count), then write GeoTIFF and PNG masks. If a reference raster is given, also score the result.
Details worth knowing
- Terrain is measured before it is resampled. Slope is computed on the terrain model at its own resolution, and only then resampled onto the analysis grid. Resampling the elevation first would smooth away real relief and under-estimate steep slopes.
- A check on the two polarizations. If VH and VV come out almost identical (correlation above 0.995), the app logs a warning, because that points to a problem in how the two channels were read for that track.
- No data is handled per track. Each track has its own no-data mask. In the fused masks a pixel is marked no-data only where every track is invalid, and a track never counts as support for a pixel it could not see.
- The strict test is the only one the hazard prior touches. The optional hazard hint (section 3.8) changes the core threshold pixel by pixel. The relaxed, extent threshold is never changed.
4.4An example run: what "pre" and "post" really look like
For the Wayanad event (30 July 2024), the automatic app resolved two satellite tracks and delegated the following images to the manual app. Notice the 12-day spacing within each track, four pre-event scenes forming the baseline, and two post-event scenes (this run asked for NUMBER_OF_POST_IMAGES = 2 with POST_AGGREGATION = "median").
The baseline for each track is built only from that track's own pre-event scenes, so viewing geometry is identical. The two tracks are then processed independently and fused at the end. A depth of four pre-event images is the default because it gives a usable median and spread, while more images mean more data to fetch.
4.5What comes out
| Property | Value |
|---|---|
| File types | GeoTIFF (for analysis) and PNG (for a quick preview) |
| Pixel values | 1 landslide · 0 no landslide · 255 no data or insufficient history |
| Per-track mask | {BASENAME}_{TrackLabel}_landslide-mask.tif |
| Fused masks (two or more tracks) | {BASENAME}_fused-union_landslide-mask.tif, {BASENAME}_fused-intersection_landslide-mask.tif |
| Support count | {BASENAME}_orbit-support-count.tif: a product in its own right, not just a step |
| Preview colours | PNG masks: red landslide, grey no landslide, pale blue no data |
| Optional diagnostic plots | backscatter-change, anomaly-thresholds, overlap-map (only when exactly two tracks are used) and, if a reference is given, a confusion-map (green correct, red false alarm, blue missed) |
| Result record | The settings used, the list of output files and, if a reference was given, the scores. The automatic app passes on the manual app's record unchanged. |
Scoring against a reference
If a reference raster of known landslides is supplied (GROUND_TRUTH_FILE), the app scores its result with standard numbers, for each track and for the fused masks:
- Precision: of the pixels flagged, the share that are truly landslide.
- Recall: of the true landslide pixels, the share that were found.
- IoU (intersection over union): the overlap divided by the combined area of both masks. It punishes both misses and false alarms.
- F1: the balance of precision and recall in one number.
- AUC (per track): how well the raw anomaly score ranks true landslide pixels above all others, before any threshold is applied.
4.6Why have both apps?
Reach for the manual app when…
You already know the imagesReprocessing a known event with tuned parameters, comparing specific scenes, or experimenting with the thresholds, without searching the archive again.
Reach for the automatic app when…
A storm just hit and you need whatever evidence existsThe far more common real situation: nobody has pre-selected orbit numbers and file names by hand. Give it an area and a date, and it finds the data and runs the whole chain.
Tested on two real disasters
The app was run on landslides that happened in India in 2024 and in Jamaica in 2025, then checked against independent references: a second satellite sensor in the first case, and expert-drawn landslide outlines in the second. The two audits show where the method works well, and where it does not.
AThe event, and the area chosen
Extremely heavy rain, as much as 572 mm in the 48 hours before the event, saturated the slopes above the villages of Mundakkai and Chooralmala in Meppadi panchayat, Kerala. Two slope failures struck in the early hours of 30 July 2024, the second far larger, sending a rockslide-induced debris flow down the Punnapuzha river valley. ISRO's National Remote Sensing Centre measured roughly 86,000 m² of source-zone land movement and a run-out of about 8 km. Officially confirmed deaths exceed 220. It is one of the deadliest single landslide events recorded in India.
The area of interest, about 13 by 13 km (11.457 to 11.576°N, 76.094 to 76.208°E), was centred on Meppadi panchayat so that it contains the affected villages, not drawn around a single press coordinate. Three independently reported landmarks fall well inside it: the landslide headscarp, and the centres of Mundakkai and Chooralmala, with coordinates taken from sources that have nothing to do with the app.
The run
- Event date
- 30 July 2024
- Tracks used
- 165 and 63, fused
- Pre-event images
- 4 per track (12-day spacing)
- Post-event images
- 2 per track, combined by
median - Search window
- 100 days back, 100 days forward
- Terrain model
- Copernicus 30 m, generated automatically
- Thresholds
- 99.5 / 96.5 percentile, slope ≥ 8°
- Other settings
- all at their defaults
BWhat each track saw
All four panels trace the same corridor from the headscarp, through Mundakkai, toward Chooralmala, not scattered noise on the surrounding terrain. The maps sit on a hillshade of the terrain model. Red marks pixels flagged as landslide.
Point-checked against independently reported landmarks
Beyond pixel statistics, each output was sampled directly at the three landmark coordinates.
| Landmark | Track 63 | Track 165 | Either | Both agree | Optical |
|---|---|---|---|---|---|
| Headscarp top of the source zone | 0 | 1 | 1 | 0 | 0 |
| Mundakkai village centre destroyed, on the debris path | 1 | 1 | 1 | 1 | 0 |
| Chooralmala village centre | 0 | 0 | 0 | 0 | 0 |
Mundakkai is flagged by every radar product, with a support count of 2 at that pixel: both tracks agree independently. The headscarp is caught only by track 165, the same one-track-sees-it asymmetry that multi-orbit fusion exists for. Neither radar nor optical flags the Chooralmala point itself, but both pass close by. A village's nominal centre and the actual debris line are not always the same pixel.
CChecked against a second sensor
No surveyed landslide outline was available for this event. So the radar result was compared with an independent detection of the same event made from Sentinel-2 optical images by another WASDI app, a different sensor and a different algorithm. The optical result was reprojected onto the radar grid and compared pixel by pixel. This is agreement between two independent methods, not validation against surveyed ground truth, but the two share nothing except the real event.
| Radar output | Detected px | Agreeing with optical (of 1,094) | Precision | Recall | IoU |
|---|---|---|---|---|---|
| Track 63 alone | 3,285 | 879 | 0.268 | 0.803 | 0.251 |
| Track 165 alone | 5,095 | 919 | 0.180 | 0.840 | 0.174 |
| Either track (union) | 6,249 | 942 | 0.151 | 0.861 | 0.147 |
| Both agree (intersection) | 2,131 | 856 | 0.402 | 0.782 | 0.361 |
Intersection is the clean product
Precision more than doublesRequiring both tracks to agree cuts detected area from 6,249 to 2,131 pixels while giving up only 86 of 942 agreeing pixels. That is exactly what multi-orbit fusion is designed to do.
Recall is consistently high
78 to 86%Every radar product, single track, union or intersection, recovers at least 78% of the optical corridor. This is not a marginal or lucky detection.
Radar draws a wider swath
2,131 vs 1,094 pixelsEven the tightest radar map has about twice the optical pixel count. The coarser radar grid (about 20 m) plus the closing step naturally draws a thicker corridor than the fine optical line.
DThe signal itself, not just the mask
This matters because it rules out the main alternative explanation for a strong-looking mask: an unstable score distribution that a threshold happens to slice favourably. Here the signal is visibly real in the raw decibel difference, before any percentile or morphology ever runs. The debris corridor already appears as a coherent drop (blue) in both polarizations.
The score histogram shown in section 3.1 is from this same event: a clean, single-humped distribution with the core threshold sitting naturally in the tail.
EAssessment
Strengths
- Independent landmarks line up. The headscarp, Mundakkai and the debris corridor all sit where a completely separate run, on a different sensor, also put them.
- High recall in every fusion mode.
- Intersection earns its keep. Precision rose from 0.18 to 0.40 against track 165 alone, for a 6 percentage-point recall cost.
- Both tracks ran cleanly and the signal is real, not a threshold artefact.
Issues
- A wider swath than the optical map, plausibly the coarser grid plus the closing kernel widening a narrow true corridor.
- Not ground-truthed. Both compared masks are automated outputs. No surveyed inventory was found for this event.
- Single-track visibility asymmetry persists: the headscarp is seen by one track only.
- Village-centre points are a blunt landmark for a narrow debris line.
AThe storm, and how the sites were chosen
Hurricane Melissa formed in the central Caribbean on 22 October 2025 and made landfall near the Westmoreland and St Elizabeth parish border on 28 October as a Category 5 hurricane, the most powerful storm ever recorded striking Jamaica. Sustained winds near 295 km/h and up to 49 inches of rain over the interior mountains triggered widespread flooding and slope failures.
Sites were not picked freehand. Three come from the USGS rapid post-storm imagery assessment, a 5 by 5 km severity grid. Two come from the Copernicus Emergency Management Service (activation EMSR847), whose analysts drew precise landslide footprints by photo-interpretation: human-verified outlines rather than a coarse grid cell. Every area of interest was computed from those datasets, not eyeballed.
| Site | Terrain | Reference | Round 1 result |
|---|---|---|---|
| 1. Blue and John Crow Mountains Portland / St Thomas | Steep, rural | USGS grid cell: major, widespread (one of only 3 cells so rated) | Detected, strong |
| 2. Ginger Hill Westmoreland / St Elizabeth | Steep, rural | USGS grid cell: localized | Detected, one track |
| 3. Trout Hall Clarendon | Hilly, rural | USGS grid cell: localized | Detected, weak |
| 4. Hills above Montego Bay St James | Gentle, built-up | Copernicus EMS: 3 photo-interpreted polygons (454 pixels) | Missed |
| 5. Rennock Lodge, Kingston Surrey County | Urban hillside | Copernicus EMS: 1 point, 75 m buffer (50 pixels) | Missed |
BThe five sites, one by one
1. Blue and John Crow Mountains Detected, strong
This 10 by 10 km area sits inside the single highest-severity cell USGS identified anywhere in Jamaica. Both tracks (150 ascending, 40 descending, four pre-event images each) independently found several discrete, elongated clusters on rough terrain, with almost no spatial overlap between them. That is exactly the layover and shadow behaviour that multi-orbit fusion exists to catch, not a flaw.
The score histograms were clean and single-humped on both tracks, with adaptive thresholds sitting well out in the tail (DCORE of 2.49 and 2.92).
2. Ginger Hill Detected, one track
Round 1 detected this site on track 113, which delivered a tight, spatially coherent diagonal chain of clusters.
3. Trout Hall Detected, weak
The weakest of the three USGS-sited detections, appropriately so: it sits in a "localized" cell, not the "major" severity of the Blue Mountains. The core thresholds here (1.88 and 2.84) are markedly lower than at the other rural sites, tracking a genuinely smaller anomaly. The app's relative detection strength across sites 1 to 3 lines up with USGS's own independent severity ranking (4,604, then 1,776, then 841 pixels), a good sign that the anomaly score measures something real and not just noise.
4. Hills above Montego Bay Missed, then partly found
Checked, not assumed. The reference pixels read as valid data that the app simply did not flag, and the georeferencing of detection and reference match to sub-micro-degree precision. It is a real miss, correctly located.
Why, precisely. At the true polygon, track 150's score is real but weak (mean 0.53 against an area background of 0.32). Track 113 barely sees the site at all: only 5 of 454 reference pixels contain valid data, the rest is shadow or layover. The slope there is only 2 to 3°, far below the 8° gate.
This is the case that moved in round 2 (below).
5. Rennock Lodge, Kingston Missed
The clearest miss of the five. At the reference point, track 150's mean score (0.09) sits below the area's own background mean (0.17): there is no detectable backscatter-change signal there at all, on this track, at any threshold. A second track errored out server-side (the same failure pattern as site 2). Round 2 confirmed the result with a different aggregation mode and every track running, so this is a clean negative and not a partial run.
No tuning would have saved it. It is a small slide on an urban hillside, below what a 20 m radar image can resolve.
CCould tuning have saved the misses?
Rather than guess, the Montego Bay anomaly-score map was cached once (the expensive step) and only thresholding, the slope gate and shape filtering were re-run across 135 parameter combinations, sweeping the two percentiles, the slope gate, the minimum size and the elongation. Every result was checked against the real Copernicus polygon, not just whether the mask was non-empty.
| Configuration | Detected px | True-positive px (of 454) | Recall | IoU |
|---|---|---|---|---|
| Default: 99.5 / 96.5, slope ≥ 8°, size ≥ 100, elongation ≥ 2.0 | 0 | 0 | 0.0% | 0.000 |
| Slope ≥ 4°, size ≥ 20 | 35 | 0 | 0.0% | 0.000 |
| Slope ≥ 0°, size ≥ 50 | 639 | 15 | 3.3% | 0.014 |
| Core 99 / extent 95, slope ≥ 4° | 326 | 17 | 3.7% | 0.022 |
| Best found: core 98 / extent 93, slope ≥ 8° | 381 | 21 | 4.6% | 0.026 |
Best-case trade-off
Over 90% false positivesEvery configuration that recovers any of the true landslide does so by flooding the rest of the area with unrelated detections. The best IoU was 0.026.
Second track visibility
5 of 454 pixelsOnly 1% of the true polygon is valid, non-shadowed data on the second track. Multi-orbit fusion cannot rescue a site that one geometry cannot see.
Slope at the true site
2 to 3°A physically gentler slope than every rural case: not a misconfigured cut-off.
111 of the 135 combinations did give a non-empty mask, so the parameters can be loosened. But loosening them does not recover the real feature, it only lowers the bar for noise elsewhere. For these two sites the limit is the physical signal, not the thresholds that sit on top of it.
DRound 2: does combining more post-event images help?
Landslide scars do not vanish the day after the storm, so a natural follow-up is to give the app more post-event passes to choose from, combined differently. All five sites were re-run with 4 post-event images (up from 1) and a search window of 100 days forward (up from 60). The aggregation was split by hypothesis: median on the three rural hits (to suppress false positives) and max_anomaly on the two urban misses (to catch a fading signal).
Montego Bay: a real, partial recovery
With max_anomaly and 4 post-event images, track 150 went from 0 detected pixels to 344, of which 25 land on the actual Copernicus polygon. More telling than the numbers: the detected region is one large, coherent, elongated shape sitting at the downslope end of the true landslide and touching it, not scattered noise. The cost is visible too: the score distribution shifted much higher (the core threshold rose from about 2.1 to 6.16), the expected price of taking a per-pixel maximum over more images.
Kingston: confirmed, still nothing
Zero detected pixels again, but this time all three tracks completed, so it is a clean negative. It rules out "a single-image fluke" and reinforces that there is no usable backscatter drop at this site on any date checked, under any aggregation mode.
The hits, re-tested with median
| Site | closest (round 1) | median (round 2) | Change | IoU vs round 1 | Cross-track agreement |
|---|---|---|---|---|---|
| Blue Mountains | 4,604 px | 4,308 px | −6% | 0.336 | 0 px (both rounds) |
| Ginger Hill | 1,776 px (1 track only) | 2,348 px (2 tracks fused) | first full fusion | - | 161 px, first time ever |
| Trout Hall | 841 px | 645 px | −23% | 0.250 | 0 px (both rounds) |
Bottom line. max_anomaly delivered a real, if small, recovery at Montego Bay with a plausible shape: worth adopting for weak or fading signals despite its higher false-positive exposure. median never degraded a detection into noise, but its false-positive benefit remains unproven because the coarse USGS grid cannot grade a boundary-level change. And the clearest outcome of the whole round was not about aggregation at all: getting both tracks to succeed together at Ginger Hill produced the one thing no single-track run can, independent cross-track confirmation (161 pixels).
EWhat the Jamaica tests teach
What worked
- Detection strength tracked independent severity (major, then localized), without the app ever seeing that ranking.
- Multi-orbit fusion earned its keep: at the Blue Mountains the two tracks found almost entirely different clusters on the same slope.
- Self-calibrating thresholds held up: every successful run had a clean single-humped score histogram, no per-scene tuning needed.
- Failures degraded gracefully: a single track erroring out did not sink the run.
- Honest misses: where it missed, it missed cleanly, with no false detections invented to compensate.
What did not
- Blind to small, gentle, urban slides: 2 of 2 misses against precise ground truth, against 3 of 3 hits on rural mountain terrain.
- The terrain gate assumes rural-steep source areas. A fixed 8° gate excludes real slides on gentler hillsides by design.
- No feedback when a track is structurally blind: a "successful" track can see only 1% of a site, and nothing in the output says so.
- Per-track failures are silent at run level.
- Parameters cannot rescue a weak physical signal: 135 combinations, best IoU 0.026.
What to keep in mind, and where it is heading
The app produces a map of candidate landslides quickly, over a large area, whatever the weather. Treat it as a way to point people, drones and optical imagery at the right slopes, not as a final damage count. This page collects what the validation work showed about where it is strong and where it is not.
6.1Where it is strong
Terrain
Steep, rural, mountainous source areasAll three rural sites in Jamaica and the whole Wayanad corridor were detected. This is the terrain the slope gate and shape filters were designed for.
Agreement
Independent checks line upRadar matched an independent optical detection with 78 to 86% recall at Wayanad, and its relative strength followed an independent expert severity ranking in Jamaica.
Robustness
Self-calibrating and honestPercentile thresholds adapt to each scene. Where the app misses, it misses cleanly, without inventing detections.
6.2Known limits
- Small slides can hide. A slide smaller than a few radar pixels (a pixel is about 20 m) may not change the echo enough, and the minimum size filter removes tiny regions by design.
- Gentle and urban slopes. The steepness filter is set for mountain source areas. A real slide on a 2 to 3° urban hillside was below the gate. Lowering the gate recovers little without flooding the map with false alarms.
- Shadow and layover. Some slopes cannot be seen from some tracks. In one test a second track saw only about 1% of the true slide area.
- Other changes look similar. Flooded fields, ploughing or harvest can change the echo. The terrain and shape checks remove many, not all.
- Needs earlier images. The baseline needs several pre-event scenes from the same track. Where the archive is thin, the area is marked "insufficient history".
- Not a survey. Results should be confirmed on the ground or with other imagery. At Wayanad both compared maps were automatic outputs, and the radar swath is wider than the optical line.
6.3Reading an empty map
A map with no detections is not the same as "no landslide". There are at least three different reasons a pixel can be empty, and the output does not yet tell them apart:
- nothing happened there;
- something happened but it was too small, too gentle or too weak to register;
- the satellite track could not see that ground at all (shadow, layover or no data).
Rule of thumb. Treat a detection as strong evidence and an empty area as unknown when the slopes are gentle, built-up, small in scale, or seen by only one track. That is why the roadmap below includes observability masks and confidence flags.
6.4Where it is heading
Susceptibility maps
A first hazard hint is built in; more to comeThe World Bank / GFDRR global landslide hazard layer is already available as an optional soft prior that cuts false positives in unlikely terrain (see Tab 4). Next steps: let users bring their own, finer susceptibility map, and test the hint on more events.
Drone confirmation
Drone images as ground truthDrone surveys of a few slopes can confirm detections and tune thresholds for a region.
Radar exclusion maps
Say where radar cannot seeMasks for shadow, layover and no-data areas, and a report of the share of each track's eligible area actually observed, so a blank result is not mistaken for "no landslide".
Radar plus optical
Combine with the Sentinel-2 appJoint detection that uses optical imagery when the sky is clear and radar when it is not.
Ideas from the validation work
- Surface track-level observability. Report what fraction of each track's eligible area was valid data, so a "successful" track that was blind at a site does not read as a clean negative.
- Make the terrain gate context-aware. Offer a lower or user-set minimum slope profile for urban and peri-urban areas.
- Escalate per-track failures. Flag a silently dropped track at run level, so users decide whether to rerun.
- Treat "zero detections" as its own outcome, with a confidence flag tied to valid-pixel coverage and slope-gate exclusion area.
- Use precise reference data where it exists. Wire emergency-mapping outlines to the scoring option automatically after an event.
Run it on your own area
The app runs in the WASDI cloud, so there is nothing to install. All you need is a WASDI account, an area and an event date. Everything else has a sensible default, which this page lists in full so that you can tune it when you need to.
7.1Quick start
- Open WASDI and pick a workspace. Use one you already have, or create a new one from the app's own screen. Your results are saved in the workspace the app runs in.
- Find the Landslide Detection (Sentinel-1) app in the WASDI marketplace and open it.
- Draw a box around the area of interest. Start with something around 10 by 10 km.
- Enter the event date. The app looks for radar images before it and after it.
- Run the app. It picks the images and satellite tracks, runs the detection, and writes the maps to your workspace.
What you get back
- A landslide mask for each satellite track.
- A combined mask for "either track" and one for "both agree".
- An orbit support count showing how many tracks flagged each pixel.
- Optional diagnostic plots that show the radar change and the thresholds used.
Tip. Use both agree when you need few false alarms, and either track when missing a slide would be worse than checking an extra one. The support count lets you choose in between.
7.2Automatic app parameters
Only the first three are required. All of the manual app's tuning parameters (next section) are also accepted and passed through unchanged.
| Parameter | Default | Meaning |
|---|---|---|
BASENAME | required | Prefix used to name every output file. Underscores are removed automatically. |
EVENT_DATE | required | The date to search around. The manual app has no equivalent. |
BBOX | required | The area of interest, as two corner coordinates. |
DEM_FILE | none | Optional terrain model. Generated automatically (Copernicus 30 m) if omitted. |
DAYS_BACK / DAYS_FORWARD | 100 / 100 | The search window around the event date. It never extends past today. |
MIN_DAYS_DISTANCE | 0 | Minimum number of days between a pre-event image and the event date, to keep images taken just before the event, when the ground may already be changing, out of the baseline. |
MIN_ORBIT_COVERAGE | 10 (%) | Minimum share of the area that a track's scenes must cover to be kept. |
ORBITS | none | Optional: the exact tracks to use, as a list such as [63, 165] or a text such as 63,165. The easiest way to repeat an earlier run. Overrides NUMBER_OF_ORBITS. |
NUMBER_OF_ORBITS | unset | Optional: use only this many tracks, best-ranked first. If both this and ORBITS are left out, every track with a usable before and after set is used, a good way to find out which tracks cover a new area. |
NUMBER_OF_PRE_IMAGES | 4 | How many pre-event images to select per track (they build the baseline). |
NUMBER_OF_POST_IMAGES | 1 | How many post-event images to select per track. |
PROVIDER | AUTO | Which data provider to search. |
FORCE_RERUN | false | If true, re-runs instead of reusing an existing terrain model or result. |
7.3Manual app parameters
| Parameter | Default | Meaning | Explained in |
|---|---|---|---|
BASENAME, BBOX, DEM_FILE | required | Output prefix, area, and the terrain model for the slope filter. | 4.3 |
ORBITS | required | List of tracks: each with a label, its pre-event images and its post-event images. | 4.4 |
POST_AGGREGATION | closest | How multiple post-event images combine: closest, median or max_anomaly. | 2.4 |
SIGMA_FLOOR_DB | 0.35 | Minimum baseline variability (dB), to keep the Z-score stable. | 2.3 |
SCORE_SMOOTHING_WINDOW | 11 | Speckle suppression on the anomaly score. | 3.1 |
CORE_PERCENTILE | 99.5 | Strict detection threshold, as a percentile of the scene's own scores. | 3.1, 3.2 |
EXTENT_PERCENTILE | 96.5 | Relaxed threshold, used only where connected to a core pixel. | 3.2 |
MIN_SOURCE_SLOPE | 8.0° | Minimum terrain slope for a core detection. | 3.3 |
MAX_SOURCE_SLOPE | none | Optional: exclude near-vertical bare-rock false positives. | 3.3 |
MORPH_CLOSING_KERNEL | 11 | Bridges small gaps in a fragmented detection. | 3.4 |
MIN_CLUSTER_SIZE_PIXELS | 100 | Reject regions smaller than this. | 3.5 |
MIN_ELONGATION | 2.0 | Reject regions too compact to be a landslide track. | 3.5 |
CLUSTER_MODE | multi | multi, single or off. | 3.6 |
MULTI_CLUSTER_RELATIVE_STRENGTH | 0.6 | How strong a region must be, relative to the strongest, to become its own anchor. | 3.6 |
GROUND_TRUTH_FILE | none | Optional reference raster; enables scoring with IoU, precision, recall and F1. | 4.5 |
USE_HAZARD_PRIOR | false | Optional soft hint from the World Bank / GFDRR hazard map: raises the core threshold in terrain rated unlikely to slide, never lowers it. | 3.8 |
hazard_prior_rank1_factor to hazard_prior_rank4_factor | 1.5, 1.2, 1.0, 1.0 | Advanced: the threshold multiplier for each hazard rank (very low to high). | 3.8 |
ISOLATE_TO_STRONGEST_CLUSTER | unset | Older on/off switch kept for compatibility: true acts as CLUSTER_MODE = single, false as off. Used only if CLUSTER_MODE is not set. | 3.6 |
GENERATE_DIAGNOSTIC_PLOTS | true | Set to false to skip the diagnostic pictures and write only the masks, for faster, leaner runs. | 4.5 |
DELETE_INTERMEDIATE_FILES | true | Removes this run's per-image pre-processed files. The original Sentinel-1 files, the terrain model and all results are kept. | below |
DELETE | true | A broader clean-up: also removes this run's raw Sentinel-1 files, terrain model, reference raster and hazard file, leaving only the results. | below |
About clean-up. With the defaults, a finished run leaves only its results in the workspace: the downloaded Sentinel-1 files, the terrain model and the hazard file are deleted. This keeps the workspace tidy, but a re-run then has to fetch and pre-process everything again. While you experiment on the same area, set DELETE to false so those files are reused. Clean-up only ever touches files that this run used or produced, never other runs' files in the same workspace.
7.4A short tuning guide
| If you see… | Try… | At the price of… |
|---|---|---|
| Too many false alarms | Raise CORE_PERCENTILE, use the "both agree" mask, raise MIN_SOURCE_SLOPE or the size and elongation limits. | Missing weaker or smaller slides. |
| Slides that are missing or clipped | Lower CORE_PERCENTILE or EXTENT_PERCENTILE, use the "either" mask, lower MIN_CLUSTER_SIZE_PIXELS. | More false alarms. |
| A signal that fades quickly | Use several post-event images with max_anomaly. | More exposure to one-off noisy dips; a higher score distribution overall. |
| Very noisy results | Use several post-event images with median, or a larger SCORE_SMOOTHING_WINDOW. | Blurring a signal that partly recovers. |
| Gentle or urban terrain | Lower MIN_SOURCE_SLOPE. | Flat-ground false alarms such as flooding. |
| Fragmented detections | Raise MORPH_CLOSING_KERNEL. | Welding together unrelated regions. |
| False alarms in terrain the global hazard map rates unlikely | Switch on USE_HAZARD_PRIOR. | A little more required signal in low-hazard areas; a real slide there must be strong. |
| One dominant slide expected | Set CLUSTER_MODE to single. | Losing independent slides far from the anchor. |
References
The events, data sets and background material behind this page, so that every claim can be checked at its source.
8.1Sources and data
- Wayanad event details and coordinates: The Landslide Blog (AGU Eos), "The 30 July 2024 Wayanad landslides in Kerala, India"; 2024 Wayanad landslides; village coordinates from Mundakkai and Chooralmala.
- Independent radar-based damage assessment of the same event: Data for Public Good, Landslide Damage Assessment, Wayanad, Kerala (not spatially compared here).
- Jamaica: USGS, 2025 Hurricane Melissa Landslide Hazards and its rapid imagery assessment data release (DOI 10.5066/P1JMKWMF); Copernicus Emergency Management Service, activation EMSR847; NOAA National Hurricane Center report AL132025.
- Radar background: NASA ARSET, Using SAR for Landslide Detection; ESA SNAP tutorials, Sentinel-1 SAR pre-processing; WASDI Python tutorial.
- Hazard hint: World Bank / GFDRR Global Landslide Hazard Map (ranks 1 to 4, used by the optional
USE_HAZARD_PRIORsetting). - Radar data: Copernicus Sentinel-1. Terrain: Copernicus 30 m DEM. Case-study figures were produced from WASDI runs of the landslide detection apps in September 2026.