Sensors
- Sentinel-1 C-band SAR (ESA): 10 m ground range detected resolution, 6-day repeat at mid-latitudes (12-day single-satellite). VV and VH polarisations. C-band (5.4 GHz) backscatter is sensitive to surface dielectric constant changes caused by near-surface soil moisture; penetration depth in dry soil is a few centimetres, limiting subsurface inference but making it responsive to surface and near-surface wetness anomalies.
- ALOS-2 PALSAR-2 L-band SAR (JAXA): 1–10 m resolution depending on mode; 14-day repeat. L-band (1.2 GHz) penetrates dry soil to roughly 0.5–2 m, giving sensitivity to moisture at depth within the embankment toe zone and providing complementary information to C-band surface returns. Full polarimetry (HH, HV, VH, VV) in some modes aids decomposition of surface and volume scattering.
- Landsat 8/9 TIRS (USGS/NASA): 100 m thermal infrared resolution (resampled to 30 m product), 16-day repeat, two thermal bands (Band 10 ~10.9 µm, Band 11 ~12.0 µm). Seepage-induced evaporative cooling depresses land surface temperature (LST) at the toe relative to the surrounding dry embankment. Temperature depressions of 2–5 °C are detectable under clear-sky conditions; cloud cover is a hard constraint.
- ICEYE X-band SAR (commercial): Spotlight mode achieves sub-1 m resolution; tasking on demand. X-band (9.65 GHz) is highly sensitive to surface roughness and very shallow moisture changes. Rapid revisit through the full ICEYE constellation (multiple satellites) supports alert-driven tasking after a Sentinel-1 change flag, at the cost of commercial licensing.
Why wet soil at the toe is not a minor inconvenience
Tailings dam failures kill people and destroy catchments. The Brumadinho dam collapse in 2019 and the Mount Polley breach in 2014 sit at the extreme end of a long documented record of embankment failures, many of which showed seepage anomalies in the weeks or months before collapse. Seepage through or beneath an embankment raises pore-water pressure, reduces effective stress in the fill material, and can initiate piping, internal erosion or slope instability. None of those processes are visible on the dam crest. They announce themselves, if at all, as wet patches, springs or saturated ground at the toe.
Remote sensing cannot replace piezometers or toe drain flow measurement. It can, however, provide a spatially continuous, temporally consistent record of surface and near-surface moisture state across the entire downstream face, including areas that field inspectors visit infrequently. That is the honest value proposition: not a replacement for instrumentation, but a wide-area screening layer that flags anomalies worth investigating on the ground.
What a change in backscatter actually measures
SAR backscatter intensity is governed principally by surface roughness and the dielectric constant of the target material. Liquid water has a dielectric constant around 80; dry sandy tailings material sits below 5. When seepage raises near-surface moisture content, the dielectric constant rises sharply and backscatter increases in a way that is largely independent of solar illumination or cloud cover. This physical relationship is well established in the soil-moisture retrieval literature and underpins operational products such as ESA's Sentinel-1-derived Surface Soil Moisture (SSM) product.
C-band (Sentinel-1) responds to the top few centimetres of soil. L-band (PALSAR-2) reaches deeper, which matters at the toe where seepage may emerge from within the embankment body rather than at the very surface. The two bands are genuinely complementary: a C-band anomaly with no L-band signal suggests a shallow surface effect (rainfall, condensation); a coincident L-band anomaly suggests moisture at depth, which is more diagnostically significant. Distinguishing the two requires careful baseline construction. A wet-season rainfall event will brighten backscatter across the entire scene; the analysis must isolate spatially localised anomalies relative to a multi-temporal baseline built from many acquisitions under similar meteorological conditions.
Change detection rather than single-image thresholding is the operationally sound approach. A coherent change detection or log-ratio method applied to a stack of Sentinel-1 IW (Interferometric Wide) swath images, typically 20 or more acquisitions over six to twelve months, establishes a per-pixel baseline. Deviations beyond two standard deviations that are spatially confined to the toe zone and persistent across multiple consecutive acquisitions are the target signal. Transient rainfall spikes are filtered by requiring persistence.
Thermal infrared: the evaporative cooling signature
Landsat 8 and 9 TIRS measure emitted thermal radiation and, after atmospheric correction, yield land surface temperature (LST) with an absolute accuracy of approximately 1–2 °C under favourable conditions. A seepage-fed wet patch at the toe will, during daylight hours, lose energy to latent heat flux rather than sensible heat flux. The result is a localised cold anomaly relative to the surrounding dry embankment material, which can reach 3–6 °C in arid environments with high evaporative demand.
The constraint is cloud cover and the 16-day Landsat repeat. In humid tropical climates, clear-sky acquisitions over a specific site may occur only a handful of times per year. Thermal anomaly detection is therefore most reliable in semi-arid and arid mining regions, where many of the world's large tailings facilities happen to be located. Acquisition timing within the diurnal cycle also matters: the thermal contrast between wet and dry surfaces peaks in mid-afternoon, which is broadly consistent with Landsat's descending-node overpass time of approximately 10:00–10:30 local solar time, though the contrast is somewhat reduced compared to a 14:00 overpass. This is a real limitation that the analysis must acknowledge.
LST anomaly mapping is best used as a corroborating layer rather than a primary detection method. A persistent cold anomaly at the toe, coinciding with a SAR backscatter increase, substantially raises confidence that the signal is seepage-related rather than an artefact of surface material difference or shadow.
Separating seepage from everything else that makes soil wet
The central analytical challenge is specificity. Rainfall, runoff ponding, condensation, tailings supernatant pond overflow and irrigation from nearby land can all produce wet-soil signatures at or near the toe. Several discriminators help. First, spatial pattern: seepage typically produces a localised, elongated or diffuse wet patch on the downstream face or at the embankment toe, not a uniform brightening across the entire scene. Second, temporal persistence: rainfall-induced backscatter increases decay within days as soil drains; seepage-induced wetness persists or intensifies. Third, cross-sensor agreement: a signal that appears in both SAR backscatter and thermal LST anomaly, on multiple dates, is far more credible than a single-sensor single-date observation.
Precipitation data from ERA5 reanalysis or GPM IMERG should be ingested alongside the SAR and thermal time series to flag acquisitions made within 48–72 hours of significant rainfall. Those acquisitions are not discarded but are treated as potentially confounded and weighted accordingly in the anomaly scoring.
Practical limits and what this method cannot tell you
Sentinel-1's 10 m resolution means that a seepage patch smaller than roughly 20–30 m across may not produce a statistically detectable anomaly against the background of speckle noise, even after multi-look processing. Very small, early-stage seeps are below the detection floor. ICEYE in spotlight mode can reduce this to sub-1 m, but only when tasked; it is not a continuous monitoring layer.
Neither SAR backscatter nor thermal LST tells you the seepage flow rate, the internal pore-water pressure distribution, or whether piping has initiated. Those require piezometers, seepage collection weirs and geotechnical judgement. What satellite data provides is a spatial screening capability and a temporal record that is independent of site access constraints, which can be significant for remote or politically complex operations.
Vegetation on the downstream face complicates interpretation. Dense canopy attenuates C-band returns and masks the soil surface in thermal imagery. L-band penetrates moderate canopy, but dense tropical vegetation remains a challenge. Sites with established grass or shrub cover on the embankment face require additional processing steps, including vegetation index masking and canopy-corrected backscatter models.
Satellize applies this multi-sensor stack, combining open Sentinel-1 and Landsat archives with commercial ICEYE tasking on client licence, as part of its tailings monitoring service. The Tonga crop-estimation programme demonstrated the same underlying principle of multi-temporal anomaly detection on open constellations; the physics here is different, but the operational pipeline is closely related.
From pixel flags to a defensible monitoring record
Regulators and dam safety engineers need more than a colour-coded map. The output of a seepage-wetness monitoring programme should include a per-acquisition anomaly score for defined toe-zone polygons, a time-series chart showing the evolution of that score against precipitation and seasonal baselines, and a threshold-triggered alert when the score exceeds a pre-agreed level on two or more consecutive acquisitions. That alert should carry enough metadata (acquisition date, sensor, polarisation, confidence band) to allow a geotechnical engineer to make an informed decision about whether to dispatch an inspection team.
Archive depth matters for establishing baselines. Sentinel-1 data is available from 2014; Landsat 8 from 2013. A site with a five-year pre-alert archive has a credible seasonal baseline. A newly commissioned facility with six months of data does not, and the anomaly detection confidence should be communicated accordingly. Honest uncertainty quantification is not a weakness in the product; it is what distinguishes a defensible monitoring record from a false sense of security.
Typical figures
| SAR spatial resolution (Sentinel-1 IW) | 10 m range × 22 m azimuth (ground range detected); 5 × 20 m single-look complex |
| SAR spatial resolution (ALOS-2 PALSAR-2 Fine mode) | 3–10 m depending on mode; Ultra-Fine mode reaches ~3 m |
| SAR spatial resolution (ICEYE Spotlight) | Sub-1 m (0.5 m published for Spotlight XF mode) |
| Thermal resolution (Landsat 8/9 TIRS) | 100 m native; 30 m resampled product; LST absolute accuracy ~1–2 °C |
| Revisit (Sentinel-1, mid-latitudes) | 6 days (two-satellite constellation); 12 days single satellite |
| Revisit (Landsat 8 + 9 combined) | 8 days combined; 16 days per satellite; cloud cover reduces usable frequency |
| SAR frequency bands | C-band 5.4 GHz (Sentinel-1); L-band 1.2 GHz (PALSAR-2); X-band 9.65 GHz (ICEYE) |
| Minimum detectable wet patch (Sentinel-1) | Approximately 20–30 m across after multi-look processing; smaller patches below detection floor |
| Archive depth | Sentinel-1 from 2014; Landsat 8 from 2013; ALOS-2 from 2014 |
| Thermal anomaly detection threshold | LST depression of ~2–5 °C detectable under clear-sky conditions; cloud cover is a hard constraint |
Analytics Satellize can run
| SAR backscatter anomaly score, toe zone | Multi-temporal log-ratio change detection on Sentinel-1 IW VV/VH stack; per-pixel z-score against seasonal baseline; spatial clustering to suppress speckle artefacts | GIS polygon layer with per-acquisition anomaly score and confidence band; time-series chart exported as CSV and PDF |
| L-band vs C-band depth discrimination | Coincident PALSAR-2 and Sentinel-1 backscatter comparison; surface-only vs deep moisture classification using dual-frequency dielectric contrast logic | Classified raster (surface-moisture / deep-moisture / dry) for each available acquisition pair; interpretive note for geotechnical review |
| Thermal LST cold-anomaly map | Landsat 8/9 TIRS atmospheric correction (USGS Collection 2 LST product); anomaly detection relative to multi-year clear-sky median; precipitation-flagged acquisitions excluded from primary score | LST anomaly raster per clear-sky acquisition; toe-zone mean LST time series with ERA5 precipitation overlay |
| Multi-sensor seepage confidence index | Fusion of SAR backscatter anomaly score and thermal LST anomaly using a weighted evidence combination; persistence filter requiring signal on two or more consecutive acquisitions | Single composite confidence score per monitoring period; threshold-triggered alert (email or API push) when score exceeds agreed level |
| Precipitation-corrected baseline update | GPM IMERG or ERA5 precipitation ingestion; rolling baseline recalculation excluding acquisitions within 72 hours of rainfall events above a configurable threshold | Updated baseline rasters delivered quarterly; alert threshold recalibrated automatically |
| High-resolution confirmation tasking (ICEYE) | Alert-triggered ICEYE Spotlight tasking; sub-1 m backscatter image compared against Sentinel-1 flag location; spatial extent and morphology of wet patch characterised | Annotated ICEYE image with wet-patch boundary polygon; comparison report against Sentinel-1 flag |
| Historical baseline and anomaly archive | Full Sentinel-1 and Landsat archive processing from 2013/2014 to present; retrospective anomaly scoring to establish pre-monitoring baseline and identify historical seepage events | Multi-year anomaly time series; PDF report summarising historical seepage episodes and seasonal patterns |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.