Sea-ice lead detection and refreezing state classification
Linear fractures in pack ice transfer enormous heat to the atmosphere and set navigation risk. SAR polarimetry and altimeter waveforms can classify them by refreezing state, from open water through nilas to grey ice.
Sensors
- Sentinel-1 A/B (C-band SAR): Primary workhorse. Extra Wide Swath mode delivers 400 km coverage at 20 m resolution with HH+HV dual polarisation. Repeat pass over Arctic: 6 days per satellite, 3 days combined. HH backscatter is sensitive to surface roughness changes as a lead refreezes; HV cross-polarisation helps separate volume scatterers (rough multi-year ice) from specular surfaces (open water, nilas).
- CryoSat-2 SIRAL (Ku-band radar altimeter): Waveform peakiness, a dimensionless ratio of peak power to integrated power, spikes sharply over specular leads. ESA's Level-2 product uses this to retrieve freeboard, but the same peakiness signal identifies lead locations along the sub-satellite track at roughly 300 m along-track sampling. Revisit at high latitudes is around 30 days at the equator-equivalent but improves to near-daily above 88°N.
- RADARSAT Constellation Mission (C-band SAR): Three identical satellites achieve a 4-day exact repeat and daily Arctic coverage above 69°N. Compact polarimetry (RCB mode) provides pseudo-quad-pol decomposition at 50 m resolution over 350 km swaths, useful for separating ice types by scattering mechanism without full quad-pol data volume.
- TerraSAR-X / TanDEM-X (X-band SAR): Spot and stripmap modes reach 1-3 m resolution, resolving narrow leads (tens of metres wide) that are invisible at Sentinel-1 scale. X-band is more sensitive to thin ice surface texture than C-band, but the swath is small (30 km in stripmap) and tasking is commercial, so it is best used for targeted validation rather than synoptic mapping.
What a lead is, and why its age matters
A lead is a fracture in pack ice, typically linear, ranging from a few metres to several kilometres wide and potentially hundreds of kilometres long. When sea ice diverges under wind or current shear, the fracture exposes seawater at roughly minus 1.8 degrees Celsius to air that in Arctic winter can reach minus 30 or colder. The resulting heat flux can exceed 500 W per square metre, orders of magnitude above the flux through consolidated ice. That makes leads the dominant local source of atmosphere-ocean heat and moisture exchange in polar winter, with consequences that ripple into weather forecasting and climate modelling.
A freshly opened lead does not stay open. Within hours, a skin of frazil ice forms, thickening through nilas (a few centimetres, elastic, dark) into grey ice (10-15 cm, less flexible, lighter) and eventually grey-white ice above 15 cm. Each stage has a distinct radar signature. The classification problem is therefore not binary: it is a progression, and the age of the ice within the lead determines its navigational risk, its heat flux, and its relevance to atmospheric models.
How SAR backscatter reads the refreezing sequence
Open water in a lead returns almost no C-band backscatter in calm conditions: the surface is specular, and energy scatters away from the sensor. HH-polarised backscatter values typically fall below minus 20 dB in EW mode, creating a dark linear feature against the brighter surrounding ice. Wind roughens the water surface and raises backscatter, which is the main source of ambiguity. A windy lead can resemble young ice; a calm multi-year ice pond can resemble a lead. Wind speed records from ERA5 reanalysis or coincident scatterometer passes are a necessary ancillary input.
As nilas forms, surface tension suppresses capillary waves. Backscatter stays low but begins to rise slightly. Grey ice introduces a thin brine-wetted layer that absorbs rather than scatters, keeping HH low while HV cross-polarisation remains near the noise floor. Multi-year ice, by contrast, has a rough surface and bubble-rich interior that generates strong volume scattering in both polarisations. The ratio HH/HV therefore separates consolidated multi-year ice from young lead ice quite cleanly, though first-year ice in intermediate states overlaps both populations. Polarimetric decomposition methods, particularly the entropy-alpha decomposition applied to compact or quad-pol data from RADARSAT Constellation Mission, add a third axis that reduces this ambiguity.
Texture metrics matter as much as mean backscatter. Leads are geometrically linear and internally homogeneous relative to surrounding floes. Applying the Grey-Level Co-occurrence Matrix (GLCM) to Sentinel-1 imagery extracts local homogeneity and contrast features that improve classification accuracy in published studies, particularly for narrow leads below 200 m width where mixed pixels dominate.
CryoSat-2 waveform peakiness: a second opinion along the track
Radar altimeters were designed for sea-surface height, not ice classification. But a specular lead beneath the CryoSat-2 footprint returns a waveform with a sharp, narrow peak, quite unlike the broad ramp-shaped return from rough ice. ESA and several academic groups have used this peakiness signal to build lead-fraction products along the satellite track, which in turn feed into freeboard and thickness retrievals. For lead detection specifically, the altimeter provides an independent, physics-based confirmation of SAR-identified features, and it works in all weather because Ku-band penetrates cloud.
The limitation is geometry. CryoSat-2 gives a one-dimensional curtain of observations, not an image. A lead that runs perpendicular to the ground track may be missed entirely; one that runs parallel is well sampled. Combining SAR imagery (which gives spatial context) with altimeter peakiness (which gives a calibrated specular signal) is more reliable than either alone.
Operational limits the buyer should know
Cloud is not a problem for SAR or altimetry. Polar darkness is not a problem. What does degrade performance is wind. Above roughly 7-8 m per second, open-water backscatter rises into the range typical of young ice, and classification confidence drops. Any operational product should flag pixels where coincident wind speed exceeds this threshold.
Narrow leads, below 50-100 m, are systematically underdetected by Sentinel-1 EW mode at 20 m pixel spacing, because a single pixel rarely falls entirely within the lead. TerraSAR-X can resolve leads down to perhaps 20-30 m, but coverage is sparse and tasking adds latency. The practical minimum detectable lead width for a synoptic Sentinel-1 product is around 100-200 m. For navigation planning in heavily fractured ice, that resolution gap is significant.
Refreezing-state classification carries its own uncertainty. Published confusion matrices for SAR-based ice-type classification typically show 80-90% overall accuracy when validated against coincident ground truth or airborne data, but accuracy for the transitional nilas and grey-ice classes is lower, often 70-80%, because these classes are spectrally and texturally similar at C-band. Reporting classification confidence alongside the class label is not optional; it is the only honest way to present the product.
From pixels to operational intelligence
The analytic chain runs from SAR ingestion through speckle filtering, polarimetric feature extraction, supervised classification (typically a random forest or support vector machine trained on labelled Arctic scenes), and geometric post-processing to enforce linearity constraints on lead objects. The output is a vector layer of lead polygons, each attributed with estimated ice class, width, length, orientation, and a confidence score. Temporal stacking of consecutive passes reveals refreezing rate, which is a proxy for local heat flux.
Satellize runs this chain on open Sentinel-1 EW acquisitions and can add commercial SAR tasking under client licence for higher-resolution validation. The approach is the same one applied in the company's broader analytics work, including the Kingdom of Tonga crop-estimation programme, adapted to polar backscatter physics. Outputs are delivered as GeoPackage or GeoTIFF layers compatible with standard GIS environments, with an optional alert feed for leads that open or widen above a threshold in a defined area of interest.
Typical figures
| Spatial resolution (synoptic) | 20 m (Sentinel-1 EW mode); 50 m (RADARSAT-C compact pol) |
| Spatial resolution (targeted) | 1-3 m (TerraSAR-X stripmap/spotlight, commercial tasking) |
| Minimum detectable lead width | ~100-200 m for Sentinel-1 EW; ~20-30 m for TerraSAR-X |
| Revisit (Arctic, combined Sentinel-1 A+B) | 3 days at mid-latitudes; near-daily above ~75°N due to orbit convergence |
| Revisit (RADARSAT Constellation Mission) | Daily coverage above 69°N; 4-day exact repeat |
| Sensor frequency / band | C-band 5.405 GHz (Sentinel-1, RADARSAT); X-band 9.65 GHz (TerraSAR-X); Ku-band 13.575 GHz (CryoSat-2 SIRAL) |
| Polarisation modes used | HH+HV dual-pol (Sentinel-1); compact pol RCB (RADARSAT-C); HH (CryoSat-2 altimeter) |
| Latency (Sentinel-1 NRT) | 3-24 hours from acquisition to Level-1 GRD product via Copernicus Data Space |
| Archive depth | Sentinel-1A from April 2014; CryoSat-2 from 2010; RADARSAT-2 heritage from 2007 |
| Classification accuracy (published range) | 80-90% overall; 70-80% for transitional ice classes (nilas, grey ice) |
Analytics Satellize can run
| Lead detection and mapping layer | Sentinel-1 EW HH+HV backscatter thresholding with GLCM texture features and geometric linearity filtering | GeoPackage polygon layer of detected leads, attributed with width, length, orientation and detection confidence; updated per satellite pass |
| Refreezing-state classification | Supervised random forest classifier trained on HH backscatter, HH/HV ratio and GLCM homogeneity; classes: open water, nilas, grey ice, grey-white ice | Raster classification grid (GeoTIFF) with per-pixel class label and confidence score; wind-speed mask flagging uncertain pixels above 7 m/s |
| Lead fraction time series | Temporal stacking of per-pass classification layers; lead area computed per grid cell over rolling 7- and 30-day windows | CSV and NetCDF time-series table per area of interest; suitable for input to atmosphere-ocean heat-flux models |
| CryoSat-2 peakiness lead confirmation | Waveform peakiness index computed from ESA Level-1b SIRAL data; specular returns co-located with SAR lead polygons to validate or flag disagreements | Along-track point dataset (GeoJSON) with peakiness value and SAR-class agreement flag |
| Lead opening and refreezing alert | Change detection between consecutive SAR passes; new lead polygons or class transitions above a configurable area threshold trigger alert | Near-real-time alert feed (JSON webhook or email) with lead coordinates, estimated width, and ice-class transition |
| High-resolution lead validation (on request) | TerraSAR-X stripmap or spotlight tasking over selected leads; 1-3 m imagery used to validate Sentinel-1 classification and resolve sub-100 m features | Annotated TerraSAR-X scene with lead boundaries and width measurements; comparison report against Sentinel-1 product |
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.