SAR sea ice type classification and drift tracking
Synthetic aperture radar separates first-year ice, multi-year ice and open water by exploiting differences in volume scattering and surface texture, then tracks floe drift through repeat-pass image correlation. Summer melt complicates every classification scheme; knowing why matters as much as knowing the method.
Sensors
- Sentinel-1 EW (Extra-Wide Swath) mode: C-band (5.405 GHz), dual-polarisation (HH+HV). EW mode covers a 400 km swath at 20 m x 40 m ground resolution (multi-looked to roughly 40 m pixel spacing in operational products). Six-day exact repeat at mid-latitudes; near-daily coverage at high Arctic latitudes due to orbit geometry. Free and open data via Copernicus. The workhorse for operational sea ice charting.
- ICEYE SAR constellation: X-band (9.65 GHz), sub-metre to 3 m resolution in spotlight and strip modes. Revisit of specific Arctic locations can be arranged within hours through commercial tasking. Higher frequency means shallower penetration than C-band; useful for resolving individual floe edges and pressure ridges at scales Sentinel-1 EW cannot.
- RADARSAT-2 (heritage, still operational): C-band, multiple beam modes including ScanSAR Wide (500 km swath, 100 m resolution) and Fine Quad-pol (25 km swath, 8 m). Quad-polarimetric data supports decomposition methods (e.g. Freeman-Durden) that improve multi-year ice discrimination. Operated by MDA; data available commercially. Provided the backbone of many published Arctic ice-type classification studies.
- ALOS-2 PALSAR-2: L-band (1.2 GHz). Longer wavelength penetrates deeper into ice volume, increasing sensitivity to internal brine drainage structures in multi-year ice. ScanSAR mode delivers 350 km swath at 100 m resolution; stripmap at 3 m. JAXA operates the satellite; data available through research agreements. Particularly useful in distinguishing heavily deformed first-year ice from true multi-year ice, an ambiguity that C-band alone handles poorly.
Why radar sees ice differently depending on how old it is
Sea ice physics drives everything. First-year ice (FYI) forms in a single winter season. Its brine content remains high and its surface is relatively smooth at the scale of a C-band wavelength (roughly 5.6 cm). The result is specular or near-specular scattering: most of the radar energy bounces away from the sensor, producing relatively low backscatter in HH polarisation over calm, undeformed FYI.
Multi-year ice (MYI) has survived at least one summer melt. Brine drains out during that melt, leaving a network of air-filled voids and irregular internal structure. At C-band, this internal roughness produces volume scattering: the radar signal penetrates the ice surface and bounces repeatedly inside before returning to the sensor, giving MYI a characteristically higher and more spatially variable backscatter than FYI. In HV cross-polarisation the contrast is even more pronounced, because volume scattering depolarises the return signal. This is the physical basis for every operational ice-type classification algorithm.
Texture metrics and the limits of backscatter alone
Raw backscatter intensity separates the extremes well: open water (very low return in calm conditions) versus thick MYI (high, variable return). The problem sits in the middle. Heavily deformed or ridged FYI develops surface roughness that raises its backscatter toward MYI levels. Thin new ice (nilas, frazil) can appear almost as dark as open water. Classifiers that rely on intensity alone misidentify these categories at rates that matter operationally.
Texture features derived from the Grey-Level Co-occurrence Matrix (GLCM), including contrast, homogeneity, entropy and correlation computed over local windows of 15 to 50 pixels, add discriminating power. MYI tends to show high entropy and low homogeneity; smooth FYI shows the inverse. Combining dual-polarisation backscatter with GLCM texture in a supervised classifier (random forest, support vector machine or convolutional neural network trained on ice-chart labels) typically reaches overall accuracies of 80 to 90 percent in winter conditions on published benchmark datasets, with the largest confusion between deformed FYI and MYI.
L-band data from ALOS-2 PALSAR-2 helps resolve that specific confusion. The longer wavelength penetrates further into the ice column, making the internal void structure of true MYI more visible and reducing the false-positive MYI rate that plagues C-band classifiers in regions with heavy FYI deformation.
Summer melt: the season when everything becomes ambiguous
In late spring and summer, melt ponds form on the ice surface and eventually flood across it. A flooded pond surface looks almost identical to open water in SAR backscatter. MYI covered in melt ponds loses its characteristic high return. The volume-scattering signature that distinguishes MYI in winter is suppressed when liquid water fills the surface voids. Published studies using RADARSAT-2 data have documented a near-collapse of C-band MYI discrimination during the July to August melt period.
This is not a solvable problem with better algorithms alone. It is a physical ambiguity. Operational ice services, including the Canadian Ice Service and the Norwegian Meteorological Institute, address it by fusing SAR with passive microwave data (AMSR2, SSMIS) during summer, accepting lower classification confidence, or restricting ice-type products to winter months. Any client expecting year-round, high-confidence ice-type maps from SAR alone should understand this limit before commissioning a service.
Drift vectors from repeat-pass image pairs
Ice drift tracking uses a different part of the SAR signal chain: not backscatter intensity values but the spatial pattern of features across two images acquired days apart. The standard method is pattern matching, either normalised cross-correlation or, for coherent image pairs, phase-based offset tracking. A grid of search windows is placed across the first image; each window is matched to its best-fit location in the second image. The displacement vector, divided by the time interval, gives ice drift speed and direction.
Sentinel-1 EW mode, with its 400 km swath and near-daily Arctic revisit, is well suited to this. Published operational products from the Copernicus Marine Service use three-day and six-day Sentinel-1 pairs to generate drift vector fields at roughly 62.5 km grid spacing, with displacement uncertainties on the order of 1 to 3 km depending on ice conditions and image quality. For tracking individual floes rather than the large-scale drift field, higher-resolution tasking from ICEYE or RADARSAT-2 Fine mode reduces the positional uncertainty to tens of metres over a 24-hour interval.
The method fails when ice deforms rapidly between passes, when the surface pattern changes enough that no reliable match exists, or when a lead opens and refreezes, replacing the trackable texture entirely. Drift products always carry a confidence flag; gaps in the vector field are informative, not missing data.
What the products look like in practice
Operational outputs fall into two categories. Ice-type classification maps are raster or vector polygon products labelled with WMO sea ice nomenclature categories: open water, new ice, young ice, first-year ice (thin, medium, thick), multi-year ice and deformed or ridged variants. They are typically delivered as GeoTIFF or shapefile layers compatible with nautical chart systems. Confidence layers accompany them.
Drift vector fields are delivered as gridded NetCDF files or as GIS point or arrow layers, with each vector carrying a displacement magnitude, direction, time interval and quality flag. Downstream uses include routing optimisation for Arctic shipping, forcing fields for sea ice forecast models, and monitoring of ice export through straits such as Fram Strait, where the rate of MYI export to the North Atlantic has direct implications for the summer ice extent.
Satellize can run classification and drift pipelines on open Sentinel-1 EW acquisitions and add commercial ICEYE tasking for clients needing finer spatial detail on specific routes or infrastructure zones. The analytics architecture mirrors the approach used in our Tonga crop-estimation programme: open constellation data as the baseline, commercial tasking where resolution requirements exceed what open data can supply.
Honest assessment of what SAR ice products cannot do
SAR does not measure ice thickness directly. Backscatter correlates with ice type and age, not with thickness in metres. Thickness requires freeboard measurement from altimetry (covered separately in the ICESat-2 freeboard page) combined with snow-load assumptions. A classification map that says 'multi-year ice' tells you the ice is old and likely thick; it does not tell you whether it is 1.5 m or 4 m thick.
Cloud cover is irrelevant to SAR, which is one of its genuine advantages over optical sensors in the Arctic. But SAR has its own atmospheric complication: heavy precipitation can increase surface roughness on open water, mimicking ice returns. Wind roughening of open water is the better-documented problem, producing backscatter values that overlap with thin ice. Dual-polarisation ratios (HV/HH) help, but the ambiguity persists at wind speeds above roughly 10 to 12 m/s. Operational ice analysts know to flag these conditions; automated classifiers without wind-speed ancillary data will make errors in them.
Typical figures
| Primary frequency | C-band (5.405 GHz, Sentinel-1 / RADARSAT-2); X-band (9.65 GHz, ICEYE); L-band (1.2 GHz, ALOS-2 PALSAR-2) |
| Swath width (operational Arctic) | 400 km (Sentinel-1 EW); 500 km (RADARSAT-2 ScanSAR Wide); 350 km (ALOS-2 ScanSAR) |
| Spatial resolution (ice-type classification) | 40 m pixel spacing typical for Sentinel-1 EW products; 100 m for wide-swath RADARSAT-2 and ALOS-2 ScanSAR; sub-3 m for ICEYE spotlight |
| Revisit at high Arctic latitudes | Near-daily for Sentinel-1 EW above 70°N due to orbit convergence; ICEYE commercial tasking within hours for specific locations |
| Drift vector grid spacing | 62.5 km (Copernicus Marine Service operational product); 5 to 10 km achievable with high-resolution image pairs |
| Drift displacement uncertainty | 1 to 3 km over a 3-day interval (Sentinel-1 EW); tens of metres over 24 hours with fine-mode commercial SAR |
| Classification accuracy (winter conditions) | 80 to 90% overall for FYI/MYI/open water; degrades significantly during summer melt season |
| Archive depth | Sentinel-1 from 2014; RADARSAT-2 from 2007; ALOS-2 from 2014; ICEYE from 2018 |
| Polarisation modes used | HH+HV dual-pol (standard for ice); quad-pol (HH, HV, VH, VV) for decomposition methods on RADARSAT-2 |
| Delivery formats | GeoTIFF (classification raster), shapefile/GeoJSON (polygon ice charts), NetCDF (drift vector fields), confidence raster layers |
Analytics Satellize can run
| Winter ice-type classification map | Dual-polarisation backscatter (HH, HV) combined with GLCM texture features; supervised classifier (random forest or CNN) trained on operational ice-chart labels | GeoTIFF raster and polygon shapefile with WMO ice categories and per-pixel confidence layer, delivered within 6 hours of scene acquisition |
| Summer ice-type map with melt-pond flag | C-band backscatter fused with passive microwave brightness temperature (AMSR2) to flag melt-affected pixels; reduced category set with explicit uncertainty band | GeoTIFF with four-class output (open water, FYI, MYI, melt-affected uncertain) and accompanying metadata report stating confidence degradation |
| Large-scale ice drift vector field | Normalised cross-correlation pattern matching on Sentinel-1 EW image pairs at 3-day or 6-day intervals; quality filtering by correlation coefficient threshold | NetCDF drift field on 62.5 km grid with speed, direction and quality flag; also available as GIS arrow layer for chart overlay |
| Individual floe drift track | High-resolution ICEYE or RADARSAT-2 Fine mode image pairs; sub-pixel offset tracking on manually or automatically identified floe features | Time-stamped position series for named floe polygons, delivered as GeoJSON with displacement vectors and positional uncertainty estimates |
| Ridged and deformed ice extent | Backscatter intensity thresholding combined with local variance and entropy metrics to isolate high-roughness zones within FYI fields | Polygon layer of deformed ice zones with area statistics, updated per available acquisition, formatted for integration with voyage-planning software |
| Fram Strait MYI export monitoring | Repeat classification of the Fram Strait corridor combined with drift vectors to estimate areal flux of MYI southward through the strait | Weekly tabular report of MYI area flux (km² per day) with time series chart; alert trigger if flux exceeds client-defined threshold |
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.