Wind-turbine blade icing detection from passive microwave and optical data
Blade icing is a leading cause of unplanned wind-farm downtime in cold climates. Satellite SAR and passive microwave data can flag icing conditions before production losses compound, though resolution limits mean site-level interpretation requires careful method layering.
Sensors
- Sentinel-1 SAR (C-band, 5.4 GHz): Interferometric Wide Swath mode at 10 m ground range resolution, 250 km swath, 6-day repeat at mid-latitudes (12-day for a single satellite). Ice accretion on blade surfaces changes dielectric constant and surface roughness, producing measurable backscatter anomalies in VV and VH polarisations. The primary tool for localising icing events to individual turbine clusters.
- AMSR2 passive microwave (JAXA GCOM-W): Brightness temperature observations at 6.9 to 89 GHz, nominal spatial resolution 25 km at lower frequencies. Sensitive to integrated liquid water and ice in the atmospheric column and at the land surface. Coarse footprint precludes turbine-level attribution but provides regional icing-condition flags useful for screening large areas.
- MODIS Terra/Aqua Land Surface Temperature: LST product at 1 km resolution, twice-daily overpass per satellite. Near-freezing surface temperatures corroborate icing risk windows identified from SAR or passive microwave. Cloud cover is a significant limitation: MODIS LST is unavailable under cloud, which frequently accompanies icing events.
- Sentinel-3 SLSTR: Dual-view thermal infrared and shortwave infrared sensor, 1 km nadir resolution, daily global coverage. Provides surface temperature context at higher spatial resolution than MODIS and with a longer planned mission lifetime. Subject to the same cloud-masking constraints as any thermal sensor.
Why ice changes what a radar sees
A dry fibreglass blade presents a relatively stable radar cross-section to a C-band SAR. Ice changes that. Liquid water has a real dielectric constant near 80 at microwave frequencies; pure ice sits around 3.15. The transition matters because rime ice, the most common accretion type on wind turbines, forms a rough, irregular surface that increases volume scattering and alters the VH-to-VV backscatter ratio in ways that differ from the dry-blade baseline.
Wet ice, or the mixed-phase condition that precedes full accretion, produces an even stronger dielectric contrast and can cause a backscatter increase of several decibels relative to clear conditions. Glaze ice, which is smoother, may actually reduce backscatter in some geometries by specular deflection away from the sensor. Neither effect is universal: wind direction relative to the SAR look angle, turbine yaw position at acquisition time, and blade rotation all introduce variability. This is not a clean, single-threshold problem.
What passive microwave adds, and where it falls short
AMSR2's brightness temperature channels at 18.7 and 36.5 GHz are sensitive to surface emissivity changes associated with snow and ice cover. At regional scale, anomalies in the polarisation difference (vertical minus horizontal brightness temperature) correlate with ice formation on land surfaces. Several published studies have used this signal to map frost and icing events across Scandinavian wind-resource zones.
The honest constraint is the 25 km footprint. A wind farm occupying a few square kilometres sits inside a single AMSR2 pixel alongside forests, fields, and roads, each contributing their own emission signature. Passive microwave is therefore a regional condition flag, not a farm-level diagnosis. Its value is in prioritising which farms to examine with SAR, not in replacing that examination.
Layering the sensors: from regional flag to cluster-level alert
A practical detection workflow runs in three stages. First, AMSR2 brightness temperature anomalies and MODIS or Sentinel-3 LST products flag grid cells where surface temperatures are at or below freezing and atmospheric moisture conditions are consistent with icing. This step has low latency, typically available within a few hours of overpass, but high false-alarm potential.
Second, the flagged area is examined in the next available Sentinel-1 acquisition. Change detection against a multi-temporal baseline stack isolates turbine cluster pixels showing anomalous backscatter. The 10 m resolution is sufficient to distinguish turbine-containing pixels from surrounding terrain, though individual blades at 50 to 100 m length are not resolved. Third, the SAR anomaly is cross-referenced against SCADA production data where available, or against the thermal record, to assess whether the backscatter shift coincides with a generation shortfall. Without SCADA, the satellite signal alone cannot confirm production loss.
Honest limits of the detection chain
Cloud cover is the most persistent obstacle. Icing events are by definition associated with cold, moist air masses, which frequently produce cloud. MODIS and Sentinel-3 LST retrievals fail under cloud; SAR is cloud-transparent, which is precisely why it anchors the detection chain. But even SAR has limits: wet snow on the ground surrounding a turbine can dominate the backscatter signal and mask the blade-specific anomaly.
Revisit is the second constraint. A 6-day Sentinel-1 repeat means a short icing event can begin and end between acquisitions. Commercial SAR constellations offer more frequent revisit but at higher cost and with different calibration baselines. Passive microwave revisits daily, but its resolution means it catches the meteorological context, not the event itself. For operational icing management, satellite data is best framed as a climatological and condition-monitoring tool rather than a real-time alarm system. The realistic lead time for a satellite-derived icing alert is hours to a day, not minutes.
Building a site-specific icing climatology
The Sentinel-1 archive extends back to 2014 for European sites. Processing multi-year backscatter time series against coincident AMSR2 and ERA5 reanalysis temperature fields allows construction of a site-specific icing frequency map: how many days per year, in which months, and under which synoptic conditions does backscatter deviate from the dry-blade baseline. This kind of climatology is more actionable for operations planning than any single event detection.
Satellize runs this class of multi-sensor time-series analysis on open constellations, the same approach that underpins its crop-estimation work in Tonga, adapted here to the backscatter domain. The output is a seasonal icing-risk calendar per turbine cluster, with associated uncertainty bounds derived from the variability in the historical record. Operators can use it to schedule de-icing maintenance windows and to inform insurance and revenue forecasting.
What the data cannot tell you, and what it can
Satellite observation cannot measure ice thickness on a blade, cannot distinguish rime from glaze at the resolution available, and cannot predict icing onset with the precision of a blade-mounted sensor or a numerical weather model. These are not failures of the satellite method; they are physical limits of remote sensing at current resolutions.
What satellite data does provide is spatial coverage across an entire wind-farm portfolio simultaneously, with no maintenance burden, no sensor failure risk at the turbine, and a historical archive that predates most operational monitoring systems. For a developer assessing a new site, or an operator trying to understand why one cluster consistently underperforms in February, the satellite record is often the only continuous, spatially consistent source available. That is a specific and defensible use case, even if it is not a complete one.
Typical figures
| SAR spatial resolution (Sentinel-1 IW mode) | 10 m ground range, 20 m azimuth (multi-looked to ~20 m) |
| Passive microwave spatial resolution (AMSR2) | 25 km at 18.7 GHz; 10 km at 89 GHz |
| Sentinel-1 revisit (mid-latitudes, two-satellite constellation) | 6 days; degrades to 12 days for single satellite |
| AMSR2 revisit | Daily (ascending and descending passes) |
| MODIS LST resolution and revisit | 1 km; twice daily per satellite (Terra + Aqua) |
| Sentinel-3 SLSTR resolution and revisit | 1 km nadir; approximately daily global coverage |
| Minimum detectable backscatter anomaly (Sentinel-1) | Typically 1.5 to 3 dB above baseline noise floor; site-dependent |
| SAR archive depth (European sites) | Sentinel-1A from April 2014; Sentinel-1B 2016 to 2021 |
| Alert latency (SAR-based) | Hours to 1 day post-acquisition, depending on processing pipeline |
| Delivery formats | GeoTIFF anomaly maps, GeoJSON cluster alerts, CSV seasonal climatology tables |
Analytics Satellize can run
| Backscatter anomaly detection | Multi-temporal SAR change detection against site-specific baseline stack (log-ratio or z-score thresholding) | Per-acquisition GeoTIFF flagging turbine-cluster pixels with anomalous C-band backscatter, with confidence score |
| Regional icing-condition flag | AMSR2 polarisation difference anomaly combined with MODIS/Sentinel-3 LST below-zero masking | Daily gridded alert layer (25 km) indicating elevated icing probability, delivered as GeoJSON or raster tile |
| Site-specific icing climatology | Multi-year Sentinel-1 backscatter time series cross-referenced with ERA5 temperature and humidity reanalysis | Monthly icing-frequency maps per turbine cluster, with interannual variability bounds, as PDF report and GIS layer |
| Icing-event duration estimation | Sequential SAR acquisition differencing to bracket event onset and clearance within revisit window | Event log CSV with estimated start/end date range and affected cluster IDs |
| Production-loss correlation analysis | Satellite-derived icing flags matched against operator-supplied SCADA generation data using temporal regression | Correlation report quantifying fraction of generation shortfall events coincident with satellite-detected icing conditions |
| Seasonal maintenance scheduling input | Climatological icing-frequency ranking of turbine clusters by month, derived from archive analysis | Ranked cluster list with recommended inspection windows, formatted as operator-ready spreadsheet |
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.