Crop lodging detection using SAR and optical change analysis
When cereal crops collapse from wind, rain or pest damage, conventional yield models keep estimating as if nothing happened. SAR backscatter and optical texture change sharply at lodging, giving a two-to-five-day detection window before the damage compounds.
Sensors
- Sentinel-1 SAR (C-band, 5.405 GHz): 10 m ground range detected resolution in IW mode, 6-day revisit at mid-latitudes with both satellites. VV polarisation responds to surface roughness; VH responds to volume scattering within the canopy. The VV/VH ratio drops sharply when an upright canopy collapses because volume scattering is suppressed. This ratio change is the primary lodging signal.
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit with both satellites. Lodged crop shows a transient greening anomaly (increased green reflectance, depressed NIR) before chlorophyll degradation begins. Texture metrics computed on Band 8 (NIR) at 10 m are sensitive to the loss of row structure. Cloud cover is the binding constraint; a clear acquisition within the lodging window is not guaranteed.
- Landsat 8/9 OLI: 30 m resolution, 8-day revisit per satellite (16-day per platform, combined ~8-day). Useful for historical archive analysis and cross-validation of Sentinel-2 detections, particularly where Sentinel-2 is cloud-obscured and a Landsat pass happened to be clear. Spatial resolution is marginal for detecting small lodged patches in fragmented fields.
- Sentinel-1 InSAR coherence: Six-day coherence pairs lose coherence over lodged areas because the canopy geometry changes between passes. This provides a secondary confirmation signal independent of the backscatter ratio, though it requires careful baseline selection and is most interpretable on flat terrain.
What a flattened canopy does to a radar signal
An upright cereal canopy, wheat or barley at grain-fill stage, scatters C-band radar in two distinct ways. VV polarisation bounces off the soil surface and the vertical stems, returning a strong signal. VH polarisation is generated by volume scattering inside the canopy, where the random orientation of leaves and ears depolarises the wave. The ratio of VV to VH is therefore a rough proxy for how much vertical structure exists.
When lodging occurs, stems fold to near-horizontal. Volume scattering collapses. VH return drops relative to VV, compressing the ratio. Studies using Sentinel-1 IW data over European wheat fields have documented ratio changes of 2 to 4 dB at lodged pixels relative to standing neighbours acquired in the same pass. That is a large, interpretable signal. The complication is that rain itself temporarily suppresses VH by wetting the canopy, so a single-date anomaly must be confirmed on a subsequent dry-condition pass, typically within six days.
The optical side: a brief greening before the yellowing
Optical sensors catch lodging through two separate mechanisms. First, when stems collapse, the upper leaf surface faces skyward rather than sideways, momentarily increasing the green reflectance and altering the NDVI in a direction opposite to stress. This counterintuitive greening lasts days to perhaps two weeks before chlorophyll degradation and soil exposure push NDVI downward. Catching this window requires a cloud-free Sentinel-2 acquisition shortly after the event.
Second, lodging destroys the regular row texture that characterises a healthy standing crop. Texture metrics computed on the NIR band, such as the grey-level co-occurrence matrix contrast or homogeneity, change measurably in lodged patches at 10 m resolution. This is useful precisely because texture does not depend on absolute reflectance calibration and is less sensitive to illumination angle variation across a scene.
Neither optical indicator is reliable in isolation. Cloud is the obvious enemy. A prolonged overcast after a lodging event can mean the greening window passes unobserved, leaving only the post-yellowing signal, which is harder to distinguish from drought stress or disease. This is why SAR is the primary detection layer and optical the confirmation layer, not the reverse.
Building a change-detection pipeline that does not cry wolf
The practical workflow begins with a dense SAR time series from sowing to harvest, typically 15 to 25 Sentinel-1 acquisitions per season at mid-latitudes. A per-pixel baseline is established for the VV/VH ratio during the vegetative phase. Anomalous drops after heading, when lodging risk peaks, are flagged against this baseline using a threshold derived from the local standard deviation rather than a fixed dB value. Fixed thresholds fail because backscatter varies with soil moisture, incidence angle and crop variety.
False positives come from several sources: field operations such as rolling or spraying that temporarily flatten the crop, irrigation events that wet the canopy, and wind-induced motion during the SAR acquisition itself. Temporal consistency filtering, requiring the anomaly to persist across at least two consecutive passes, eliminates most transient events. Spatial filtering removes isolated single pixels, since real lodging patches are rarely smaller than a few hundred square metres.
The minimum detectable lodged area is not a fixed number. At 10 m Sentinel-1 resolution, patches of roughly 0.1 hectares are detectable in principle, but the false-alarm rate rises sharply below about 0.5 hectares because speckle noise at that scale competes with the signal. For national-scale production estimates, this is acceptable. For farm-level insurance verification, it may not be.
Why yield models need this information and what happens without it
Most operational crop yield models, whether empirical or process-based, are calibrated on standing-crop observations. They ingest NDVI or LAI time series and assume the canopy geometry is upright throughout the season. Lodged areas violate this assumption silently. The model continues to predict normal development from a canopy that has, in effect, stopped functioning normally.
The consequence for national production estimates is systematic overestimation in years with widespread lodging. The 2007 UK wheat harvest, for example, saw significant lodging losses that were poorly captured by area-based estimates until ground surveys were completed. Satellite-derived lodging masks, ingested as a correction layer, allow yield models to flag affected pixels and apply a lodging-specific yield penalty drawn from agronomic literature rather than pretending the crop is fine.
For farm managers, the value is more immediate. Lodged areas are prone to fungal disease, particularly Fusarium, because air circulation is reduced and humidity rises inside the collapsed canopy. Early identification allows targeted fungicide application or, in severe cases, early harvest decisions before quality deteriorates further.
Honest limits and where the method breaks down
Cloud cover over temperate agricultural regions, where lodging risk is highest, is a genuine operational constraint. In a wet summer in northern Europe, there may be only two or three usable Sentinel-2 acquisitions during the critical six-week window from heading to harvest. SAR is cloud-independent, but even SAR has limits: heavy rainfall during acquisition degrades the VV/VH signal, and the six-day revisit means a lodging event can be between 0 and 6 days old before the next pass.
Variety and growth stage matter. Shorter, stiffer modern wheat varieties lodge less and produce weaker signals when they do. Maize lodging, which involves stem breakage rather than stem bending, has a different backscatter signature and is less well characterised in the published literature than small-grain cereal lodging.
Finally, distinguishing partial lodging from full lodging is difficult at 10 m resolution. A pixel flagged as lodged may contain a mixture of standing and collapsed stems, and the backscatter change will be proportional to the lodged fraction in a non-linear way. Quantifying lodged fraction within a pixel requires either higher-resolution commercial SAR or ground truth from field surveys.
Satellize applies this detection pipeline within its crop analytics work, including the methodological framework developed for the Kingdom of Tonga crop-estimation programme, and can configure field-boundary-aware lodging alerts for national agriculture ministries or commodity traders requiring sub-national production intelligence.
Typical figures
| Primary SAR resolution | 10 m (Sentinel-1 IW mode, ground range detected) |
| SAR revisit at mid-latitudes | 6 days (Sentinel-1A + 1B combined; single-satellite 12 days) |
| Optical resolution | 10 m visible/NIR (Sentinel-2); 30 m (Landsat 8/9) |
| Optical revisit | 5 days (Sentinel-2A + 2B combined); ~8 days (Landsat 8 + 9 combined) |
| Primary SAR bands used | C-band VV and VH polarisation; VV/VH ratio as primary lodging index |
| Optical bands used | Sentinel-2 Band 3 (green, 560 nm), Band 4 (red, 665 nm), Band 8 (NIR, 842 nm) |
| Minimum detectable lodged patch | ~0.5 ha at acceptable false-alarm rates; ~0.1 ha in principle with speckle filtering |
| Detection latency after event | 0 to 6 days (SAR revisit limited); optical confirmation subject to cloud clearance |
| SAR archive depth | Sentinel-1 from 2014; Landsat optical from 1972 (USGS archive) |
| Delivery formats | GeoTIFF lodging-probability raster, field-boundary polygon layer with lodged-fraction attribute, CSV alert feed |
Analytics Satellize can run
| SAR ratio anomaly map | Per-pixel VV/VH baseline subtraction with temporal consistency filtering across consecutive Sentinel-1 passes | GeoTIFF raster of VV/VH anomaly magnitude, updated each Sentinel-1 acquisition |
| Binary lodging mask | Threshold applied to SAR anomaly map, spatially filtered to remove sub-0.5 ha patches; confirmed against optical texture change where cloud-free imagery is available | Polygon GIS layer of lodged field zones with detection date and confidence class |
| Lodged-area fraction by field | Zonal statistics of lodging mask intersected with field boundary layer; lodged fraction expressed as percentage of field area | CSV or GeoPackage table joinable to farm or cadastral identifiers |
| Optical texture change layer | GLCM contrast and homogeneity computed on Sentinel-2 NIR band before and after lodging event; change image thresholded and compared to SAR mask | GeoTIFF confirmation layer flagging pixels where both SAR and optical signals agree |
| Seasonal lodging timeline | Time-series stack of per-field lodged-fraction values from heading to harvest, assembled from all available SAR passes | Per-field time-series chart and CSV, suitable for ingestion into yield-model correction workflows |
| Sub-national production-impact estimate | Lodged-area mask multiplied by crop-type layer and agronomic yield-penalty coefficients drawn from published literature; aggregated to administrative unit | Tabular report of estimated production loss by district, with uncertainty range |
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.