Sensors
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral including a dedicated SWIR stack. At this resolution, individual tunnel bays and frame shadows are resolved, allowing object-based delineation of structures as narrow as 4 m. Revisit roughly 1–4.5 days depending on latitude and tasking priority.
- Planet SuperDove: 3 m, 8-band multispectral, daily revisit globally. Adequate for mapping greenhouse clusters larger than roughly 10 m wide; frame-level detail is lost, but the high revisit makes it useful for change detection between underwriting cycles.
- Sentinel-1 C-band SAR: IW mode at 10 m ground range resolution, 6-day revisit (12-day single satellite). C-band penetrates cloud, which matters in humid growing regions. Metallic greenhouse frames produce a measurable double-bounce return, but at 10 m the signal mixes with surrounding clutter for small structures; best used for coherence-based change detection rather than first-pass delineation.
- COSMO-SkyMed X-band SAR: Spotlight mode reaches 1 m resolution; Stripmap HH at 3 m. X-band double-bounce from steel or aluminium greenhouse frames is stronger and more spatially confined than C-band, making it the preferred SAR mode for delineating individual structures in dense growing districts. Revisit can be scheduled to days with the commercial constellation.
What a floating roof gives away
A polyethylene greenhouse tunnel is, from above, a near-perfect radiometric anomaly. The film reflects strongly in the near-infrared while absorbing in the visible red, producing a spectral ratio that is the inverse of healthy vegetation. Glass greenhouses add specular glint in the shortwave infrared. Both materials sit on a background of soil or low vegetation, making them separable by a supervised classifier trained on even a modest set of manually labelled samples.
The structural regularity matters as much as the spectral signal. Greenhouse districts in Almería, the Nile Delta, or the Westland region of the Netherlands are laid out in near-orthogonal grids. Object-based image analysis (OBIA) exploits shape metrics, specifically elongation ratio, rectangularity and compactness, alongside spectral values, to discriminate tunnels from roads, water bodies or bare soil that might otherwise confuse a pixel-level classifier. At WorldView-3 resolution, individual bay widths of 6 to 8 m are resolved cleanly.
The resolution floor that medium-resolution sensors cannot clear
Sentinel-2 at 10 m and Landsat at 30 m are adequate for field-crop mapping at scale, but they fail greenhouse mapping in two ways. First, a standard polytunnel bay is 6 to 9 m wide; at 10 m resolution the tunnel occupies roughly one pixel, and mixed-pixel effects suppress the spectral contrast. Second, the geometric shape metrics that anchor OBIA classification disappear entirely at coarse resolution.
Published studies using Sentinel-2 in Almería, one of the world's most densely covered greenhouse districts, report commission and omission errors that are acceptable for regional statistics but too large for per-farm underwriting. The practical threshold for reliable individual-structure delineation is sub-5 m optical resolution or X-band SAR in Spotlight mode. This is not a processing problem that better algorithms can fully overcome; it is a physics constraint on what information the sensor encodes.
Cloud cover compounds the issue in humid regions such as the Netherlands, Belgium or South Korea's Chungcheongnam-do. A SAR coherence stack from Sentinel-1 or COSMO-SkyMed can identify persistent high-coherence structures, including greenhouse frames, through overcast conditions, though the 10 m Sentinel-1 product requires careful spatial filtering to avoid merging adjacent structures.
From pixels to insured area: the delineation pipeline
Automated greenhouse mapping follows a well-documented sequence. A multispectral or SAR image is segmented into objects using a multiresolution algorithm (eCognition-style or equivalent open-source implementations). Each object is scored against spectral indices, shape descriptors and, where SAR is available, backscatter intensity or coherence. A random forest or support vector machine classifier, trained on labelled samples from the target region, assigns a greenhouse or non-greenhouse label. Post-classification, adjacent objects are merged into individual structures and their footprint areas computed in a projected coordinate system.
Accuracy depends heavily on training-sample quality and on how well the classifier generalises across structure types. Glass and polyethylene have different spectral profiles; older weathered film differs from new. A classifier trained on Almería may underperform in the Nile Delta without retraining. Reported overall accuracies in peer-reviewed literature for high-resolution optical data range from roughly 88 to 96 percent, with the lower end typical of heterogeneous or partially shaded scenes. For underwriting, the operationally relevant figure is not overall accuracy but the systematic bias in total area: a classifier that correctly labels 92 percent of pixels can still over- or under-estimate district area by 5 to 10 percent if errors are spatially correlated.
What the underwriter actually needs, and where ambiguity remains
Crop-credit underwriting for protected horticulture requires three numbers: total insured planted area, the crop type inside each structure, and an indication of operational status at the time of cover. Satellite imagery answers the first question well. It answers the second poorly, because the spectral signature of the greenhouse film dominates the signal from the crop beneath. Thermal infrared can infer whether heating systems are active in winter, suggesting occupied structures, but crop-type identification inside a greenhouse from external imagery is not reliable at individual-farm resolution.
Operational status is partly inferrable from change detection. A greenhouse that has been dismantled, left open-ended, or whose film has degraded shows a different spectral and coherence profile than an active structure. Comparing two WorldView-3 acquisitions six months apart will flag structures that have changed materially. What satellite data cannot resolve is whether a structurally intact greenhouse is planted, fallow, or in a between-crop gap at the moment of survey. That residual ambiguity is honest, and underwriters should price it accordingly rather than treat the satellite area figure as a ground-truth substitute.
Satellize applies this pipeline in contexts ranging from sovereign agricultural census support to commercial credit analytics; the crop-estimation methodology developed for the Kingdom of Tonga programme uses a related object-based workflow adapted to open-field rather than protected horticulture.
Archive depth and the fraud-detection dividend
One underappreciated advantage of satellite-based area estimation is the archive. WorldView-3 has been collecting since 2014; Planet's daily archive runs from 2016. An underwriter reviewing a new credit application can commission a retrospective time series over the claimed farm location and check whether the greenhouse footprint existed before the application date, whether it has grown or shrunk, and whether the claimed expansion matches the construction timeline in the borrower's draw-down schedule.
This is straightforward to do and catches a specific class of fraud: inflated area claims on structures that do not exist or were built after the cover was sought. It does not catch misrepresentation of crop type or yield, which requires ground truth or independent crop-monitoring data. The archive is also the basis for multi-year exposure aggregation, allowing a portfolio underwriter to estimate total insured greenhouse area across a region and its year-on-year change, which is relevant for catastrophe modelling in hail, wind or frost-event scenarios.
Typical figures
| Best optical spatial resolution | 0.31 m pan / 1.24 m MS (WorldView-3); 3 m (Planet SuperDove) |
| Best SAR spatial resolution | ~1 m Spotlight (COSMO-SkyMed); 10 m IW (Sentinel-1) |
| Minimum reliably delineated structure width | ~4–6 m (sub-metre optical or X-band SAR); ~10–12 m (Sentinel-1 C-band) |
| Revisit (commercial tasking) | 1–4.5 days (WorldView-3, latitude-dependent); daily (Planet SuperDove) |
| Revisit (open SAR) | 6 days (Sentinel-1 two-satellite); schedulable days (COSMO-SkyMed) |
| Spectral bands used | NIR, Red, SWIR (optical); C-band 5.4 GHz or X-band 9.6 GHz (SAR) |
| Cloud penetration | SAR only; optical data blocked by cloud cover |
| Typical area-mapping accuracy (peer-reviewed range) | 88–96% overall accuracy; systematic area bias 5–10% in heterogeneous scenes |
| Archive depth | WorldView-3 from 2014; Planet daily from 2016; Sentinel-1 from 2014 |
| Delivery formats | GeoJSON / Shapefile polygon layer, area-summary CSV, change-detection report |
Analytics Satellize can run
| Greenhouse footprint polygon layer | Object-based image analysis (OBIA) with random forest classification on WorldView-3 or SuperDove multispectral data; shape metrics (elongation, rectangularity) plus NIR/SWIR spectral indices | GeoJSON polygon layer with per-structure area in hectares, confidence score and structure type (glass vs. polyethylene where distinguishable) |
| District-level planted-area estimate with uncertainty bounds | Aggregation of delineated polygons with bootstrap-resampled accuracy correction derived from validation sample; uncertainty expressed as 90% confidence interval on total area | PDF underwriting report with tabulated area by administrative unit and map inset |
| Year-on-year change detection for fraud screening | Bitemporal OBIA comparison between archive and current acquisition; change polygons flagged where structure appeared or disappeared between stated application date and prior image | Change-detection GIS layer with flagged parcels and acquisition-date metadata for due-diligence review |
| SAR coherence-based structure persistence map | Multi-temporal Sentinel-1 or COSMO-SkyMed coherence stack; high coherence over multiple passes indicates persistent rigid structures (greenhouse frames) regardless of cloud cover | Raster coherence map overlaid with optical-derived polygons; useful for cloud-affected growing regions |
| Operational-status indicator | Spectral change index comparing sequential Planet SuperDove acquisitions; degraded or removed film shows measurable NIR reflectance drop relative to baseline | Per-structure status flag (active / degraded / removed) appended to polygon layer |
| Portfolio exposure aggregation | Spatial join of insured-parcel registry against delineated greenhouse layer; area discrepancy computed per policy and ranked by magnitude | Ranked discrepancy table for portfolio review; input to catastrophe-model exposure file |
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.