Unlicensed tobacco plantation detection for excise enforcement
Unlicensed tobacco cultivation supplies untaxed cigarette manufacturing across the EU, Ukraine and South America. Multispectral imagery can fingerprint tobacco's canopy and phenology, flagging unregistered plots for excise investigators.
Sensors
- Sentinel-2 MSI: 10 m visible and NIR bands, 20 m red-edge bands (B5 704 nm, B6 740 nm, B7 783 nm). Free, 5-day revisit over Europe. The 10 m pixel floor limits confident detection to plots larger than roughly 0.5 ha; smaller patches are spectrally mixed and unreliable.
- PlanetScope SuperDove: 3 m resolution, 8 spectral bands including two red-edge channels (at 705 nm and 740 nm), near-daily revisit globally. Sufficient to resolve plots down to approximately 0.05 ha and to distinguish row structure within a canopy.
- Airbus Pléiades: 50 cm panchromatic, 2 m multispectral. Tasked on demand. Used for confirmation of candidate sites identified by coarser sensors: row spacing, canopy texture and field-edge infrastructure (curing barns, irrigation risers) are visible at this resolution.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit. Too coarse for small plots but useful for building multi-year phenological baselines across large regions and detecting year-on-year changes in agricultural land use at low cost.
Why tobacco is spectrally distinguishable
Tobacco (Nicotiana tabacum) is a broadleaf annual with a high leaf-area index and unusually thick, waxy leaves. That physiology produces a distinctive red-edge slope in reflectance spectra: the transition between red absorption and near-infrared reflectance is steeper and shifts to slightly longer wavelengths compared with cereals, oilseed rape or sugar beet at equivalent growth stages. Sentinel-2's B5 (704 nm), B6 (740 nm) and B7 (783 nm) bands capture this slope directly, and indices derived from them, particularly the Red-Edge Chlorophyll Index (CIre) and the MERIS Terrestrial Chlorophyll Index (MTCI), have been used in published crop-mapping literature to separate tobacco from other broadleaf crops with accuracy above 85 percent in controlled study areas.
Phenology adds a second discriminator. In central and eastern Europe, tobacco is transplanted in May, reaches peak canopy closure in July and August, and is harvested progressively through August and September, leaving bare or lightly residued soil well before maize harvest. This temporal pattern is specific enough that a time-series of four to six Sentinel-2 acquisitions across a growing season can distinguish tobacco from sunflower, maize and potato with fewer misclassifications than any single-date image. The combination of spectral signature and phenological timing is the foundation of the detection approach.
What the enforcement problem actually looks like
OLAF, the EU's anti-fraud office, has documented illicit tobacco cultivation as a supply-side input to counterfeit and contraband cigarette manufacturing. The plantations are typically registered as other crops, or not registered at all, to avoid excise obligations that attach to tobacco at the point of cultivation in several member states. Field sizes range from under one hectare to tens of hectares. Operators often rotate locations between seasons and favour areas with weak cadastral enforcement.
That rotation is both the operational challenge and an argument for satellite monitoring. Ground inspection is slow and geographically sparse. A satellite-based screening layer that flags candidate tobacco plots each season, ranked by confidence and cross-referenced against agricultural subsidy registrations, gives investigators a prioritised target list rather than a needle-in-a-haystack search. The approach is analogous to methods published for coca and poppy monitoring, adapted to a temperate agricultural context.
Detection method and honest limits
The standard workflow begins with a Sentinel-2 time-series classification. A supervised classifier (random forest or support vector machine are the most commonly published choices in remote-sensing crop-mapping literature) is trained on confirmed tobacco parcels from cadastral records or ground truth, then applied to candidate areas. Red-edge-derived indices, NDVI trajectory shape and harvest-timing signals are the primary features. Where Sentinel-2 flags a candidate plot, PlanetScope imagery at 3 m is pulled for the same date range to verify canopy texture and row geometry.
The limits are real and worth stating plainly. The 10 m Sentinel-2 pixel means plots smaller than 0.5 ha produce too few pure pixels for reliable classification; those require PlanetScope as the primary sensor, which raises cost. Cloud cover over central Europe in May and June can interrupt the time-series at the critical transplanting window, requiring either SAR gap-filling (Sentinel-1 C-band backscatter correlates weakly with canopy biomass but cannot reproduce red-edge discrimination) or acceptance of reduced confidence for that season. Tobacco grown under shade nets or polytunnels is largely invisible to spectral classification. And a detection is a candidate, not a conviction: ground verification or cross-referencing with tax and subsidy databases is always required before enforcement action.
Turning detections into an enforcement intelligence product
Raw classified pixels are not useful to an excise investigator. The analytic chain needs to convert them into georeferenced parcels with confidence scores, estimated area in hectares, year-on-year change flags and proximity to known processing infrastructure. Each candidate parcel can be cross-referenced against the EU's Land Parcel Identification System (LPIS) declarations, where available, to identify the declared crop and the registered claimant for agricultural subsidies. A mismatch between declared crop and satellite-detected crop is the enforcement trigger.
Satellize structures this kind of output as a seasonal GIS layer delivered after each growing-season classification run, with a supplementary alert feed for any parcel where PlanetScope confirmation changes the confidence rating materially. The Tonga crop-estimation programme demonstrated that smallholder-scale crop mapping from open constellations is operationally viable; the tobacco detection problem is methodologically similar but adds the cross-referencing step against administrative registers.
Archive depth matters here. Sentinel-2 data is available from 2017 and Landsat from the 1980s. A five-year retrospective analysis can establish which parcels have hosted tobacco repeatedly, which is a stronger indicator of deliberate unlicensed cultivation than a single-season detection.
What the method cannot see, and what fills the gap
Spectral classification identifies the plant. It says nothing directly about ownership, intent or the downstream supply chain. A plantation that is genuinely mis-declared on a subsidy form is a different enforcement matter from one that is deliberately concealed. Satellite data establishes the physical fact of cultivation; investigators supply the legal and commercial context.
Very small plots, plots under shade structures, and plots in persistently cloudy regions remain hard targets. For those, the practical supplement is not a better satellite but better tip integration: combining the satellite screening layer with customs seizure data, known processing-site locations and transport network analysis. The satellite layer is most powerful as a systematic first filter across large agricultural regions, not as a standalone case-builder.
Typical figures
| Primary spatial resolution | 10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 red-edge), 3 m (PlanetScope SuperDove) |
| Confirmation resolution | 50 cm pan / 2 m multispectral (Pléiades, tasked) |
| Revisit frequency | 5 days (Sentinel-2, Europe); near-daily (PlanetScope); 16 days (Landsat 8/9) |
| Key spectral bands | Red-edge B5 704 nm, B6 740 nm, B7 783 nm (Sentinel-2); red-edge 705 nm and 740 nm (PlanetScope SuperDove) |
| Minimum detectable plot (Sentinel-2) | Approximately 0.5 ha (below this, pixel mixing degrades classification reliability) |
| Minimum detectable plot (PlanetScope) | Approximately 0.05 ha under clear-sky conditions |
| Seasonal acquisitions required | 4 to 6 cloud-free images spanning May to September for phenological classification |
| Archive depth | Sentinel-2 from 2017; Landsat from 1984 (30 m, for multi-year trend baselines) |
| Cloud limitation | Persistent cloud cover in May/June can break the transplanting-window observation; SAR provides partial gap-fill only |
| Delivery formats | GeoTIFF classified raster, GeoPackage or Shapefile parcel polygons with confidence scores, CSV cross-reference table for LPIS matching |
Analytics Satellize can run
| Seasonal tobacco candidate parcel map | Supervised random-forest classification on Sentinel-2 red-edge time-series (CIre, MTCI, NDVI trajectory); published crop-mapping methodology | GIS polygon layer with per-parcel area (ha), classification confidence score and harvest-timing flag, delivered post-season |
| PlanetScope canopy-texture confirmation | Object-based image analysis on 3 m SuperDove imagery to verify row structure and canopy architecture at flagged sites | Confirmation report per candidate site with annotated image chips and revised confidence rating |
| LPIS declaration mismatch table | Spatial join of satellite-detected tobacco parcels against client-supplied LPIS or cadastral declared-crop data | CSV table of parcels where satellite detection conflicts with declared crop, ranked by area and confidence |
| Multi-year recurrence analysis | Change detection across Sentinel-2 archive (2017 to present) and Landsat OLI time-series to identify parcels with repeated tobacco cultivation | Recurrence map and ranked parcel list flagging sites with two or more seasons of detected tobacco |
| Pléiades confirmation tasking | On-demand tasking of 50 cm imagery for highest-priority candidate parcels; visual and semi-automated assessment of curing barn presence, irrigation infrastructure and row geometry | Site-level intelligence report with annotated very-high-resolution image and infrastructure inventory |
| Regional screening baseline | Unsupervised clustering on Landsat 8/9 OLI phenological composites to identify agricultural areas with anomalous broadleaf crop signatures for prioritising Sentinel-2 tasking | Priority-area polygon set for focused Sentinel-2 and PlanetScope analysis, reducing processing area and cost |
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.