Illicit crop cultivation monitoring: coca and cannabis
Coca and cannabis carry spectral and phenological signatures that distinguish them from surrounding vegetation. UNODC's ICMS programmes in Colombia, Peru, Bolivia and Afghanistan have demonstrated that Sentinel-2 time-series, combined with VHR optical confirmation, can map cultivation at national scale with sub-hectare sensitivity.
Sensors
- Sentinel-2 MSI (ESA): Primary time-series workhorse. Ten-metre resolution in four visible and NIR bands, twenty-metre in red-edge and SWIR. Five-day revisit at the equator with both satellites. Red-edge bands (B5, B6, B7) are particularly sensitive to the dense, waxy canopy of Erythroxylum coca. Free archive from 2015.
- Landsat 8/9 OLI (USGS/NASA): Thirty-metre resolution, sixteen-day revisit per satellite (eight days combined). Extends the historical record to 1972 via earlier Landsat missions. Useful for multi-year trend analysis and baseline land-cover classification before Sentinel-2 coverage begins.
- Maxar WorldView-2 (commercial): Forty-six-centimetre panchromatic, 1.85-metre multispectral including a dedicated red-edge band. Used to resolve ambiguous small plots (under 0.5 ha) that fall below Sentinel-2's reliable detection floor, and to visually confirm classification outputs for UNODC ground-truth workflows.
- Planet SuperDove: Three-metre resolution, eight spectral bands including red-edge, near-daily revisit over most cultivation zones. Bridges the gap between Sentinel-2's spectral depth and VHR's spatial detail. Useful for intra-annual harvest-cycle tracking where cloud cover breaks allow.
What a coca canopy gives away
Erythroxylum coca is an evergreen shrub with a dense, glossy canopy and high chlorophyll content. In near-infrared wavelengths (roughly 750 to 900 nm), it reflects strongly, producing a normalised difference vegetation index (NDVI) that is consistently higher than the surrounding secondary forest or pasture it often displaces. That contrast is the spectral handle every classification algorithm grabs first.
Cannabis is more seasonal and more variable in canopy structure, but outdoor plots in Colombia's Sierra Nevada or Afghanistan's Helmand province still show elevated red-edge reflectance during the vegetative growth phase. The challenge is separating it from other broadleaf crops at the same phenological stage. That is where time-series analysis earns its keep: the planting and harvest cadence of illicit crops rarely matches the calendar of legal agriculture in the same landscape.
How UNODC's ICMS actually works, and what the satellite record shows
UNODC's Integrated Illicit Crop Monitoring System has published annual coca surveys for Colombia, Peru and Bolivia since the late 1990s, with Afghanistan covered separately under the Afghanistan Opium Survey. The methodology is publicly documented: wall-to-wall satellite classification is cross-validated by stratified ground-truth sampling and, in Colombia, by aerial survey data from the national police. The satellite classification step uses supervised maximum-likelihood and, increasingly, random-forest classifiers trained on confirmed coca plots.
Colombia's 2022 survey reported approximately 230,000 hectares under coca cultivation, a figure derived primarily from Landsat and Sentinel-2 analysis. Peru's survey has consistently used Landsat 8 as its backbone. These are not estimates in the loose sense; they carry documented confidence intervals and known omission errors, mostly attributable to persistent cloud cover over Andean slopes during the growing season. UNODC publishes those error bounds, which is more candour than most national crop surveys manage.
The practical detection floor in UNODC's published workflows is roughly 0.5 hectares for Sentinel-2 classification and around 1 hectare for Landsat. Plots smaller than that are systematically under-counted. In fragmented landscapes where growers deliberately keep plots small to evade detection, this is a meaningful gap. VHR imagery closes it, but not at national scale, and not cheaply.
Cloud cover is the real adversary, not the growers
The Andean coca-growing regions sit inside some of the cloudiest terrain on Earth. The Colombian departments of Putumayo, Nariño and Norte de Santander routinely see fewer than thirty cloud-free Sentinel-2 acquisitions per year. That constrains the time-series density available for phenological analysis and forces analysts to composite images across multi-month windows, blurring the intra-annual signal.
Synthetic aperture radar (SAR) is the obvious complement, since C-band Sentinel-1 penetrates cloud. But coca and cannabis do not produce a distinctive backscatter signature that separates them reliably from other dense vegetation. SAR is more useful here for detecting land-clearing events that precede planting than for classifying the crop itself. The honest position is that cloud cover imposes a structural latency on detection: in the worst-affected zones, a new coca field may not appear in a clean optical composite until three to six months after planting.
Phenological time-series as a classification signal
Because coca is evergreen and harvested multiple times per year (typically three to four leaf harvests annually in Colombia), its NDVI trajectory through a calendar year is flatter and higher than that of annual crops or recovering secondary forest. A pixel that maintains NDVI above roughly 0.6 across all cloud-free observations in a year, without the seasonal dip that pasture or maize would show, is a strong candidate for perennial shrub cultivation.
Cannabis, being annual, shows a sharper growth-and-senescence curve. Detecting it requires denser time-series coverage timed to the vegetative phase, which is exactly the period most likely to be cloud-obscured in tropical growing regions. In Afghanistan's Helmand, where cloud cover is less severe, Sentinel-2 time-series have been used to map cannabis alongside poppy, though the poppy work is more mature and better documented in the public literature.
Random-forest classifiers trained on multi-temporal feature stacks (NDVI, EVI, red-edge chlorophyll index across multiple dates) consistently outperform single-date classification in published comparisons. Accuracy figures in the peer-reviewed literature on coca mapping typically range from 85 to 92 percent overall accuracy, with producer accuracy for the coca class somewhat lower due to confusion with other perennial crops and agroforestry.
Honest limits of the method
Sub-hectare plots are systematically missed at Sentinel-2 and Landsat resolution. Intercropping of coca within legitimate agroforestry (a documented evasion strategy) reduces spectral separability significantly. VHR imagery resolves individual plant rows but cannot be tasked at national scale without prohibitive cost. And any classification is only as good as its training data: if ground-truth samples are biased toward accessible areas, the classifier will under-perform in remote, high-altitude plots.
Temporal latency is a second hard limit. This method produces area estimates, not real-time alerts. The UNODC annual survey cycle means that policy responses are working from data that is, at best, six to twelve months old by the time it reaches decision-makers. Shortening that cycle requires more frequent tasking, faster processing pipelines and, in cloud-prone regions, acceptance of higher classification uncertainty.
Satellize runs Sentinel-2 and Landsat time-series analytics on open constellations for government clients, adding commercial VHR tasking where plot-level confirmation is required. The crop-estimation methods developed for the Kingdom of Tonga programme share the same phenological feature-extraction pipeline used in illicit-crop workflows, adapted to different target species and legal contexts.
What a monitoring programme actually delivers
A well-designed national coca or cannabis monitoring programme produces three things that matter to a government client: an annual area estimate with documented uncertainty bounds, a change map showing where cultivation expanded or contracted relative to the previous survey period, and a hotspot layer identifying the districts where new clearing activity is concentrated.
The change map is often more operationally useful than the absolute area figure. Enforcement agencies care less about whether there are 180,000 or 200,000 hectares nationally than about which specific valleys saw a 40 percent increase in the past twelve months. That is the layer that directs resources. Producing it reliably requires consistent methodology across survey years, which is an argument for maintaining a stable processing chain rather than chasing the latest classifier each season.
Typical figures
| Primary spatial resolution | 10 m (Sentinel-2 visible/NIR), 30 m (Landsat 8/9), 3 m (Planet SuperDove), 0.46 m pan (WorldView-2) |
| Revisit cadence | 5 days (Sentinel-2 combined), 8 days (Landsat 8+9 combined), near-daily (Planet SuperDove) |
| Minimum reliably detectable plot | ~0.5 ha at Sentinel-2 resolution; ~1 ha at Landsat resolution; <0.1 ha with WorldView-2 VHR confirmation |
| Key spectral bands | Red-edge (705–783 nm), NIR (842 nm), SWIR (1610 nm, 2190 nm) for Sentinel-2; comparable OLI bands for Landsat |
| Classification accuracy (published range) | 85–92% overall accuracy in peer-reviewed coca-mapping studies; producer accuracy for coca class typically 80–88% |
| Cloud-cover constraint | Fewer than 30 usable Sentinel-2 scenes per year in worst-affected Andean zones; compositing windows of 3–6 months required |
| Archive depth | Sentinel-2 from 2015; Landsat from 1972 (earlier missions); Planet SuperDove from approximately 2021 |
| Survey latency (UNODC model) | 6–12 months from image acquisition to published national estimate |
| Typical delivery formats | GeoTIFF classification rasters, GeoPackage/Shapefile change polygons, district-level CSV area tables, PDF survey report |
Analytics Satellize can run
| Annual cultivation area estimate | Supervised random-forest classification on multi-temporal Sentinel-2 and Landsat feature stacks (NDVI, EVI, red-edge chlorophyll index); stratified accuracy assessment against ground-truth samples | National and sub-national area tables with confidence intervals, GeoTIFF classification layer, PDF survey report |
| Year-on-year change map | Post-classification change detection comparing consecutive annual classification outputs; minimum mapping unit 0.5 ha | GeoPackage polygon layer of expansion and reduction areas, district-level change statistics CSV |
| New-clearing alert layer | NDVI time-series breakpoint detection (BFAST or equivalent) applied to Sentinel-2 dense stack; flags abrupt vegetation-loss events followed by perennial-shrub regrowth signature | Monthly GIS alert layer of candidate new-plot locations, flagged for VHR follow-up tasking |
| Harvest-cycle frequency estimate | Intra-annual NDVI phenology decomposition on cloud-gap-filled Sentinel-2 composites; number of harvest events inferred from partial-canopy reflectance dips | Per-pixel harvest-frequency raster, aggregated to plot and district level |
| VHR plot-level confirmation | Object-based image analysis on WorldView-2 or comparable VHR imagery; row-spacing and canopy-texture features used to confirm or reject candidate plots from medium-resolution classification | Confirmed-plot shapefile with species-confidence score, analyst notes |
| Multi-year trend and hotspot report | Time-series regression on annual area estimates by district; Mann-Kendall trend test to identify statistically significant expansion zones | Hotspot ranking table, trend maps, annual briefing document for enforcement planning |
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.