Commodity-driven deforestation attribution and supply-chain traceability
Detecting deforestation is solved. Attributing it to cattle, soy, palm oil or timber concessions requires layering spectral trajectories, cadastral boundaries and concession maps, and the answer is often probabilistic rather than certain.
Sensors
- Planet NICFI Basemaps: Monthly and biannual 4.77 m RGB+NIR mosaics covering tropical forest nations under the Norway NICFI programme. Fine enough to distinguish early pasture grass regrowth from soy seedling rows, though spectral depth (four bands only) limits crop-type discrimination.
- Sentinel-2 MSI: 10 m resolution in visible and NIR bands, 20 m in red-edge and SWIR. Five-day revisit at the equator (combined Sentinel-2A and 2B). SWIR bands 11 and 12 separate bare soil from green cover and help distinguish soy from pasture during the growing season. Cloud contamination in humid tropics can reduce usable acquisitions to fewer than six per year in some locations.
- Landsat-8/9 OLI: 30 m multispectral, 16-day single-satellite revisit. Forty-plus years of archive enables disturbance-history reconstruction back to the 1980s. Thermal band adds a proxy for land surface temperature that can distinguish irrigated cropland from dry pasture, though at 100 m resampled resolution the signal is coarse.
- MapBiomas Brazil annual mosaics: Annual land-cover classifications at 30 m derived from Landsat time series, using Random Forest trained on Brazilian biomes. Provides post-clearing land-use labels (pasture, soy, sugarcane, etc.) from 1985 onward. Coverage is Brazil-centric; analogous programmes exist for the Amazon, Cerrado and Pantanal but quality varies by biome.
What a spectral trajectory gives away
A clearing event looks similar regardless of what replaces the forest: a sharp drop in NDVI followed by bare-soil reflectance. The driver reveals itself in the months after clearing. Pasture re-greens slowly and unevenly, with a spectral signature close to C4 grasses. Soy produces a dense, spectrally uniform green flush within sixty to ninety days of planting, then senesces abruptly at harvest. Oil palm shows a distinctive canopy geometry visible in very-high-resolution imagery within two to three years of planting. Timber concessions often leave slash and residual woody debris that elevates SWIR reflectance for months.
Sentinel-2's red-edge bands (705 nm and 740 nm) are particularly useful here. Chlorophyll content in soy differs measurably from that in grass pasture during peak greenness, and the red-edge chlorophyll index separates them more reliably than NDVI alone. This is not a clean binary: speculative clearing, where land is held bare awaiting a buyer, produces no post-clearing signal at all, and attribution must then rest on spatial context rather than phenology.
Concession layers and the cadastral problem
Global Forest Watch publishes concession boundaries for palm oil, timber and mining across much of Southeast Asia and Central Africa, derived from government disclosures and investigative reporting. Overlaying deforestation polygons with these layers immediately flags clearing inside a licensed area, which is a different legal and commercial risk from clearing on unregistered land. Brazil's SIGEF rural cadastre covers most private landholdings and is publicly downloadable, making it possible to assign a clearing event to a specific registered property.
The practical difficulty is that cadastral data and concession maps are rarely current. Boundaries shift through subdivision, informal transfer and deliberate misregistration. A clearing that falls one hundred metres outside a palm concession boundary may still be causally linked to that concession. Attribution methods in peer-reviewed supply-chain research, including work published in journals such as Science and Nature Food, typically use spatial buffers and probabilistic scoring rather than hard boundary intersections, precisely because the underlying geodata is imperfect.
Soy in the Brazilian Cerrado presents a particular challenge. Much of the expansion since 2008 has occurred on land that was already degraded savanna rather than closed-canopy forest, which means it falls outside some deforestation definitions. Whether that clearing counts as commodity-driven forest loss depends entirely on the definition applied, and different buyers, certifiers and regulators apply different ones.
The probabilistic attribution framework
No single image or overlay produces a definitive driver label. The published approach, used in studies underpinning the EU Deforestation Regulation (EUDR) due-diligence requirements, combines four evidence streams: the post-clearing spectral trajectory, the spatial overlap with commodity-specific concessions or farm boundaries, the proximity to existing commodity infrastructure (mills, slaughterhouses, ports), and the regional commodity expansion trend in the same period.
Each stream contributes a likelihood score. A clearing inside a registered soy farm, within fifty kilometres of a crushing plant, with a soy-like phenological trajectory, scores high confidence. A clearing on unregistered land with no detectable crop regrowth scores low confidence for any specific commodity. Outputs are typically expressed as probability distributions across driver classes, not point estimates. Buyers and compliance teams need to understand this: the satellite data does not certify deforestation-free supply chains; it quantifies residual risk.
Where Planet NICFI changes the geometry
At 30 m, a small clearing of one or two hectares, common in smallholder palm or cocoa expansion in West Africa, is detectable but its internal land-use pattern is not. Planet NICFI's sub-5 m basemaps resolve individual tree rows, access tracks and field drainage patterns that are diagnostic of specific commodities. The trade-off is spectral: four bands versus Sentinel-2's thirteen. In practice, the two sources are complementary. Sentinel-2 provides the spectral depth for crop-type discrimination; NICFI provides the spatial detail to confirm field geometry and to catch small clearings that fall below Landsat's minimum mapping unit.
NICFI coverage is restricted to tropical forest nations, and the monthly compositing means a clearing that occurred and was replanted within a single month could be missed entirely. Biannual basemaps, which go back to 2015, are more useful for historical attribution than for near-real-time work. That historical depth matters for supply-chain due diligence, where buyers are increasingly asked to demonstrate that land was not cleared after a specific cut-off date, often 31 December 2020 under the EUDR.
Honest limits: what attribution cannot do
Cloud cover in humid tropical regions is the most persistent technical constraint. In parts of the Congo Basin and Indonesian Borneo, optical sensors may return fewer than four cloud-free observations per year. SAR (Synthetic Aperture Radar) from Sentinel-1 can detect the structural change of clearing through cloud, but SAR backscatter does not carry the spectral information needed for crop-type attribution. The two sensor types are complementary but not interchangeable.
Attribution confidence degrades sharply where multiple commodities co-occur spatially, where land is cleared speculatively, or where informal smallholder expansion dominates. Smallholder mosaics, common in much of Southeast Asia and West Africa, mix subsistence crops, cash crops and fallow in patches too small and too mixed for reliable commodity labelling from orbit. Published accuracy assessments for driver-attribution models in the Amazon typically report overall accuracies of 70 to 85 percent at the polygon level; in more heterogeneous landscapes, that figure is lower and less well documented.
Satellize can run commodity attribution workflows on Sentinel-2 and Landsat archives, integrating Global Forest Watch concession layers and MapBiomas classifications. The output is a scored GIS layer with explicit confidence intervals, not a binary certification. For teams building EUDR compliance systems or investor-grade supply-chain risk assessments, that distinction between evidence and certification is the one that matters most.
Typical figures
| Primary optical resolution | 4.77 m (Planet NICFI); 10–20 m (Sentinel-2 MSI); 30 m (Landsat-8/9 OLI) |
| Revisit frequency | Monthly composites (NICFI); 5 days at equator (Sentinel-2A+2B combined); 16 days per satellite (Landsat-8 or 9) |
| Spectral bands used | RGB + NIR (NICFI); 13 bands including red-edge (705, 740, 783 nm) and SWIR (1610, 2190 nm) for Sentinel-2; 7 OLI bands plus thermal for Landsat |
| Minimum detectable clearing | ~0.1 ha with Planet NICFI; ~0.5 ha with Sentinel-2; ~1 ha reliably with Landsat 30 m |
| Archive depth | 2015 to present (NICFI biannual); 2017 to present (Sentinel-2); 1972 to present (Landsat); 1985 to present (MapBiomas Brazil) |
| Attribution confidence (published benchmarks) | 70–85% overall accuracy at polygon level in Amazon studies; lower in heterogeneous smallholder landscapes |
| Cloud impact | Fewer than 4 usable optical acquisitions per year possible in Congo Basin and parts of Indonesian Borneo |
| Key ancillary layers | Global Forest Watch concession maps; Brazil SIGEF cadastre; MapBiomas annual land-cover classifications; EUDR commodity cut-off date (31 Dec 2020) |
| Delivery format | GeoPackage or Shapefile with per-polygon driver probability scores; GeoTIFF time-series stacks; tabular CSV for compliance reporting |
Analytics Satellize can run
| Deforestation polygon layer with commodity driver scores | Change detection on Sentinel-2 or Landsat NDVI/NBR time series, followed by Random Forest classifier trained on post-clearing spectral trajectories and spatial covariates | GIS polygon layer with probability scores for cattle, soy, palm, timber and speculative clearing; updated quarterly or on demand |
| Concession-level clearing attribution report | Spatial intersection of deforestation polygons with Global Forest Watch concession boundaries and national cadastral data, with 500 m buffer analysis for boundary uncertainty | PDF and CSV report per concession or farm unit, flagging clearings by confidence tier for compliance teams |
| Post-clearing land-use phenological profile | Harmonic analysis and time-series segmentation (BFAST or equivalent) on Sentinel-2 red-edge and SWIR bands to extract crop-cycle timing and amplitude | Per-polygon phenological fingerprint chart and crop-type probability table |
| Historical clearing timeline for EUDR due diligence | Landsat archive stack analysis from 1985 onward using MapBiomas classifications; change-point detection relative to 31 December 2020 EUDR cut-off | Binary pre/post cut-off clearing flag with date-of-first-disturbance estimate and confidence interval, delivered as GIS layer and tabular summary |
| Supply-chain risk score by sourcing region | Aggregation of polygon-level attribution scores weighted by commodity infrastructure proximity (mill, slaughterhouse, port buffers); method follows published supply-chain deforestation risk literature | Regional risk dashboard with commodity-specific exposure metrics, suitable for investor ESG reporting or procurement screening |
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.
