Oil palm expansion deforestation attribution for supply-chain carbon claims
Zero-deforestation commitments live or die on whether a specific concession cleared forest before or after a cut-off date. Satellite time-series can establish that date, classify oil palm from its geometry and spectral signature, and distinguish new clearing from replanting.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in red-edge and shortwave infrared. Five-day revisit at the equator (combined Sentinel-2A and 2B). The red-edge bands (705 nm, 740 nm, 783 nm) are particularly useful for separating mature oil palm from other broadleaf crops and for detecting the early green flush after clearing. Free archive from 2015.
- Landsat 8/9 OLI and Landsat archive: 30 m multispectral, 16-day single-satellite revisit. The archive runs continuously from 1972 (MSS) and consistently from 1984 (TM), making it the only freely available source for establishing pre-2015 forest baselines and historical clearing chronologies. USGS Collection 2 provides analysis-ready surface reflectance.
- Planet SuperDove: 3 m resolution, eight spectral bands including red-edge, daily revisit over most tropical latitudes on a commercial licence. At 3 m, individual palm crowns (typically 5–9 m diameter at maturity) are resolvable, allowing direct crown-count validation and detection of early-stage replanting versus new clearing. Not freely available; requires client licensing.
- Sentinel-1 SAR (C-band): 10 m ground range detected, six-day revisit in tropical regions. C-band backscatter penetrates cloud cover that routinely blankets Borneo and Sumatra for weeks at a time, maintaining change-detection continuity where optical sensors go dark. SAR backscatter from oil palm canopy is measurably lower than from closed-canopy forest due to differences in vertical structure, though the contrast diminishes as palms mature beyond roughly 10 years.
What a palm crown gives away
Oil palm (Elaeis guineensis) has a radially symmetric crown of stiff pinnate fronds arranged in a near-perfect rosette. Viewed from above at 3–10 m resolution, mature palms produce a repeating dot pattern with predictable inter-crown spacing of roughly 8–9 m in commercial plantations planted on a triangular grid. No other tropical crop matches this geometry at scale. Sentinel-2 at 10 m resolves the canopy texture as a characteristic mottled surface; Planet SuperDove at 3 m resolves individual crowns. Both are sufficient for automated classification using convolutional neural networks or object-based image analysis trained on labelled plantation polygons.
Spectral behaviour adds a second discriminator. Mature oil palm has a high normalised difference vegetation index (NDVI) that overlaps with forest, which is why NDVI alone cannot separate them. The red-edge bands on Sentinel-2 (particularly the ratio of Band 6 at 740 nm to Band 5 at 705 nm) and the shortwave infrared response differ enough between closed-canopy forest and plantation that a multi-feature classifier achieves reported overall accuracies above 90 percent in published studies using Southeast Asian training data. The honest caveat: young palms in the first two to three years after planting have low canopy cover and are spectrally closer to bare soil or scrub, making them harder to classify and creating a detection gap precisely when the clearing event is most recent.
Establishing the clearing date, not just the clearing
A zero-deforestation claim is not a static map question. It is a temporal question: was this land forested on a specific reference date, and if so, when was it cleared? The Landsat archive, processed through the USGS Collection 2 surface reflectance pipeline, provides annual or sub-annual composites from 1984 onward. The standard approach builds a forest/non-forest time-series using spectral indices and change-detection algorithms such as LandTrendr or CCDC (Continuous Change Detection and Classification), both published and freely available through Google Earth Engine. These algorithms fit piecewise regression models to the spectral history of each pixel and identify the date of abrupt change with a precision of roughly one to three months for clear-cut events in years with adequate cloud-free observations.
In persistently cloudy regions, cloud-free Landsat observations can drop to two or three per year, which degrades temporal precision. Sentinel-1 SAR fills part of this gap: C-band backscatter drops sharply when forest is cleared, and the change is detectable even under complete cloud cover. Combining optical and SAR change detection in a fusion approach narrows the clearing-date uncertainty compared to either sensor alone. The practical floor for clearing-date precision in high-cloud regions is approximately plus or minus one quarter-year, which is usually sufficient for supply-chain cut-off compliance but should be stated explicitly in any audit report.
Replanting versus new clearing: the hardest distinction
Commercial oil palm has an economic lifespan of roughly 25 years. At the end of a rotation, palms are felled, the land is briefly bare or covered with debris, and new seedlings are planted. From a spectral and structural standpoint, this replanting event looks almost identical to a new forest-clearing event: canopy disappears, bare soil is exposed, NDVI collapses. Misclassifying a replanting cycle as deforestation would falsely attribute carbon loss; missing a new clearing by assuming it is replanting would understate it. Both errors matter for carbon claims.
The distinguishing evidence is the prior land-cover history. If the pixel was classified as oil palm plantation for the preceding 10 or more years, a canopy-loss event is almost certainly replanting. If the prior classification is closed-canopy forest or mixed forest, the same spectral signal is almost certainly clearing. This means the attribution logic depends entirely on the quality of the historical time-series going back at least one full rotation before the reference date. Where the archive is ambiguous, for instance because a concession was granted on land that had already been partially degraded, the uncertainty should be reported as a range rather than a single date.
Concession-level attribution and the supply-chain link
Individual mill supply sheds in Indonesia and Malaysia typically cover hundreds of thousands of hectares and aggregate fruit from dozens of concessions. Tracing a carbon claim to a specific concession requires matching the satellite-derived clearing polygon to a concession boundary. Concession boundary data quality varies enormously by jurisdiction. Indonesia's One Map Policy has improved public availability of palm oil concession boundaries, but gaps and disputed boundaries remain. Malaysia's Sabah and Sarawak states have more limited public disclosure. Where official boundaries are unavailable, plantation perimeters can be approximated from the classified palm extent itself, though this introduces circularity into the analysis.
The satellite analysis produces a per-polygon clearing date and a forest/non-forest status at the chosen reference date. That output can be compared directly against a company's stated cut-off date (commonly 2016, 2018 or 2020 depending on the certification scheme). Polygons that show clearing after the cut-off date and within a concession linked to a specific mill are the flagged events. The analysis does not resolve commercial relationships between mills and buyers; that chain-of-custody data must come from trade documentation. Satellite evidence establishes the physical fact of clearing; it cannot establish who ordered it.
Honest limits of the method
Resolution is the first limit. At 30 m (Landsat), clearings smaller than roughly one to two hectares may be missed or their area underestimated. At 10 m (Sentinel-2), the practical detection floor drops to around 0.1–0.25 hectares for distinct clear-cuts. Gradual or selective clearing within a plantation mosaic is harder to detect than a sharp forest-to-bare-soil transition.
Cloud cover is the second. Tropical peatland regions in Kalimantan and Sumatra can have fewer than 20 cloud-free Sentinel-2 observations per year, and some years produce almost none. SAR partially compensates, but SAR change detection has its own false-positive rate from flooding, wind damage and other non-deforestation disturbances. Any audit report should state the number of cloud-free observations used to establish the baseline and the clearing date.
The third limit is that satellite data cannot determine the legal status of a clearing or whether it was authorised under local land-use permits. It establishes the physical and temporal fact. Legal and commercial attribution requires additional documentation.
What an audit package should contain
A credible satellite-based deforestation attribution report for a supply-chain carbon claim needs at minimum: a classified oil palm extent map with confidence scores, a forest/non-forest time-series for each concession polygon from at least 1990 to the present, a clearing-date estimate with stated uncertainty for each flagged polygon, a replanting-versus-new-clearing determination with the evidence basis, and a cloud-cover and data-availability log for the analysis period. Deliverables should be in formats that auditors and certification bodies can ingest directly, typically GeoTIFF rasters and GeoJSON or shapefile polygon layers with attribute tables, accompanied by a methodology note specifying the algorithms, training data vintage and validation accuracy.
Satellize builds these packages on open constellations (Sentinel, Landsat) with optional commercial tasking through client licence for the high-resolution validation layer. The workflow is the same one that underpins the crop-estimation programme in Tonga, adapted for the considerably larger spatial scales and longer time horizons that palm supply-chain work demands.
Typical figures
| Primary optical resolution | 10 m (Sentinel-2 visible/NIR); 30 m (Landsat); 3 m (Planet SuperDove, commercial licence) |
| SAR resolution | 10 m ground range detected (Sentinel-1 IW mode) |
| Revisit frequency | 5 days at equator (Sentinel-2A+B combined); 16 days (single Landsat); 6 days (Sentinel-1); daily (Planet SuperDove) |
| Archive depth | 1984 to present (Landsat TM/ETM+/OLI); 2015 to present (Sentinel-2); 2014 to present (Sentinel-1) |
| Minimum detectable clearing | ~0.1–0.25 ha at 10 m optical; ~1–2 ha at 30 m; cloud cover and scene geometry affect this floor |
| Clearing-date temporal precision | ±1–3 months in low-cloud years; ±1 quarter-year typical in high-cloud tropical regions |
| Key spectral bands | Red-edge (705 nm, 740 nm, 783 nm), SWIR (1610 nm, 2190 nm), NIR (842 nm) on Sentinel-2; C-band 5.4 GHz on Sentinel-1 |
| Delivery formats | GeoTIFF (classified raster), GeoJSON/Shapefile (concession polygons with attributes), CSV (clearing-event table), PDF methodology note |
| Coverage | Global; primary focus regions are Indonesia (Kalimantan, Sumatra, Papua) and Malaysia (Sabah, Sarawak) |
| Analysis latency | Historical audit packages: 4–8 weeks depending on concession count and archive depth; near-real-time monitoring updates: configurable, typically monthly |
Analytics Satellize can run
| Oil palm plantation extent map | Object-based image analysis or CNN classification on Sentinel-2 red-edge and SWIR composites, with geometric crown-pattern features from high-resolution imagery where licensed | GeoTIFF classified raster and polygon layer with confidence scores, delivered per concession or per mill supply shed |
| Historical forest/non-forest time-series | LandTrendr or CCDC spectral change detection on Landsat Collection 2 surface reflectance, extended with Sentinel-2 from 2015 | Annual forest-cover stack (GeoTIFF) per concession from 1990 to present, with data-availability log |
| Clearing-date attribution per polygon | Breakpoint detection from spectral time-series, cross-validated with Sentinel-1 SAR backscatter change for cloud-affected periods | Shapefile or GeoJSON polygon layer with clearing-date field, uncertainty range, and prior land-cover classification |
| Replanting versus new-clearing determination | Prior land-cover history lookup against classified time-series; flagging logic based on plantation age and rotation-cycle priors | Attribute column in clearing-event table with determination (replanting / new clearing / ambiguous) and evidence basis |
| Cut-off date compliance assessment | Intersection of clearing-date polygons with concession boundaries and client-specified reference date; flagged events tabulated by concession ID | Compliance report (PDF and CSV) listing flagged concessions, clearing area in hectares, and estimated carbon stock at time of clearing using published biomass density ranges for the relevant forest type |
| Ongoing monitoring feed | Monthly Sentinel-2 and Sentinel-1 change detection against the baseline classification, with alert generation for new canopy-loss events above a configurable area threshold | Monthly GeoJSON alert layer and summary report, suitable for ingestion into supply-chain due-diligence platforms |
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.