- Dataset Availability
- 1990-01-01T00:00:00Z–2021-12-31T23:59:59Z
- Dataset Producer
- Biopama Programme
- Tags
Description
Global oil palm extent for 2016-2021 at 10 m resolution and oil palm planting year from 1990 to 2021 at 30 m resolution. The dataset consists of 609 100x100 km grid tiles covering tropical and subtropical oil-palm-growing regions worldwide, mapping 23.98 Mha of closed-canopy oil palm (16.69 Mha [69.6%] industrial and 7.29 Mha [30.4%] smallholder), with a sample-based area estimate of 16.82 ± 0.19 Mha industrial and 7.37 ± 0.25 Mha smallholder. Note that although the catalog time range spans 1990-2021 to reflect the planting year range, this collection is a single static mosaic rather than a multi-temporal time series.
Methodology
- Oil Palm Extent (
classification, 10 m): Annual closed-canopy oil palm maps for 2016-2021 were generated from Sentinel-1 C-band SAR GRD dual-polarization (VV and VH) annual composites using a DeepLabv3+ (MobileNetV2 backbone) convolutional neural network (CNN) trained on 296 10x10 km regions. Daily Sentinel-1 scenes were corrected for local incident angle using the SRTM v4 DEM, aggregated into annual median composites separately for ascending and descending orbits, and averaged across both orbits to avoid terrain artifacts when one orbit had more passes. Unlike v1 (2019), Sentinel-2 red reflectance (Band 4) was excluded because systematic Sentinel-2 Level-2A coverage of oil palm regions was unavailable prior to 2019; to reuse the 3-channel DeepLabv3+ architecture from v1, the VV and VH composites were stacked with a zero-filled third channel. - To reduce false positives (such as coconut plantations and paddy fields), overlapping cropland, built-up, waterbody, herbaceous wetland, and mangrove classes from ESA WorldCover 10 m v200 were masked out, followed by manual visual screening. In post-processing, 82 of the 100x100 km grid cells from the 2019 v1 dataset were discarded (mostly false-positive coconut regions in Mexico, India, the Philippines, and Indonesia) while new hotspot grid cells (Andhra Pradesh in India, Congo Basin, Thailand, and Central/South America) were added to arrive at the 609 tiles.
- The six annual maps (2016-2021) were merged into a single static layer
prioritizing oil palm presence: a pixel is labeled industrial (
1) if classified as industrial in any year during 2016-2021; otherwise smallholder (2) if classified as smallholder in any year during 2016-2021; and other land cover (0) otherwise. - Planting Year (
planting_year, 30 m): Derived for all pixels mapped as oil palm (1or2) using 30 m Landsat-5, Landsat-7, and Landsat-8 Collection 2 Tier 1 Level-2 surface reflectance time series from 1990 to 2021. A Random Forest classifier labeled each clear Landsat observation as vegetation, bare land, or cloud to identify the Last Date of Land Preparation (LDLP)—the final bare-land observation corresponding to land clearing before canopy closure—smoothed with a spatial mode filter. The planting year is defined as the calendar year with the minimum(NIR - SWIR2) / (NIR + SWIR2)within the 5-year window preceding the LDLP (referred to as Normalized Difference Water Index,NDWI (Gao, 1996), in the paper, though computed withSWIR2such as Landsat-8 Bands 5 and 7, which is the standard Normalized Burn Ratio [NBR] formula and namedminNBRin the source GeoTIFF band description). In theplanting_yearband, values span1989to2022(rather than1990to2021): because the algorithm looks back up to 5 years before the LDLP, plantations established before 1990 (see the note below in limitations about left-censoring) are recorded as1989(~7.5% of oil palm pixels globally, the single most frequent value in the dataset), and a small fraction of pixels (~0.04%) have value2022.
Limitations
- Closed-canopy detection lag and 2019-2021 omission: Sentinel-1 SAR
reliably detects oil palm only after stands reach full canopy closure
(typically ~3 years after planting). Consequently, newly established young
plantations planted during 2019-2021 are largely omitted from the extent
layer. Pixels with
planting_yearin 2019-2021 primarily represent replanting (rotation) of older plantations that already had a closed canopy during 2016-2018 before being clear-cut, and must not be used to measure new oil palm expansion or deforestation between 2019 and 2021. - Last rotation vs. initial establishment and 1989 left-censoring:
Because the algorithm detects the last bare-land clearing event prior to
2021,
planting_yearrecords the establishment year of the current stand as of 2021. In plantations that have undergone rotation (typically every ~25 years), this reflects the most recent replanting year rather than initial conversion to oil palm. Additionally,planting_yearis left-censored at1989; this means pixels withplanting_year == 1989(~7.5% of oil palm pixels globally) represent plantations established in or before 1989 and should not be treated as literally planted in 1989 when calculating plantation age. - Post-clearing land-use conversion: Because the 2016-2021 merge rule retains any pixel detected as oil palm in at least one year, plantations clear-cut between 2016 and 2021 and subsequently converted to a different land use remain classified as oil palm.
- Omission of open-canopy and semi-wild oil palm: The CNN only maps closed-canopy stands and largely omits traditional, scattered, or subsistence semi-wild oil palm groves (particularly in West Africa, such as Nigeria) as well as stands with open canopies caused by fire, flooding, disease, or agroforestry intercropping.
- Lower accuracy in smallholder plantations: Across all 6,843 ground-truth planting-year observations, overall planting-year RMSE is 2.65 years (Mean Error = -0.24 years, R^2 = 0.86). Against visual interpretation of Landsat time series (n = 841 points), exact-year agreement of the Random Forest LDLP was only 59.6% overall (64.8% industrial, 44.0% smallholder; 76.3% within ±3 years). Smallholder plantations have lower classification accuracy (producer's accuracy 71.7%, user's accuracy 72.4%) and higher planting-year error (RMSE = 4.89 years, reaching 7.01 years in Cameroon) than industrial plantations (producer's accuracy 91.0%, user's accuracy 91.8%, planting-year RMSE = 2.02 years). Smallholder planting-year accuracy is degraded by mixed 30 m pixels along irregular plot boundaries, multi-year sequential planting, and "under-planting" (planting seedlings beneath old palms prior to felling, which avoids a bare-soil stage).
- Narrow geographic coverage of field validation: Field validation of
planting_yearhad narrow geographic coverage: only 3 countries for industrial plantations (Brazil, Gabon, Indonesia) and 2 countries for smallholders (Cameroon and Indonesia, 93.5% collected by a single company via owner interviews, which are also subject to owner recall error on older stands), with no field validation in Malaysia, Thailand, India, or other producing regions. - Resolution mismatch and grid offset between bands: Because
classificationis at 10 m andplanting_yearis at 30 m (with a 20 m West-edge origin offset between the source GeoTIFF grids), ~5% of 10 m oil palm pixels along plantation edges lack aplanting_yearvalue, and 30 mplanting_yearpixels along boundaries can spill onto adjacent 10 m non-oil-palm (0) pixels. - Regional artifacts from clouds and seasonality: Persistent cloud cover
and sparse pre-2013 Landsat archive coverage in Equatorial and West Africa
(e.g., Cameroon, Gabon, Congo Basin) increase planting-year uncertainty.
In regions with strong dry seasons (notably Andhra Pradesh in India,
Thailand, and Sulawesi in Indonesia), seasonal drops in NDWI can trigger
false bare-land detections, biasing
planting_yeartoward overly recent years.
Bands
Bands
Pixel size: varies
| Name | Min | Max | Pixel Size | Description |
|---|---|---|---|---|
classification |
10 meters | Oil palm plantation type classification |
||
planting_year |
1989 | 2022 | 30 meters | Estimated year of plantation establishment |
classification Class Table
| Value | Color | Description |
|---|---|---|
| 0 | #696969 | Other land covers that are not oil palm |
| 1 | #ff0000 | Industrial closed-canopy oil palm plantations |
| 2 | #ef00ff | Smallholder closed-canopy oil palm plantations |
Terms of Use
Terms of Use
Citations
Citations:
Descals, A., Gaveau, D. L., Wich, S., Szantoi, Z., and Meijaard, E. (2024). Global mapping of oil palm planting year from 1990 to 2021. Earth System Science Data, 16(11), 5111-5129. doi:10.5194/essd-16-5111-2024
DOIs
Explore with Earth Engine
Code Editor (JavaScript)
var dataset = ee.ImageCollection('BIOPAMA/GlobalOilPalm/v1_2'); // Band 1: Oil palm plantation type classification (10 m resolution) // 0: Other land covers that are not oil palm // 1: Industrial closed-canopy oil palm plantations // 2: Smallholder closed-canopy oil palm plantations var classification = dataset.select('classification').mosaic(); var classificationVis = { min: 0.0, max: 2.0, palette: ['696969', 'ff0000', 'ef00ff'], }; // Create a mask to add transparency to non-oil palm plantation class pixels. var mask = classification.neq(0); mask = mask.where(mask.eq(0), 0.6); // Band 2: Year of plantation establishment (1989-2022, 30 m resolution) var plantingYear = dataset.select('planting_year').mosaic(); var yearVis = { min: 1989.0, max: 2022.0, palette: ['ffffb2', 'fecc5c', 'fd8d3c', 'f03b20', 'bd0026'], }; Map.setCenter(101.7, 3.1, 10); Map.addLayer( classification.updateMask(mask), classificationVis, 'Oil Palm Classification'); Map.addLayer(plantingYear, yearVis, 'Planting Year (1989-2022)', false);