Extensive Unreported Non-Plantation Oil Palm in Africa

1. Simple Random Sampling for Country-Wide Estimation of Oil Palm Area

1.1. Delimitation of the Sampling Area

We defined a sampling area based on the potential oil palm distribution delimited in a previous study [18]. This potential distribution represents regions where bioclimatic variables indicate that the climate is suitable for oil palm cultivation. Within this potential distribution, we excluded areas where oil palm occurrence is unlikely by masking non-vegetated land cover, including water, bare land, and impervious surfaces in the European Space Agency WorldCover 2021 v2 land cover map. This mask aimed to constrain the sampling area and reduce uncertainty in the area estimates. Then, we subdivided the potential distribution into primary forests, inhabited and uninhabited areas. First, we defined the primary forests using the JRC Tropical Moist Forest (TMF) dataset [19], which defines primary forests as undisturbed evergreen or semi-evergreen forest, with no signs of degradation or deforestation, as observed in the Landsat data. Then, outside the primary forest, we defined inhabited areas as pixels within a 100-meter radius of those classified as inhabited in any of two datasets, the Global Human Settlement Layer and the World Settlement Footprint 2015. Areas outside the inhabited buffer zone were classified as uninhabited. In addition, we buffered the potential oil palm distribution area by 100 kilometers and extended the buffering area in West Africa across the coast up to Senegal, where non-plantation oil palm has been reported [5]. In summary, the sampling area consisted of four classes: 1) primary forest, 2) inhabited and 3) uninhabited areas within the potential distribution, and 4) buffering area of the potential distribution. The total sampled area was 618.83 Mha.

The purpose of subdividing the sampling area was to use different sampling densities. Inhabited areas are more likely to contain human-managed oil palm than uninhabited areas, and primary forests and areas outside the potential distribution are expected to have less oil palm. Thus, we applied a higher sampling density to inhabited areas. This varied sampling density aimed to reduce uncertainty in area estimates in regions with higher oil palm coverage. The minimum sampling density per country was set to 20 points per Mha in inhabited areas, 10 points per Mha in uninhabited areas and buffered areas of the potential distribution, and 5 points per Mha in primary forests.

Despite the extensive study area, we acknowledge the existence of oil palm in Africa outside the sampled area. For example, we did not include isolated oil palm populations that have been reported in Angola, Zambia, and Mauritania [1]. These regions represent sparse populations with a negligible oil palm area compared to the areas found in the potential distribution of oil palm.

1.2. Definition of oil Palm Types

We defined plantation and non-plantation oil palm based on spatial features that can be recognized in sub-meter satellite images. While plantation oil palm is a specific type of plantation where oil palm is intensively managed, non-plantation oil palm includes a heterogeneous group of oil palm types:

1)

Plantation oil palm: Structured monoculture plantations with homogeneous oil palm age. The plantation has a regular planting pattern and optimal spacing to maximize productivity. These plantations can be corporate-managed (industrial plantations) or owned by small businesses, families or local communities (smallholder plantations). Plantation oil palm can be identified in sub-meter satellite images because of the high planting density and a linear arrangement of palms (Supplementary Figure S2). While industrial plantations show similar characteristics such as large coverage, long linear well-defined boundaries, and dense and equidistant trails placed for optimal harvesting, smallholders are a heterogeneous group with different planting settings. Here, in this category, we include any smallholder contexts with a systematic layout, planting rows, and standardized spacing.

2)

Non-plantation oil palm: oil palm that grows dispersed or disorganized in clusters, with irregular arrangement and lack of strict planting patterns (Supplementary Figure S2). This includes: a) wild oil palm, which grows in naturalized settings coexisting with native vegetation or other plant species, creating a mixed structure that resembles a natural vegetation type rather than a cultivated plantation. This includes truly wild oil palm which was mapped elsewhere [14]; b) smallholder or community-based oil palms that grow close together but lack systematic layout, planting rows, or standardized spacing. In these groves, oil palms grow in irregular patterns due to spontaneous seedling emergence or informal planting methods without planned organization; c) Dispersed oil palms in anthropogenic landscapes: individual or scattered oil palm growing within human-modified environments, such as agricultural fields, villages, roadsides, or mixed-use lands. Our definition of non-plantation oil palm does not distinguish between land owned and managed by smallholders and land where oil palm grows entirely wild, as this information cannot be directly determined from satellite data.

1.3. Visual Interpretation of Satellite Data

Oil palm can be distinguished from other palm and plantation types based on specific features visible in sub-meter resolution satellite images. To identify oil palm and differentiate it from other species, we collected 175 sub-meter resolution images paired with Google Street View data. Supplementary Figure S3 provides examples from six locations. First, we determined the palm or crop type in the 175 locations using Google Street View. Then, we analyzed the sub-meter resolution satellite images and qualitatively identified the key features that distinguish oil palm from other palms and crop types.

Sub-meter resolution satellite images capture the crowns of palms and trees. Palms typically appear as asterisk-like shapes when viewed from above (Supplementary Figure S3), formed by fronds growing outward in all directions from the top of the trunk. For oil palms, these fronds are usually longer than those of other palms, such as coconut or areca palms. Long fronds are the first indicator of an oil palm, although this is less reliable for older oil palms, which grow taller with shorter fronds. Other palms, such as Raphia sp. and Phoenix sp., also have long fronds, but their leaflets are smaller and thinner compared to oil palm. Thus, the second distinguishing feature of oil palm is broader, well-defined fronds visible in sub-meter resolution satellite images. Fronds of Raphia sp. and Phoenix sp. are often indistinct in such images. However, this is not a definitive indicator, as some Google Earth images are slightly blurred, making the fronds also indistinguishable in oil palm. The color of satellite imagery can also help distinguish oil palms, although it is the least reliable indicator due to variations in Google Earth imagery. Oil palms typically appear dark green, although young oil palms often look brighter. In comparison, other palms show brighter green hues.

Plantation and non-plantation oil palm can be identified based on the spatial arrangement of the oil palms (Supplementary Figure S2). Plantation oil palm presents a homogeneous density and is typically planted in rows using a triangular spacing arrangement [1]. Rectangular spacing is also used in oil palm, although this spacing type is more common in coconut plantations. Non-plantation oil palm, on the contrary, does not follow a strict plantation pattern and occurs at variable densities or scattered across the land.

We visually inspected sub-meter resolution satellite images in Google Earth to determine if oil palm was present at the sampled points. We classified a point as having oil palm only if the point was located within the oil palm canopy area. The oil palm canopy area is the ground area covered by the vertical projection of the oil palm fronds. If sparse oil palms were present but the point didn’t fall within the canopy area, we considered that oil palm was not present in the sampled point. If the point fell within the oil palm canopy area, we further categorized the point as plantation or non-plantation oil palm. This approach ensured that the oil palm area estimates represented the canopy area rather than areas with varying palm densities.

The sub-meter resolution images available on Google Earth are updated regularly, but there are regions where images are not available. When the imagery was unavailable for a given point, we labelled the point as ‘no satellite data’ and considered the points as class ‘Other’ in the subsequent analyses. We did this reclassification because most points without satellite data were in isolated, forested areas with a low likelihood of containing oil palm. When sub-meter satellite images are available, they show different years depending on the location. For this reason, our estimates reflect the oil palm area over the past few years rather than a specific year. This uncertainty is minimal because oil palm is perennial, and its spatial extent does not exhibit substantial year-to-year variation.

A single interpreter conducted the visual interpretation of satellite images, performing the task independently five times to ensure consistency. The final assigned class for each point was determined using a majority vote among the results from the five replications.

1.4. Estimation of Oil Palm Canopy Area

We used the Wilson score to estimate the proportion of points with plantation and non-plantation oil palm. The Wilson score provides accurate approximations to the binomial distribution when dealing with low proportions [20], which is the case in our study; the proportion of points labelled as oil palm was below 5% in many countries. The center of the Wilson score interval $\check{p}$ was calculated with Equation (1):

$$\check{p}={\displaystyle \frac{p+\frac{z^2}{2n}}{1+\frac{z^2}{n}}}$$

(1)

where p is the observed proportion of points with oil palm (either plantation or non-plantation oil palm separately), n is the total number of points sampled, and z is the critical value from the standard normal distribution for a desired confidence interval. In this study, we used a 95% confidence interval and, thus, a z value of 1.96. The formula for the Wilson Score Interval (Equation (2)) provides the confidence interval of $\check{p}$.

$$CI={\displaystyle \frac{\check{p}+\frac{z^2}{2n}\pm z\sqrt{\frac{\check{p}(1-\check{p})}{n}+\frac{z^2}{{4n}^2}}}{1+\frac{z^2}{n}}}$$

(2)

The oil palm area estimate (AE) was calculated by multiplying the total sampled area by the proportion of oil palm (Equation (3)). Similarly, the confidence interval of the oil palm area estimate (${CI}_{AE}$) was obtained by multiplying the total sampled area by the confidence interval of the proportion of oil palm (Equation (4)). Sampled points were classified as oil palm only if they fell within the oil palm canopy. Thus, our area estimates represent the oil palm canopy area, defined as the ground area covered by the vertical projection of the oil palm fronds.

$$AE=Totalsampledarea\times \check{p}$$

(3)

$${CI}_{AE}=Totalsampledarea\times CI$$

(4)

We used different density of points that differed by country and sampling area classes. Thus, the estimation of the oil palm area (Equation (1) and 3) and its confidence interval (Equation (2) and 4) were calculated separately for each country and sampling area class. To obtain country-wide area estimates (AEcountry), we summed the oil palm area estimates for the inhabited area (AEinhab), uninhabited area (AEuninhab), buffered potential distribution (AEbuffer), and primary forest (AEforest) (Equation (5)). For the confidence intervals, we used the propagation of uncertainty to aggregate the uncertainty at the country scale (Equation (6)):

$$AEcountry=AEinhab+AEuninhab+AEbuffer+AEforest$$

(5)

$$CItotal=z\sqrt{{\displaystyle \frac{CIinhab}{z^2}}+{\displaystyle \frac{CIuninhab}{z^2}}+{\displaystyle \frac{CIbuffer}{z^2}}+{\displaystyle \frac{CIforest}{z^2}}}$$

(6)

where CIinhab, CIuninhab, CIbuffer, and CIforest represent the confidence intervals for the inhabited area, uninhabited area, buffered potential distribution, and primary forest for a given country. CItotal represents the confidence interval of the country-wide area estimate (AEcountry).

1.5. Comparison with Official and Satellite-Derived Statistics

We compared our oil palm area estimates with official statistics from FAOSTAT and USDA Foreign Agricultural Service, as well as satellite-derived estimates from a global oil palm map [10]. We extracted the oil palm area from these datasets for the year 2021. These datasets have different definitions of what constitutes oil palm. The official statistics from FAO and USDA include oil palm areas that have been harvested. In contrast, the global oil palm map only includes closed-canopy monoculture oil palm plantations, whether it was harvested or not, but excludes young plantations and oil palm that grow sparsely. For Nigeria, we also compared our estimates with previous studies on plantation and non-plantation oil palm (Supplementary Text S1).

2. Determining the Occurrence of Non-Plantation Oil Palm Within the Proximities of Villages

The second approach focused on determining the proportion of villages with non-plantation oil palm growing in their surrounding areas. We mapped the surroundings of the villages using OpenStreetMap (OSM) data. Specifically, we used all settlement boundaries in West and Central Africa included in OSM. To enhance the reliability of the village boundaries, we also used the Open Buildings 2.5D Temporal Dataset and excluded the village boundaries with fewer than 10 buildings or more than 1,000 buildings.

Because the density of villages varies across Africa, we created a 100 x 100 km grid and randomly selected 100 villages within each grid cell. This approach prevented oversampling areas with high village densities. Finally, from the selected villages, we randomly selected 400 village boundaries in West Africa and 700 in Central Africa. The boundaries in Central Africa were subdivided into 500 within the Congo forest and 200 outside it. The villages in the Congo forest were selected based on the region with tropical & subtropical moist broadleaf forests in the RESOLVE Ecoregion dataset. For each selected village, we visually inspected sub-meter resolution imagery within the village boundary and within a 100-meter buffer surrounding the boundary. We considered that oil palm was present if it appeared in natural groves, clustered formations, around houses, or in non-plantation settings. Supplementary Figure S4 shows a village with non-plantation oil palm. In cases where oil palms appeared scattered, we considered that the village contained oil palm if there were at least five palms within the inspected area.

This analysis focused on West and Central Africa, excluding East Africa for two reasons. First, the oil palm area in East African countries, based on the sampling points, was much smaller than in West and Central Africa. This suggests that oil palm is less relevant in East Africa, although it may have local importance in regions like Kigoma in western Tanzania. Second, preliminary analysis indicated that oil palm is absent around villages in East Africa. We observed, however, a high occurrence of coconut palms in villages along the coasts of Tanzania and Mozambique.