Increasing global demand of vegetable oils and biofuels
results in significant oil palm expansion in southeastern Asia, predominately
in Malaysia and Indonesia. The land conversion to oil palm plantations has posed
risks to deforestation (50 % of the oil palm was taken from forest during
1990–2005; Koh and Wilcove, 2008), loss of biodiversity and greenhouse
gas emission over the past decades. Quantifying the consequences of oil palm
expansion requires fine-scale and frequently updated datasets of land cover
dynamics. Previous studies focused on total changes for a multi-year
interval without identifying the exact time of conversion, causing
uncertainty in the timing of carbon emission estimates from land cover
change. Using Advanced Land Observing Satellite (ALOS) Phased Array type
L-band Synthetic Aperture Radar (PALSAR), ALOS-2 PALSAR-2 and Moderate
Resolution Imaging Spectroradiometer (MODIS) datasets, we produced an annual
oil palm area dataset (AOPD) at 100 m resolution in Malaysia and
Indonesia from 2001 to 2016. We first mapped the oil palm extent using
PALSAR and PALSAR-2
data for 2007–2010 and 2015–2016 and then applied a
disturbance and recovery algorithm (Breaks For Additive Season and Trend – BFAST) to detect land cover change
time points using MODIS data during the years without PALSAR data (2011–2014
and 2001–2006). The new oil palm land cover maps are assessed to have an
accuracy of 86.61 % in the mapping step (2007–2010 and 2015–2016). During
the intervening years when MODIS data are used, 75.74 % of the detected change
time matched the timing of actual conversion using Google Earth and
Landsat images. The AOPD revealed spatiotemporal oil palm dynamics
every year and shows that plantations expanded from 2.59 to 6.39
The global demand for vegetable oil and its derivative products calls for an
increase in palm oil production, leading to oil palm expansion and
intensification in southeastern Asia (Sayer et al., 2012).
According to the Food and Agriculture Organization (FAO), Malaysia and
Indonesia account for 81.90 % of the global oil palm fruit production in
2017, an increase of 179.72 % from 2000 to 2017 (see
Quantifying the spatiotemporal details of oil palm expansion is important for understanding the deforestation process and its impacts on ecosystem services and promotes progress in environmental governance and policy decisions (Gibbs et al., 2010; Koh and Wilcove, 2008). However, annual information on the expansion of oil palm plantations is poorly documented in Malaysia and Indonesia. The statistical records (e.g. FAO, United States Department of Agriculture – USDA) give neither the detailed spatial distribution nor the young oil palm trees and smallholder plantations. Many efforts have been made to characterize the oil palm extent (Cheng et al., 2018; Gaveau et al., 2016; Miettinen et al., 2017). For example, the Roundtable on Sustainable Palm Oil (RSPO), whose members manage one-third of the world's oil palm, provided spatial information on oil palm distribution in Malaysia and Indonesia (Gunarso et al., 2013). The continuous mapping of oil palm on peatland in 1990, 2000, 2007 and 2010 described the dynamic change of oil palm on peat during the past 30 years (Miettinen et al., 2012). But these maps are given for a certain year or several time phases without capturing the exact time of oil palm changes. Dynamic global vegetation models use gross land-use change and thus require high-resolution grid-cell-based annual oil palm conversion maps rather than country-level inventories and bi-decadal land cover maps (Yue et al., 2018a, b). A lack of continuous change information may cause a wrong interpretation of land cover change time and significant bias in global carbon dynamic studies (Zhao and Liu, 2014; Zhao et al., 2009). As a result, oil palm plantation maps at high temporal and spatial resolutions in Malaysia and Indonesia are urgently needed.
Remote sensing has been used in oil palm monitoring since the 1990s. Progress has been made in oil palm mapping and change detection, including (1) data sources from optical satellite earth observations (Lee et al., 2016; Srestasathiern and Rakwatin, 2014) to microwave datasets such as the Phased Array type L-band Synthetic Aperture Radar (PALSAR; Cheng et al., 2018; Dong et al., 2015), (2) spatiotemporal resolutions from the regional to national scale (Miettinen et al., 2017) and from single to multi-decadal mapping (Gaveau et al., 2016; Miettinen et al., 2016), (3) interpretation methods from manual to semi-automatic and fully automatic identification (Baklanov et al., 2018; Cheng et al., 2019; W. Li et al., 2017; Mubin et al., 2019; Ordway et al., 2019), and (4) products going from oil palm land cover maps to more detailed datasets on plantation structure, e.g. tree counting (Li et al., 2019; Cheang et al., 2017) age, yield estimation (Balasundram et al., 2013; Tan et al., 2013), etc. A few studies also focused on the continuous oil palm change detection (Carlson et al., 2013; Gaveau et al., 2016; Vijay et al., 2018). These studies adopted visual or semi-automatic interpretation for oil palm plantations, which is labour-intensive and not appropriate for long-term annual oil palm plantation monitoring. Automatic identification can overcome this difficulty by using classification algorithms based on Landsat and PALSAR and PALSAR-2 data, which were successfully applied to produce the 2015 land cover map of insular southeastern Asia with discrimination of oil palm plantations (Miettinen et al., 2017). So far, however, the annual dynamics of oil palm plantations (expansion and shrinkage) remains unquantified for Malaysia and Indonesia.
The annual oil palm mapping in tropical areas such as insular southeastern Asia is a challenge due to the persistent cloudy conditions (Gong et al.,
2013; Yu et al., 2013). Multi-temporal optical images can help reduce cloud
effects (Yu et al., 2013), but it is still difficult to obtain
effective optical observations in Malaysia and Indonesia (51.88 % of the
region is without annual Landsat images; Fig. S1 in the Supplement). Microwave remote
sensing is not affected by clouds and is considered to be the most
efficient source in separating forested vegetation and oil palms
(Ibharim et al., 2015; Teng et al., 2015). The long time span of 25 m
resolution PALSAR and PALSAR-2 data provides opportunities for mapping oil palm
at high spatiotemporal resolutions. Recently the PALSAR and PALSAR-2 data have
been successfully used in characterizing oil palm change for the whole country of
Malaysia for 6 years using PALSAR (2007–2010) and PALSAR-2 (2015–2016; Cheng et al., 2019). However, the gap years (2011–2014) between
PALSAR and PALSAR-2 hampered continuous tracking of oil palm plantation
dynamics. One potential way to achieve annual mapping is to use optical
earth observation data, e.g. Landsat images for the PALSAR gap period
(Chen et al., 2018; Shen et al., 2019). However, this requires abundant
Landsat images (
The objectives of this study are (i) to develop a robust and consistent approach capable of detecting annual oil palm changes in southeastern Asia using multiple remote-sensing datasets based on image classification and break point detection, (ii) to produce a nominal 100 m annual oil palm area dataset (AOPD) in Malaysia and Indonesia from 2001 to 2016, and (iii) to quantify the spatial and temporal patterns of oil palm change dynamics since 2001. Specifically, we developed the annual oil palm plantation dataset in Malaysia and Indonesia by using a two-stage method. The first step is random forest-based image classification using PALSAR during 2007–2010 and PALSAR-2 data during 2015–2016 (the periods with PALSAR and PALSAR-2 data available). Combined with the oil palm maps produced in the first step during the years with PALSAR coverage, MODIS NDVI was used in a change-detection algorithm called Breaks For Additive Seasonal and Trend (BFAST; Verbesselt et al., 2010a) to fill the data-gap years (2011–2014) outside the PALSAR years and extend the oil palm land cover mapping period back to 2001. Oil palm in this study refers to both young and mature oil palm trees from industrial plantations and smallholders with the minimum size of 1 ha (oil palm smallholders are defined as 50 ha or less of cultivated land producing palm oil controlled by smallholder farmers – the definition used by the RSPO – with an average of 2 ha; World Bank; 2010).
Insular southeastern Asia was originally occupied by evergreen moist tropical forest, which is one of the most biologically diverse terrestrial ecosystems on Earth. The natural environment, with humid tropical climates and low-lying topography, is suitable for the oil palm (Elaeis guineensis; Fitzherbert et al., 2008). Since 1911, when the first commercial oil palm plantation in southeastern Asia was founded in Sumatra, oil palm plantations expanded rapidly in Sumatra and peninsular Malaysia and then spread to Sarawak and Sabah in Malaysia and Kalimantan in Indonesia (Corley and Tinker, 2008). Industrial oil palm plantations spurred the economic sectors in southeastern Asian countries but also raised concerns on the negative social and environmental impacts (Obidzinski et al., 2012; Sayer et al., 2012). Recently, oil palm plantation expansion became one of the dominant drivers of deforestation in Malaysia and Indonesia (Austin et al., 2018; Gaveau et al., 2016). Thus, we chose as a study area the whole area of Malaysia and Sumatra and Kalimantan in Indonesia, which encompass 96 % of the total oil palm production in Indonesia (Petrenko et al., 2016). Oil palm plantations in these two countries account for 67.51 % of world's total oil palm plantation area (FAOSTAT, 2017), and dramatic land cover conversion happened in this region due to human-induced modifications.
The development of AOPD includes two major stages: (1) oil palm mapping using PALSAR and PALSAR-2 data (Sect. 2.3) and (2) change-detection-based oil palm updating using MODIS NDVI during the gap years in operation between Advanced Land Observing Satellite (ALOS) and ALOS-2 (Sect. 2.4). The first stage aimed at producing the oil palm maps for 2007, 2008, 2009 and 2010 using PALSAR and 2015 and 2016 using PALSAR-2 datasets. The detailed procedures include the pre-processing of the original PALSAR and PALSAR-2 data, training sample collection and image classification, and final production of oil palm maps for the target years after post-processing using ancillary datasets. In the second stage, we combined oil palm maps produced in the first stage with MODIS NDVI data. Time series of MODIS NDVI data and change maps were prepared in the data preparation step, followed by the break point test using the change-detection algorithm, BFAST, to detect the change year (change from other land cover types to oil palm and the reverse) in the PALSAR and PALSAR-2 missing-data period. After the post-processing, we derived the oil palm maps in these gap years and traced the oil palm distribution back to 2001. Combining the results from the two stages, we obtained the annual oil palm plantation maps from 2001 to 2016 at 100 m spatial resolution, forming the AOPD. The whole workflow is shown in Fig. 1.
Workflow of the annual oil palm mapping procedure. Stage 1 stands for oil palm mapping using PALSAR and PALSAR-2 data, and Stage 2 stands for change-detection-based oil palm maps updated using MODIS NDVI.
We used multi-source remote-sensing images to fully cover the study
area; these included images from ALOS PALSAR, ALOS-2 PALSAR-2 and MODIS NDVI. The Landsat
archives were not used because of the low data availability in this region
caused by frequent thick cloud cover (Fig. S1 in the Supplement).
The Japan Aerospace Exploration Agency (JAXA) provided the 25 m resolution
global PALSAR and PALSAR-2 mosaic by mosaicking SAR images of the backscattering
coefficient (
Although ALOS PALSAR and ALOS-2 PALSAR-2 have different satellite microwave sensor properties (e.g. frequency, off-nadir angle), the backscatter signals are relatively stable for the given period (2007–2010 and 2015–2016), as seen by comparing the distribution of backscattering values (HH and HV) of 250 000 randomly generated pixels (using ArcGIS 10.3) in the study area between different years (see Fig. S2). The similar findings for the stability of PALSAR and PALSAR-2 data were also given in previous studies (Cheng et al., 2019; Qin et al., 2017). Meanwhile, the HH and HV values for oil palm and forest are also shown in Fig. S3 and indicate the separability between the two land cover types for both PALSAR and PALSAR-2 data. Therefore, the consistency between ALOS PALSAR and ALOS-2 PALSAR-2 allows tracking the oil palm changes in the study period. One problem of using PALSAR and PALSAR-2 data, however, is the “salt-and-pepper” noise (Zhang et al., 2019), which may cause misclassification and false changes in the subsequent process. Previous studies showed that the resampling method reached higher accuracy and better visual results in oil palm mapping compared to the commonly used filter method (Cheng et al., 2018). The identification and area estimation of oil palm plantations have also been proven to perform better at 100 m resolution (Cheng et al., 2018). Therefore, we resampled the original 25 m PALSAR and PALSAR-2 images to 100 m resolution for every year to reduce salt-and-pepper noise.
In this study, a multi-year training sample set (2007–2010, 2015 and 2016)
was used to map the oil palm extent in Indonesia and Malaysia from 2007 to
2016. We used the training sample set for Malaysia from our previous study
(Cheng et al., 2017) and interpreted the training datasets for
Indonesia using the same interpretation method. The sample collection was
mainly based on the high-resolution (
The distribution of training data (unit: pixel). Malay.: Malaysia. Indon.: Indonesia.
Thereafter, we used a random forest (RF) classifier in the image
classification step. The HH and HV digital number of decibels and the derived
difference (HH
Post-processing after the initial results is necessary because of the limitation in the training set, unavoidable classification errors and the difficulty in describing heterogeneous real land surface. To obtain a reliable oil palm dataset, we adopted several steps, including mode filtering, terrain filtering, and intact forest and mangrove filtering in post-processing to improve the final oil palm maps in Stage 1 for 2007, 2008, 2009, 2010, 2015 and 2016.
Mode filtering is used for the very small patches (mainly single
pixels) in the initial results, since it is more likely to be errors or noise
induced by PALSAR and PALSAR-2 data rather than real oil palm plantations. The
topographic factors such as slope and elevation will cause the confusion of
backscattering signals from satellite sensors, particularly in area with
undulating terrain. Therefore, we applied terrain filtering to reduce the
confusion by the topographic factor using the Shuttle Radar Topography Mission
(SRTM) 30 m digital elevation model (DEM). The altitude threshold of 1000 m
was applied, since the oil palm is mainly distributed in the lowlands (mostly
Another problem when developing oil palm maps is the replantation of oil palm trees. Oil palm has a long life cycle of 25 to 30 years. After that, the trees will be cleared and replaced because of a decrease in palm oil yield (Röll et al., 2015). However, from the satellite observations, the land cover type is bare land at the time of oil palm logging, whereas the land-use property remains unchanged as oil palm plantations when checked before and after oil palm logging. Given the limitation of satellite observation, we provided two versions of our oil palm datasets. The first version is the oil palm datasets after the post-processing mentioned above. Here replantation is not considered, and this version includes conversion from other land cover types to oil palm (oil palm expansion) as well as the opposite (oil palm shrinkage). In the second version, we assumed that oil palm expansion is a unidirectional activity due to the growing demand of palm oil. The time-series filtering was conducted by using the 2007 oil palm extent to filter all pixels classified as “non-oil palm” in the subsequent years. As a result, this version of the oil palm plantation dataset has shown continuously expanding areas from 2007 to 2016. The second version includes the impact of oil palm replantation and the thriving oil palm industry in southeastern Asian countries but ignored any possible decrease in oil palm (e.g. abandonment, conversion to cropland) in some areas.
MODIS NDVI is an important index of vegetation conditions and has been
widely used in vegetation and land cover change studies (Clark et al.,
2010; Ding et al., 2016; Estel et al., 2015). NDVI in the recent updated
MODIS vegetation index data (MOD13Q1) collection 6 from 2000 to 2007 and from
2010 to 2015 (downloaded from
A change map for the microwave data gap period between PALSAR and PALSAR-2 (2011–2014) was extracted using the change pixels in 2010 and 2015 oil palm maps with spatial locations and “from–to” types. Here, we assumed that the change from classification was reliable because of the high resolution of PALSAR data. We then sought the exact change year within the intervals in the next step (Sect. 2.4.2) using temporal NDVI files extracted from each change pixel. Frequent changes such as two or three shifts during the gap years were assumed to be of low probability and thus not considered in this study. For the period during 2001–2006 without PALSAR and PALSAR-2 data and oil palm distribution in 2000, we assumed a unidirectional expansion of oil palm, and the oil palm extent in 2007 was used as the potential change regions in the next step. In total, we derived two versions of change maps (one with bi-directional change and the other with only unidirectional oil palm expansion) for the two periods.
Change-detection analysis was conducted in the change pixels derived from the last step to identify the exact change time within the two periods (2011–2014 and 2001–2006) based on the time series MODIS NDVI from 2010 to 2015 and 2000 to 2007, respectively. Here we aimed to capture abrupt NDVI changes (break points) in the two given periods, which are assumed to be caused by the conversion of the original land cover type to the oil palm cultivation. Many change-detection algorithms and their derivatives have been developed in recent years to detect subtle or abrupt changes in dense time-series satellite profiles (Broich et al., 2011; Kennedy et al., 2010; Verbesselt et al., 2010b). Most of these algorithms were applied in forest change monitoring, and all reach high consistency in detecting significant change (Cohen et al., 2017). A recent algorithm, the Bayesian estimator of abrupt change, seasonal change and trend (BEAST), aggregating the competing models and then the conventional single best model, performed well in capturing multiple and subtle phenological changes (Y. Zhao et al., 2019b). Here we used BFAST to capture the oil palm conversion time within the two study periods (2011–2014 and 2001–2006).
BFAST has been successfully applied in monitoring forest disturbance and
regrowth and has proved robust with different sensors (DeVries et al.,
2015; Verbesselt et al., 2012). Based on the structural change methods, the
BFAST algorithm is able to find the structural break point between different
segments in the observation time series (DeVries et al., 2015)
and thus can be used to detect the time and number of abrupt or gradual
changes as well as to characterize the magnitude and direction. The BFAST
method decomposes the time series into trend, seasonality and residual
sections (Verbesselt et al., 2010b). The model can be
expressed as
An ordinary least-square-residual-based moving-sum test (OLS-MOSUM; Zeileis, 2005)
was used to test whether break points occurred in the trend or seasonal
components. Then, the test was conducted to determine the number and optimal
position of the breaks using the Bayesian information criterion (BIC) and the minimum
of the residual sum of squares. The trend and seasonal coefficients were
then computed using robust regression. A harmonic seasonality model (with
three harmonic terms) was used to describe the seasonality of the satellite
data (Eq. 6; Verbesselt et al., 2010b). For each
piecewise linear model (
For the
Examples of the break point detection in the MODIS time series
using the BFAST algorithm.
The previous steps generated annual oil palm maps for 6 years (Sect. 2.3) and the oil palm change time in the missing periods (2011–2014 and
2001–2016; Sect. 2.4.1 and 2.4.2). In the final step, all these data were
combined to update the continuous oil palm dataset from 2001 to 2016
following Xu et al. (2020).
For the gap period from 2011 to 2014, the oil palm updating was based on the
from–to land cover types (
Our product of annual oil palm maps, AOPD, was evaluated for three aspects: (1) the independent annual oil palm sample set for Malaysia (2007, 2008, 2009, 2010, 2015 and 2016) and Indonesia (2010–2016) to evaluate the annual mapping results for the classified maps using PALSAR and PALSAR-2 data and gap years using the change-detection method, (2) a change sample set aimed at assessing the accuracy of detected change years, and (3) comparison with statistical inventories (e.g. FAO, USDA, Malaysian Palm Oil Board – MPOB – 2011–2016, Badan Pusat Statistik – BPS-Statistics Indonesia – 2011–2015, the existing oil palm maps from Gaveau et al., 2016, and the Landsat-based deforestation maps – Hansen et al., 2013). FAO and USDA agricultural statistical data provided the harvested area of oil palm using data collected by official and unofficial outlets. MPOB is a government agency providing oil palm plantation area in Malaysia based on the data reported by state agencies, institutions, private estates and independent smallholders. BPS-Statistics Indonesia, a non-ministry government agency, provided statistical data for the public including oil palm plantation area compiled from the quarterly (SKB17-Oil Palm) and annual (SKB17-Annual) plantation estate survey, custom documents from the Directorate General of Customs, and secondary data from the Directorate General of Estate Crops.
Two sets of annual oil palm samples set were used to validate the mapping
results in Malaysia and Indonesia according to the sampling protocol of
Gong et al. (2013). The independent annual sample set in Malaysia was
from the previous studies (Cheng et al., 2019, 2017). All
pixel-based samples were randomly produced in an equal-area hexagonal grid
(95.98 km
Spatial distribution of oil palm samples in the two validation datasets. The annual sample set contains 2986 (in 2016) samples in Malaysia, which were interpreted for 2007, 2008, 2009, 2010, 2015 and 2016, and 7667 (in 2016) samples in Indonesia, interpreted from 2010 to 2016. These samples were used to validate the annual maps developed from PALSAR and PALSAR-2 data. Of the annual sample set in Malaysia, oil palm samples consist of 16.92 % (505), while the forest, water and others consist of 78.16 %, 2.48 % and 2.44 %, respectively. The Indonesian annual sample set contains 601 (7.84 %) oil palm samples, and the rest (92.16 %) were other types. The change sample set includes 370 oil palm samples which were converted in the interpolated period (2001–2006 and 2011–2014). This sample set, with change year labelled, is used to assess the change detection result in the gap years.
The distribution of annual validation sample set for Malaysia and Indonesia (unit: pixel).
The change sample set was developed to evaluate the detected change year by
the break point detection analysis. Time lapses of high-resolution imagery
from Google Earth covering the change period were used to check the change
time detected by the BFAST algorithm. We randomly selected 5000 points
(implemented with ArcGIS 10.3 software) in the change area, but there were
only limited samples (370, 25.07 % of the total 1476 oil palm samples)
with continuous high-resolution images from Google Earth and cloud-free
Landsat time series. We compared our detected change years with the actual
oil palm conversion time for these test samples. A confidence interval of
The annual changes of oil palm plantations from 2001 to 2016 are shown in
Fig. 4. The spatial and temporal dynamics of oil palm changes vary in
Malaysia and Indonesia. In the study area, most oil palm plantations are
located in lowland areas (elevation
Year of oil palm change at 100 m resolution in the study area from
2002 to 2016.
Light colours in Fig. 4 indicate the oil palm changes (expansion and shrink) in early years, while the dark colours are the changes in more recent years. Oil palm plantations expanded rapidly during the study period in peninsular Malaysia and Sumatra and Borneo. In Indonesia, rapid expansion first occurred in Sumatra and was then surpassed by Kalimantan (Gunarso et al., 2013; Petrenko et al., 2016). This can also be observed in our maps, where more changes happened in earlier years in Sumatra (lighter colours in Fig. 4) and later in Kalimantan (darker colours). The decrease in oil palm plantations was also detected (Fig. 4b), although it is difficult to separate the oil palm replantation after one rotation (i.e. still oil palm in land use) from the permanent oil palm loss (i.e. change to other land-use types). Compared to the period before 2007 using change detection in NDVI data, our data product in the gap period of 2011–2014 would be of higher quality, since the net changes were constrained by the oil palm maps in 2010 and 2015 derived from PALSAR and PALSAR-2 data, respectively.
Figure 5 displays the annual total area of oil palm in Malaysia and Indonesia from 2001 to 2016, with uncertainty ranges (shaded area with boundary lines) during 2001–2006 and 2011–2014. This uncertainty range is from the change-detection step; 9.45 % of the total changes from 2010 to 2015 were not captured in the MODIS NDVI using the BFAST algorithm because of the low resolution, cloud contamination, the mapping error from the base maps, etc. Assuming that these missing changes all happened from 2010 to 2011, the oil palm area of the gap years should follow the trajectory of the upper boundary line. If all the missing changes happened in the last year of the period, the oil palm area curve would be the lower boundary line. Since the distribution of oil palm in 2001 was unknown, large uncertainty may exist before 2007. Here, the uncertainty range during 2001–2006 was determined based on the data availability of MODIS NDVI and consistency of change time detection from the quality maps (Figs. S6 and S7). The oil palm area before 2007 follows the upper boundary curve if the same breaks were detected in all three structural change methods (OLS-MOSUM, SupLM, BIC) and there are more than 60 % valid NDVI values available in this time period. If all the breaks were counted regardless of the number of valid MODIS NDVI values and the consistency of change methods, the oil palm area would be the lower boundary line.
Comparison of the annual oil palm plantation area among FAO and
USDA statistics, MPOB records for Malaysia, BPS-Statistics and oil palm
concessions from GFW for Indonesia and our mapping results in
Generally, the net oil palm plantation area shows a monotonous increasing
trend from 2001 to 2016 for Malaysia (Fig. 5a) and Indonesia (Fig. 5b)
in both the bi-directional (green lines) and unidirectional (blue lines)
versions. During the past 16 years, the net oil palm area across Malaysia
increased from
The mapping performance of AOPD was evaluated first using an independent annual
oil palm sample set for 2007, 2008, 2009, 2010, 2015 and 2016. The mapping
accuracy from the previously developed datasets over Malaysia
(Cheng et al., 2019) was also compared. The results of the annual
accuracy (
The comparison of the oil palm accuracy between our mapping results and Cheng et al. (2019) for the 6 mapping years in Malaysia. UA: user's accuracy. PA: producer's accuracy.
The oil palm accuracy in Indonesia from 2010–2016. UA: user's accuracy. PA: producer's accuracy.
Figure 6 shows the direct comparison of the change maps with the images from
Google Earth and Landsat, which document the change process. We use time
lapse of images when the annual high-resolution images from Google Earth
were not available. Here time lapse means the images obtained for intervals of
Visual comparison of the detected change years with the
high-resolution images and medium-resolution Landsat images from Google
Earth. The colour of the first column represents the change-detected time in
our results. The red shape highlights the change areas. Panels
Our detected change time is also consistent with the timing of change interpreted from Google Earth and Landsat images. The deviation of the detected change years – during 2001–2006 (grey) and 2011–2014 (blue) – from the validation samples (change sample set) is shown in Fig. 7. Limited change samples from 2001 to 2006 were collected because of few high-resolution images being available during the early years. Overall, an agreement between the detected and the actual change time was found in 75.74 % of the samples (two-thirds of the detected change time matched the actual change time, while one-third was within a 1-year interval). Further, the change time tended to be more accurate during 2011–2014 (78.20 %) compared to 2001–2006 (67.07 %), given the constraints of the from–to type and the range of exact change area of oil palm from 2011 to 2014.
Difference between the detected change years using MODIS NDVI
dataset and the exact change years from the reference dataset (Google Earth
and Landsat). Negative values on the
We first compared the oil palm plantation area from our AOPD product with oil palm harvested area from FAO and USDA and the oil palm plantation area from MPOB (data available from 2011 to 2015) and BPS-Statistics Indonesia (available from 2011 to 2016; Fig. 5). Note that the FAO inventory data for Malaysia from 2011 to 2015 and the USDA statistics from 2011 to 2014 were derived from MPOB (mainly mature area). The FAO statistics included both mature and immature oil palm area during 2011–2013 but only mature oil palm area during 2014–2015, resulting in an abrupt decline in area in the FAO inventory in 2014 (the orange line in Fig. 5a). Therefore, the areas from FAO inventory should be used with caution due to the lack of reliable on-field data sources (Ordway et al., 2019).
Compared to FAO and USDA statistics, the annual mean differences from 2001
to 2016 of our results in Malaysia and Indonesia are positive and amount to
2.00 and 1.18
Trends of oil palm expansion in our mapping results (upper and lower
boundary lines) are also compared with statistical data (FAO and USDA from
2001 to 2016, MPOB and BPS-Statistics from 2011 to 2015; Table S1 in the Supplement).
Generally, the overall trends of our mapping results (0.758–0.941
An industrial oil palm plantation dataset developed by a previous study
(Gaveau et al., 2016; Fig. 8) was also used to compare
our mapping results. The oil palm plantation in Gaveau's dataset was
visually interpreted using Landsat datasets in 1973, 1990, 1995, 2000, 2005,
2010 and 2015 in Borneo. The overall distribution of oil palm extent in
Borneo is similar between our mapping results (the unidirectional version)
and Gaveau's results (Fig. 8a and b). The differences were scattered
across the whole island, with more oil palm plantation areas in our results
than in Gaveau's results in the south of Borneo (Fig. 8c; aggregated to
proportional maps at
Comparison with existing oil palm datasets in Borneo (Gaveau et
al., 2016) for 2010, 2015 and 2016. The oil palm maps were aggregated to
proportional maps at
The oil palm concession area for Indonesia and Malaysia (Sarawak) for 2014
from Global Forest Watch (
Comparison with oil palm concession from Global Forest Watch (GFW) for 2014. The PALSAR-2 images were composited in RGB format (HH, HV, HV).
Oil palm expansion is one of the major drivers of deforestation in the studied region (Austin et al., 2018). Therefore, the forest area loss map from Hansen et al. (2013) was overlaid with the AOPD map, and the results are shown for selected areas in Fig. 10a and c, where the year of oil palm expansion is roughly coincides with the year of forest clearance. In other cases such as in Fig. 10b, a larger discrepancy was found in the two maps because of different causes. For example, forest loss is not always caused by oil palm expansion but timber plantations, logging, fires, and conversion from forest to grassland and agriculture (Austin et al., 2018; Kamlun et al., 2016). Meanwhile, expansion of oil palm plantations occurred not only in forest area but also in non-forest area. In some regions, the oil palm was planted after the logging of forest immediately (area filled with same colour in Fig. 10), but in other regions, areas may first experience a forest clearance and then become oil palm plantations several years later (indicated by the patches filled with darker colour in AOPD than in the forest loss map; Fig. 10). However, the difference in the spatial resolution (30 m vs 100 m) may also cause some differences, particularly in smallholder and newly developed oil palms. According to our result, 28.20 % of total oil palm expansion area overlapped with Hansen's forest loss area (5.38 % with the exact same change time, 15.37 % later than forest loss year and the remaining 7.46 % earlier than the forest loss time). Among the overlapped area, 19.16 % of the area has the same change time, 23.67 % in 1-year intervals (may be caused by the time lag between clearance and cultivation), and 38.11 % of oil palm expanding areas in AOPD coincide with forest area loss, with a lag of at least 2 years. These 38.11 % areas may experience first forest clear-cutting for other applications or are logged and remained unused for several years and then converted to oil palm plantations.
Comparison of oil palm expansion map in this study with the Landsat forest area loss map (Hansen et al., 2013).
Mapping annual oil palm plantations using remote-sensing data in Malaysia and
Indonesia is challenging. We developed the first annual oil palm land cover
maps (AOPD) from 2001 to 2016 at 100 m resolution, combining optical and
microwave satellite observations. However, the uncertainties of AOPD, coming
from both mapping and change detection, should be acknowledged for the
future applications of our dataset. In the mapping procedure, our results
showed a good separation between primary forest and oil palm trees, but
confusion may occur in some impervious area and plantations of other species
such as coconuts. As a result, the accuracy of the change detection in the
second step was also influenced by the oil palm maps generated from
PALSAR and PALSAR-2 data in the first stage. Although oil palm maps for the 6
years of PALSAR and PALSAR-2 data reached high accuracy, at nearly 90 % in
Malaysia and
Despite of these uncertainties, the AOPD annual oil palm maps integrated the strengths of microwave (SAR) and optical satellite observations. SAR has the capability to identify the oil palm from forest regardless of the weather conditions, and MODIS time series has a hyper-temporal density and long time span. Also, our study gives a good example of integrating fine and coarse datasets. Instead of directly using the coarse dataset, the oil palm maps combined the overall change information for the whole data gap period from fine PALSAR and PALSAR-2 data and the detection of exact change year using coarse MODIS data. In recent years, there is a transition from annual classification to change information mining in remote-sensing interpretation to reduce the false changes (Xu et al., 2018b). This method can be used not only in monitoring global oil palm dynamics but also in producing annual land cover maps where only discrete high-resolution observations are available. Since the data scarcity of successive Landsat imagery is common across the world, the algorithm described in this study provides an effective way of combining coarse data to update the annual land cover change. Further, inventory compilation and manual visualization of oil palm change to a large extent would remain labour-intensive and time-consuming (Gaveau et al., 2016; Miettinen et al., 2016; Vijay et al., 2018). Our semi-automatic algorithm in oil palm mapping may thus help to establish long-term monitoring for oil palm that can be improved over time with regular validation using ground-based observation or very high resolution images such as Google Earth.
The 100 m annual oil palm maps from AOPD produced in this study can be used
in a number of applications. First of all, it can be used as
cross-validation reference data for other regional oil palm datasets (e.g.
FAO inventory). Second, the annual data can be further used to quantify the
spatiotemporal characteristics of oil palm change, estimate the annual oil
palm yields, identify the potential area planted with oil palm and predict the
boundary of oil palm expansion in the future, and so on. Overlapping the AOPD
with forest maps, peatland maps and other land cover maps can give a clue to
how the oil palm expansion influences different ecosystems and their carbon
balance. For example, oil palm expansion is the largest single driver of
deforestation in Indonesia, which contributed to 2.08
Another vision lies in the sustainable future of oil palm industry. As the major contributor to the economy that supports thousands of people in the tropical countries, the developing oil palm industry has been one of the priorities in these countries (Mahmud et al., 2010; Sayer et al., 2012). At the same time, the possible environmental and ecological consequences of monocultures need to be taken into account for the sustainable development of oil palm industry. For example, the Roundtable on Sustainable Palm Oil (RSPO) was established to formulate the standards for the industrial oil palm plantations in southeastern Asia, followed by the foundation of the Africa Palm Oil Initiative. Voluntary zero-deforestation commitments in the palm oil industry have also been implemented since 2010 (Focus, 2016). However, the number of large corporations and the extent to which they pay real attention to the rights of local populations remain unknown (Barr and Sayer, 2012).
It is crucial to find a balance between the rural economic development and environmental protection, especially in the regions with high-biodiversity primary forest and carbon-rich peatlands like southeastern Asia. More complete information on oil palm plantations (e.g. spatiotemporal changes of oil palm and its consequences) would help to reduce the disputes and provide strategies for oil palm's sustainable development. Our annual oil palm maps would thus contribute to the policy formulation as well as policy evaluation (e.g. national moratorium on new permits for the oil palm conversion from primary natural forests and peatlands; Busch et al., 2015).
The AOPD in Malaysia and Indonesia from 2001 to 2016 at 100 m resolution is
available to the public at
Combining the optical and microwave satellite observations, we developed the first annual oil palm maps (AOPD) in Malaysia and Indonesia from 2001 to 2016 at 100 m resolution using the image classification and change-detection analysis. The dataset reached high accuracy in both annual classification and change detection. As a result, this dataset provided insights and details on dynamic oil palm changes for Malaysia and Indonesia from the perspective of remote sensing and can serve as a supplement for statistics. Further applications of the dataset include but are not limited to regional carbon studies, water and agricultural management, biodiversity and conservation protection, and the sustainable development of the oil palm industry. The annual updating method in this study that fully used information from discrete high-resolution data and continuous low-resolution data is also expected to be applicable in other regions facing data scarcity.
The supplement related to this article is available online at:
YX and LY conceived the study. YX analysed the data and drafted the paper. WL, PC, YC and PG supported the generation of the dataset and the analysis of the results.
The authors declare that they have no conflict of interest.
This research has been supported by the National Key R&D Program of China (grant nos. 2017YFA0604401 and 2019YFA0606601) and the Tsinghua University Initiative Scientific Research Program (grant no. 2019Z02CAU).
This paper was edited by David Carlson and reviewed by two anonymous referees.