What causes deforestation in Indonesia?

We used 30 m resolution Global Forest Change (GFC) maps representing tree cover loss annually between 2001 and 2016 (Hansen et al 2013). We used tree cover loss data 2001–2015 from GFC Version 1.3 and, as updated data became available, we included loss events from 2016 from GFC Version 1.5. All GFC data versions 1.1 onwards include reprocessed data beginning in 2011 that could have introduced time series inconsistencies in our analysis. Once available, a planned reprocessing of all years 2001 onwards will allow us to assess the impact of these inconsistencies.

Hansen et al defined tree cover loss as the transformation of forests to a non-forest state at the mapping scale; we refer to this interchangeably as deforestation and forest loss. We defined a deforestation event as at least two contiguous loss pixels, according to an eight-neighbor rule, where tree cover loss was reported in a given calendar year. This resulted in a minimum change area of 1800 square meters, or 0.18 hectares (ha). We excluded single pixel events from our analysis, which are more likely associated with errors. As the focus of our analysis is forest loss in natural forests, rather than in plantations undergoing harvesting, we omitted GFC tree cover loss events where the majority of the event was outside intact or degraded primary natural forests prior to the start of our analysis (Margono et al 2014).

We randomly selected deforestation events from the full GFC dataset, stratified by deforestation event size. We used three size class strata representing small (<10 ha), medium (10–100 ha) and large (>100 ha) deforestation events (Austin et al 2017a). Size class stratification is required when using events as the unit of analysis, rather than pixels, as there were relatively few medium- and large-scale events, but these comprised a large area of deforestation. Our sample size was n = 4590 events, representing 455 765 ha of deforestation, or 5% of the total area deforested in primary natural forests over the study period.

We imported the selected deforestation events into Google Earth Pro for visual interpretation. Two interpreters classified the sampled deforestation events into 12 driver categories based on the land cover observed following a tree cover loss event (table 1, figure 1). The interpreters used features in the landscape within and adjacent to the event to determine the subsequent land cover. To identify drivers of deforestation the interpreters viewed all available high spatial resolution imagery and annual cloud-minimized Landsat composites in the four years following the event (for example, over 2006–2009 for an event that occurred in the year 2005). In most cases we were able to identify a driver within a two-year window after the event, but we used the four-year time frame to allow for sufficient tree growth for confident plantation species categorization. In a small number of cases we observed more than one land cover within the four-year window; in these cases we assigned a driver category based on the latest available high-resolution image. In just under one-quarter of all samples there was no good quality high-resolution imagery during the window, and in these cases we either based our assessment on Landsat imagery if possible or, in just over 15% of cases, assigned a 'no data' classification. The four-year time frame is not available for more recent years, and as a result we assigned 'no data' to just over 20% of the events in 2015–2016.

Zoom In Zoom Out Reset image size

We checked the stability of our driver classification by additionally examining longer term land cover transitions. For deforestation events that occurred 2001–2005, we recorded both the initial driver of change occurring within four years of tree cover loss, and a secondary driver occurring at least ten years after the event. This allowed us to examine longer term trajectories of land cover change, and to gauge the extent of more complex land cover transitions.

To assess the reliability of our visual interpretation method, the two expert interpreters classified 100 of the same samples within each size class category. We compared the assignment of drivers for this subsample, and report user's and producer's agreement for each driver category (table S2). Overall agreement was 91%, 90%, and 95% for small, medium and large patches, respectively. Agreement was >90% for all individual driver categories except for large-scale plantations of unknown species and timber plantations in the small size class, and small-scale oil palm plantations and small-scale mixed plantations across all size classes (table S2). Due to the difficulty of distinguishing small-scale plantations from small-scale oil palm plantations, we combine these two categories in the presentation of our results.

We calculated the variability of our estimates by creating a bootstrap distribution via resampling from each of our stratified samples. We used the boot package in R version 3.5.0 to randomly sample 1000 times with replacement, calculated the area and relative proportions of each driver for each iteration, and calculated confidence intervals based on the resulting distributions (Canty and Ripley 2017, R Core Team 2018).

To calculate the proportion of each driver in each year, we weighted our sample results in each size class by the proportion of total deforestation in each size class occurring nationwide in each year (figure S1) (Austin et al 2017a). To explore the heterogeneous nature of drivers of deforestation across different landscapes of potential interest, we repeat this process using the proportion of deforestation in each size class across major islands, permit types (Global Forest Watch 2015), protected areas (IUCN and UNEP-WCMC 2015), intact forest landscapes (IFLs), and peat lands (MoA 2011). IFLs are contiguous forests with no signs of anthropogenic disturbance detectable via remotely sensed imagery and large enough to maintain populations of wide-ranging species (Potapov et al 2008).

We observed substantial variation in drivers of deforestation across the archipelago (figure 3, figure S2). Although relatively smaller contributors to overall deforestation, Sulawesi, Java and Nusa Tenggara are characterized by a significantly larger proportion of deforestation driven by small-scale agriculture than elsewhere in the country. In Papua, logging roads and small-scale clearings followed by secondary forest regrowth were responsible for more than one-third of deforestation, a much larger proportion than observed nationwide. Importantly, deforestation driven by oil palm and other large-scale plantations grew in importance over the study period and together became the dominant driver of deforestation in Papua after 2013.

Zoom In Zoom Out Reset image size

In Kalimantan, oil palm plantations were the dominant driver of deforestation from 2005 to 2013, with a peak in 2009 followed by a steady decline (figure S2). Forest conversion to grassland/shrubland was also an important driver of forest loss in Kalimantan, and the notable peaks in deforestation driven by grassland expansion observed at the national level are attributable in large part to the grassland expansion dynamics in this region. For example, more than two-thirds of the area of national forest to grassland conversion in 2016 occurred in Kalimantan. On the other hand, Sumatra has a more varied driver profile, with roughly equal areas of deforestation driven by oil palm plantations, grassland/shrublands, and small-scale agriculture. The island experienced a peak in oil palm plantation expansion into forests somewhat earlier than in Kalimantan, from 2005 to 2009, followed by a peak in timber plantations expansion from 2010 to 2012. Notably, the Sumatra has a substantially higher rate of deforestation driven by small-scale agriculture and small-scale plantations than Kalimantan.

Government permitting agencies grant concessions for specific types of land use, including for oil palm plantations, pulpwood plantations, and selective logging activities. As expected, oil palm expansion dominated deforestation within permits granted for oil palm cultivation, and timber plantations dominated deforestation within permits granted for pulpwood cultivation, each causing more than half of deforestation in these permit types (table 3). While not the dominant driver of deforestation in permits for selective logging, logging roads were more prominent in this category of permits than elsewhere, contributing more than one-quarter of deforestation in this category. Thus, while we observed a general trend toward deforestation drivers corresponding to their expected permit type, a large fraction deforestation in each permit type was followed by a different land cover than expected. This is not surprising, given that previous research in Indonesia has illustrated a significant mismatch between land cover and permit allocations in Indonesia, with at least 33% of oil palm and pulpwood permits in one region occupied by independent farmers (Gaveau et al 2017).

Not unsurprisingly, logging roads and small-scale clearings followed by secondary forest regrowth were responsible for almost 40% of deforestation in IFLs (table 3). Oil palm, timber and other large-scale plantations were responsible for the remainder. In peat forests, oil palm, timber, and other large-scale plantations drove more than half of deforestation, with most of the remainder attributable to grassland/shrubland expansion. And in protected areas, conversion to grasslands/shrublands was the dominant driver of deforestation, causing almost two-thirds of deforestation, with most of the remainder attributable to small holder agriculture.

Conversion of forest to one land cover category during 2001–2005 is rarely followed by another land cover transition during the time frame of our study. Just 8% of all deforestation events 2001–2005 were assigned one driver based on the initial land cover following forest loss, and then converted to a different land cover ten years later (figure S3). Forest loss events caused by mining and oil palm, timber, and small-scale plantations almost always continued in these same land cover categories after 10 years. On the other hand, one-fifth of deforestation that occurred for small-scale agriculture and one-third of deforestation for grassland/shrubland expansion were later converted to another land cover category. The dominant follow-up transition in these cases was conversion to small-scale mixed plantations and oil palm plantations. Although relatively minor in terms of overall deforestation, roughly half of the area deforested to make way for logging roads was subsequently converted to another land cover category, usually small-scale agriculture, small-scale plantations and oil palm plantations. This suggests that over longer time frames these driver categories may be intermediate steps towards more intensive land cover conversion and resource extraction. The timing and nature of these transitions merit further investigation, particularly given our observation of increasing conversion of forest to grasslands in recent years.

Our findings confirm that large-scale plantation agriculture is an important driver of deforestation in Indonesia. However, an important new finding is that the role of these large-scale plantations has diminished since the end of the 2000s, when they were responsible for more than half of deforestation, to 2014–2016, when they were responsible for an average of one-quarter of deforestation. This observed decline could be due, at least in part, to the impacts of forest conservation interventions aimed specifically at preventing large-scale plantations from causing forest loss. These include for example Indonesia's national moratorium on new permits for the conversion of primary natural forests and peat lands, established in May of 2011 (Austin et al 2014, Busch et al 2015), and corporate voluntary zero-deforestation commitments in the palm oil industry since 2010 (Climate Focus 2016). While outside the scope of the present study, our map of annual deforestation drivers will enable more refined evaluations of policy impacts, including for example causal analysis of the influence of these policy interventions specifically on deforestation driven by large-scale plantations.

Conversion to grasslands was another important driver of deforestation nationally, is the most important driver of deforestation in protected areas and peat forests, and appears to correspond to dry seasons with observed spikes in fire activity. In particular, grassland expansion into forests increased dramatically during the 2014–2016 period, particularly in Kalimantan, and was a key contributor to the increasing rate of forest loss nationally up to 2016. The severe dry seasons in 2015 may have set the stage for the high rate of observed grassland expansion in 2016 and resulted in acute levels of transboundary air pollution and associated public health crises. In response to this burning event the national government put in place several measures to prevent fires on peatlands, including regulations for restoring degraded peat lands and swapping timber plantation permits on peat lands for alternative areas on mineral soils (Jong 2017). Further assessment of the role of fire in deforestation trajectories, including through incorporation of improved estimates of burned area (Giglio et al 2016), will strengthen our understanding of these linkages.

Given that large-scale plantation expansion was a substantial driver of forest loss in Indonesia, there has been a justifiable focus among forest conservationists on counteracting these agents of change. However, we note that small holder agriculture and plantations continued to contribute to forest loss over the study period, suggesting the importance of designing forest management interventions that take the values and requirements of these agents of change into account. This is challenging, however, due to the large number of small-scale farmers across Indonesia with diverse motivations and governing principles. Evidence from Brazil suggests that, while policy interventions aimed at addressing large-scale drivers of forest loss have been at least partially successful at slowing deforestation (Nepstad et al 2014), interventions addressing small-scale drivers of change have had relatively little impact (Godar et al 2014).

The context of smallholder plantation agriculture in Indonesia is also unique, in that there is a direct link between some forms of small holder plantation management and their large-scale industrial counterparts. For example, small-scale oil palm cultivation comprises 40% of national production, and is closely tied to large-scale oil palm enterprises via outgrower schemes (DJP 2016). Mosnier et al report that, as large-scale oil palm plantations developers increasingly comply with sustainability standards, the area cultivated by small-scale producers will likely increase (Mosnier et al 2017). Given that compliance with emerging sustainability standards may be less enforceable and more challenging for these small-scale plantation managers, this increase could lead to a larger proportion of deforestation driven by small-scale plantations. On the other hand, there may also be a window of opportunity to facilitate small holder compliance with sustainability standards and reduce deforestation, by leveraging extension services and informal information networks.

There is potential to augment our method for characterizing drivers of deforestation by using crowdsourcing approaches, which leverage citizen scientists to scale up data collection efforts. Such efforts have successfully collected data on, for example, global agricultural lands and field sizes (Fritz et al 2015), and tree cover in dry lands (Bastin et al 2017). There is also potential to use sample data such as those collected in this study to parameterize a statistical model that predicts the drivers of deforestation, by considering the characteristics of the events (e.g. size, shape, location relative to infrastructure). Such models could facilitate more rapid and systematic assessment of deforestation drivers over time (Finer et al 2018). For example, Richards and Friess trained a model to predict drivers of mangrove forest loss in Southeast Asia (Richards and Friess 2016), and recently Curtis et al trained a model to predict drivers of deforestation globally (Curtis et al 2018).

We note that there are important considerations of spatial scale when conducting evaluations of drivers of deforestation using any of these methods. For example, Curtis et al use 10 km² grids for their global assessment of deforestation drivers, which is several orders of magnitude larger than the unit of analysis in the present study. This is one factor contributing to their estimate of commodity driven deforestation in Indonesia of 88%, a value nearly twice as large as our corresponding estimate of 46% over the same period (table S1). This highlights the influential role of spatial resolution in assessments of direct deforestation drivers.

We acknowledge that a better understanding of the direct drivers of forest loss is just one step towards developing robust policy interventions to manage forests, mitigate GHG emissions, conserve biodiversity, and protect ecosystem services. To design, implement, and enforce effective policies, the forest conservation community needs to improve understanding of the complex suite of underlying factors that lead to the rise or fall of specific direct drivers of deforestation. These include population, income, market demand, infrastructure, tenure security, and law enforcement (Geist and Lambin 2002, Busch and Ferretti-Gallon 2017). A comprehensive dataset of direct drivers of deforestation, such as the data developed in this analysis, provides a key element needed to investigate the pathways and linkages between these underlying factors and the agents directly responsible for deforestation.