Rates and drivers of mangrove deforestation in Southeast Asia, 2000–2012

Significance

This study quantifies the proximate drivers (i.e., replacement land uses) of mangrove deforestation across Southeast Asia between 2000 and 2012. Mangrove forests in the region were lost at an average rate of 0.18% per year. Aquaculture was a major pressure on mangrove systems during this period, but its dominance was lower than expected, contrary to popular development narratives. Rice agriculture has been a major driver of mangrove loss in Myanmar, and oil palm expansion is a key but under-recognized threat in Malaysia and Indonesia. The threat of oil palm to mangroves is likely to increase in the future as new frontiers open up in Papua, Indonesia. Future research and policy responses must consider the diversity of drivers of mangrove deforestation.

Abstract

The mangrove forests of Southeast Asia are highly biodiverse and provide multiple ecosystem services upon which millions of people depend. Mangroves enhance fisheries and coastal protection, and store among the highest densities of carbon of any ecosystem globally. Mangrove forests have experienced extensive deforestation owing to global demand for commodities, and previous studies have identified the expansion of aquaculture as largely responsible. The proportional conversion of mangroves to different land use types has not been systematically quantified across Southeast Asia, however, particularly in recent years. In this study we apply a combined geographic information system and remote sensing method to quantify the key proximate drivers (i.e., replacement land uses) of mangrove deforestation in Southeast Asia between 2000 and 2012. Mangrove forests were lost at an average rate of 0.18% per year, which is lower than previously published estimates. In total, more than 100,000 ha of mangroves were removed during the study period, with aquaculture accounting for 30% of this total forest change. The rapid expansion of rice agriculture in Myanmar, and the sustained conversion of mangroves to oil palm plantations in Malaysia and Indonesia, are identified as additional increasing and under-recognized threats to mangrove ecosystems. Our study highlights frontiers of mangrove deforestation in the border states of Myanmar, on Borneo, and in Indonesian Papua. To implement policies that conserve mangrove forests across Southeast Asia, it is essential to consider the national and subnational variation in the land uses that follow deforestation.

Global demand for food, biofuels, and raw materials continues to drive land use and land cover change (LULCC), particularly in the tropics (1, 2). Demands are expected to further intensify as populations and global affluence increase, and further LULCC is expected, given that future demand cannot be met by yield increases of currently cropped lands alone (3). LULCC enables large-scale commodity production, but can have substantial negative impacts on biodiversity (4) and the provision of ecosystem services (5). To manage LULCC-driven deforestation and conserve tropical forested landscapes in the future, it is important to understand spatial and temporal variation in the land uses that replace forests, both locally and at regional scales.

Coastal mangrove forests grow in the intertidal zone in tropical and subtropical regions (6). Southeast Asia contains the greatest diversity of mangrove species (7) and more than one-third of the world’s mangrove forest extent (8). Estimates of historical mangrove deforestation are unreliable in many instances (9), but Asia may have lost more than one-third of its mangrove area between the 1980s and 1990s (10). Such deforestation has had substantial negative impacts on biodiversity, with 16% of the world’s mangrove vegetation species now at an elevated risk of extinction (11). Mangrove deforestation also has implications for the provision of ecosystem services. For example, mangroves store disproportionately high densities of carbon compared with other ecosystems (12), so LULCC in mangroves may result in carbon emissions equal to 2–8% of emissions from terrestrial deforestation, despite the fact that this ecosystem represents only 0.7% of the global tropical forest area (13, 14).

The aquaculture industry has been held primarily responsible for mangrove deforestation in Southeast Asia over the past 30 y (10, 15). Agriculture, forestry, and urbanization were generally considered less important drivers at the regional scale (10, 16), but rice agriculture and urbanization have recently been shown to be locally important (17, 18). The demand for alternative land uses, such as oil palm plantations, is a major driver of terrestrial deforestation in the region (19, 20), but oil palm has been considered in tropical coastal habitats only rarely. In general, the current importance of different drivers of mangrove conversion is not clear, in part because previous analyses focused on longer-term changes over several decades (10, 16, 21), hindering assessment of recent and contemporary trends. In addition, several previous studies used national-level government statistics (10, 15, 16), but these data may be unreliable (9, 22), and do not allow analysis of spatial variation in drivers within countries.

Different policy interventions may be needed to combat mangrove deforestation in Southeast Asia, depending on the proximate drivers that are responsible. For example, legislation may be led by different government ministries depending on whether the replacement land use is urban, forestry, agriculture, or aquaculture, especially because mangrove management is commonly spread over multiple agencies in Southeast Asian countries (23). Improvements in remote sensing technology and image analysis have allowed us to systematically monitor changes in the distribution of forests at regional and global scales (6, 8, 24), annually and at a high spatial resolution (24). Comparable approaches have not been applied to identify the land uses that replace forests, however, making it difficult to quantify the major proximate drivers of deforestation over large areas. Some previous studies of LULCC at global and regional scales have used intermediate-resolution remote sensing imagery and, correspondingly, have analyzed broad land cover categories, such as “plantation” or “lowland forest” (25). Such broad land cover categories do not allow us to identify the particular commodities being produced. Remote sensing LULCC studies using finer-resolution imagery are typically geographically restricted to smaller case studies (18, 26), making it difficult to identify general patterns and compare the relative importance of different replacement land uses across countries.

In this study, we applied a systematic remote sensing method across Southeast Asia to quantify the LULCC that occurred in mangrove forests annually between 2000 and 2012, within all deforested patches larger than 0.5 ha. We identified the replacement land uses that followed deforestation by classifying patches into categories linked to proximate drivers: aquaculture, rice-dominated arable, oil palm plantation, urban, and other terrestrial forest (including possible rubber plantations). We also classified categories for mangrove regrowth, coastal erosion, and recently deforested mangrove with no clear replacement land use. We analyzed national and subnational variation in the rates and drivers of mangrove deforestation across Southeast Asia, and compared temporal trends in the three major replacement land uses between 2000 and 2012.

Results

Mangrove deforestation remains substantial across Southeast Asia, with more than 100,000 ha of mangrove forest lost between 2000 and 2012 (Table 1). Approximately 2% of the mangroves present in Southeast Asia in 2000 were lost during the study period, at an average rate of 0.18% per year. There is considerable spatial variation in the degree of deforestation, with hotspots in Myanmar, particularly in Rakhine state, in Indonesian Sumatra and Borneo, and in Malaysia (Fig. 1 and Table 1). The rate of mangrove deforestation was higher in these regions and was considerably lower in Thailand, Vietnam, and the Philippines (Table 1).

Table 1.

Mangrove area and loss statistics for Southeast Asian countries between 2000 and 2012

Country	Total mangrove in 2000, ha	Mangrove deforestation, ha	Mangrove habitat area lost, ha	Percentage mangrove loss 2000–2012, %
Indonesia	2,788,683	60,906	48,025	1.72
Myanmar	502,466	27,957	27,770	5.53
Malaysia	557,805	18,836	15,809	2.83
Thailand	245,179	3,504	3,344	1.36
Philippines	257,575	1,423	1,296	0.50
Cambodia	47,563	1,218	1,086	2.28
Vietnam	215,154	531	528	0.25
Brunei	11,054	48	41	0.37
Timor-Leste	1,066	2	2	0.19
Singapore	583	0	0	0
Southeast Asia	4,626,545	114,424	97,901	2.12

Countries are ordered by total mangrove lost. Mangrove habitat lost takes into account mangrove regrowth in deforested areas during the period.

Fig. 1.

Mangrove deforestation between 2000 and 2012. Deforestation is summarized within each 1 decimal degree square.

There were three main mangrove replacement land use types identified during the study period: aquaculture, rice, and oil palm (Table 2). Over the entire study period, the single most important replacement land use was aquaculture (30% of total area; SE, 2%), which was particularly dominant in Indonesia, Cambodia, and the Philippines (Fig. 2 and Table 2). Conversion to rice agriculture was important at the regional scale (22% of total area; SE, 1.5%), but this figure is skewed heavily by rice expansion in Myanmar (Fig. 2 and Table 2). Oil palm plantations also accounted for a considerable area (16% of total area; SE, 1.6%), particularly in Malaysia and Indonesia (Fig. 2 and Table 2). Urbanization had a small regional impact on mangrove extent, but locally important impacts in the Bangkok region, southern Malaysia and Vietnam (Fig. 2 and Table 2). Approximately 15% (SE, 1.5%) of the mangrove lost during the study period was classified as mangrove regrowth in 2012. Mangrove regrowth was particularly apparent in Indonesia (Table 2) and northwest Malaysia (Fig. 2).

Table 2.

Percentage of the total deforested mangrove (2000–2012) converted to different land uses

Country	Aquaculture	Rice	Oil palm	Mangrove forest	Urban	Other category
Indonesia	48.6	0.1	15.7	22.6	1.9	11.2
Myanmar	1.6	87.6	1.1	0.5	1.6	7.6
Malaysia	14.7	0.1	38.2	17.6	12.8	16.7
Thailand	10.8	5.6	40.0	5.1	14.4	24.1
Philippines	36.7	0.9	11.1	7.3	2.7	41.3
Cambodia	27.7	1.5	8.9	9.8	4.6	47.6
Vietnam	21.0	10.4	0.5	0.6	62.5	4.9
Brunei	29.2	0	27.7	12.5	15.9	14.8
Timor-Leste	0	26.1	0	0	0	73.9*
Singapore	0	0	0	0	0	0
Total	29.9	21.7	16.3	15.4	4.2	12.3

Countries are ordered by total mangrove lost. Percentages might not sum to 100 owing to rounding.

^*The small amount of mangrove deforestation in Timor-Leste is due mainly to shoreline erosion.

Fig. 2.

Percentage mangrove deforestation between 2000 and 2012, and dominant land uses of deforested areas in 2012. Land uses are summarized as the converted land use with the greatest area within each 1 decimal degree grid square. Circles are located in the center of each grid square, and circle size represents the percentage of the mangrove area in 2000 that has been lost.

Land use conversion did not occur at a constant rate between 2000 and 2012, and the relative importance of the different drivers varied temporally (Fig. 3). The percentage conversion to aquaculture ponds declined from 2000, before rising to the 2000 level in 2010 and 2011 (Fig. 3A). The percentage of mangrove converted to rice fields increased steadily between 2000 and 2009, before falling rapidly during 2010 (Fig. 3B). The rate of conversion to oil palm showed a sustained increase throughout the study period (Fig. 3C).

Fig. 3.

Temporal trends in the conversion of mangrove habitats to aquaculture (A), rice agriculture (B), and oil palm plantation (C), between 2000 and 2012. Black lines indicate error-corrected estimates of the proportional coverage of each land use. Gray shading indicates the SE of the areal estimates.

Discussion

Deforestation Rates Are Lower Than Previously Thought.

The rate at which mangroves present in 2000 were deforested up to 2012 (average 0.18% per year) is lower than that of previous estimates across Asia (10) and insular Southeast Asia (25), which have estimated rates of at least 1% per year. It is possible that the rate of mangrove conversion has slowed since the 1990s, but the discrepancy may also be related to methodological differences, given that the higher estimates of mangrove deforestation were based on analyses of national-level government statistics or literature reviews (10, 16) or on satellite imagery with relatively coarse resolution (25). The rates of mangrove loss in Southeast Asia reported in this study are lower than previously thought, but nonetheless a substantial area of mangrove (on average 9,535 ha year⁻¹) is lost annually. Continued mangrove deforestation will have further negative impacts on biodiversity and ecosystem service supply.

There was substantial spatial variation in the rate of mangrove forest change between 2000 and 2012, with >10% of mangrove forest lost per 1° grid square in parts of Rakhine state in Myanmar and in Sumatra, Borneo, and Sulawesi in Indonesia. In contrast, several countries that once were considered hotspots of mangrove deforestation, such as Vietnam and Thailand, showed relatively slower rates of deforestation between 2000 and 2012. The history of commodity production in these areas means that their coastlines have long been heavily managed, and the smaller areas of mangrove that remain may be more strongly protected (27).

Although most deforested mangrove was replaced with agriculture or aquaculture, a considerable deforested area was classified as mangrove in 2012, particularly in Malaysia and Indonesia. This regrowth may occur after illegal logging of mangrove wood, or after tree removal in sustainably managed mangrove forestry schemes, such as the Matang Mangrove Forest Reserve in northwest Malaysia. The Matang Reserve is largely managed as a sustainable monoculture of one mangrove tree species (Rhizophora apiculata) to provide mangrove charcoal (28). It is likely that mangrove forest has established in areas where it was not present in 2000, but such expansions would not be recorded using our method. The rate of mangrove forest expansion is considerable in South Asia (6), so it is possible that the percentage net loss in mangrove forest area in Southeast Asia between 2000 and 2012 may be less than 2%.

Aquaculture as a Driver of Mangrove Loss.

Previous studies in Southeast Asia and globally have focused on the role of aquaculture in driving mangrove deforestation (10, 15, 29). Although aquaculture was still the dominant driver of mangrove deforestation between 2000 and 2012, the percentage converted to fish or shrimp ponds was approximately one-half that estimated during the 1980s–1990s (10), a period of rapid expansion of tropical coastal aquaculture. During the 1980s–1990s, aquaculture accounted for as much as 54% of all mangrove deforestation in a survey of eight major aquaculture-producing countries (29). The percentage converted annually to aquaculture was lower during most of the second half of our study period compared with the first half (although the percentage rose during 2010 and 2011). Since the 1960s and 1970s, conversion of mangrove forests to aquaculture ponds has been encouraged by the governments of Thailand, Indonesia, Vietnam, and the Philippines to enhance food security and improve livelihood (27, 30). These countries are now some of the largest aquaculture producers in the world (31). However, policies that encouraged expansion rather than intensification have now been reversed, and there are increased environmental regulations for new aquaculture development (27). Intensive production now accounts for the majority of production in Thailand (32).

Mangrove conversion to aquaculture now occurs mainly in Kalimantan and Sulawesi, Indonesia. Aquacultural expansion in these provinces largely drove the regional increase in mangrove conversion to aquaculture observed in 2010 and 2011. Aquacultural expansion has been driven in part by the recent Indonesian Government Regulation Per.06/MEN/2010, a policy that aims to position Indonesia as the world’s largest aquacultural producer by 2015 (33). Thus, while Indonesia’s aquacultural production was only slightly higher than other regional producers in 2006 (at 2.48 million tons), in 2012 Indonesia’s production (9.6 million tons) was almost threefold larger than that of other regional aquaculture producers, such as Vietnam (3.3 million tons) (34). Indonesian government departments continue to encourage growth in the industry as a means of improving livelihood, generating foreign currency, and providing protein (30), so further mangrove conversion may be expected in the future.

Rice Expansion in Myanmar.

Agricultural expansion for rice production, primarily in Myanmar, accounted for more than 20% of the total mangrove change in Southeast Asia over the study period. The local impact of rapid rice expansion on mangrove extent in the Ayeyarwady Delta has been described previously (18), but the present study shows that the expansion of rice agriculture across the whole of Myanmar is responsible for driving the fastest rate of mangrove deforestation of any country in Southeast Asia. Furthermore, our findings indicate that the rate of mangrove replacement with rice agriculture was lower in the agricultural hotspot of the Ayeyarwady Delta, and that recent rice expansion into mangroves has largely occurred in the state of Rakhine, an outlying region of the country with poor connections to the center (35). The government of Myanmar has historically aimed to increase rice production through engineering assistance and village-level expansion targets to enhance national food security (36). Reforms of the rice market in 2003, and accompanying suggestions of further liberalization by the government, might have stimulated some increased activity in the private market owing to price increases (18, 35).

Increasing rice production is considered critical for national food security (35), and it is likely that economic diversification from rice to such products as shrimp and oil palm will occur as export restrictions are eased in the future (18). The Myanmar government provides few environmental safeguards for mangrove forests; for example, the current protected area network is poorly enforced and covers little mangrove forest (37). As a result of the lack of environmental safeguards and continuing economic transformation in Myanmar, we may expect mangrove conversion to rice and other agriculture to continue to displace large areas of mangrove in this country in the future.

The Rise of Oil Palm as a Proximate Driver of Mangrove Deforestation.

The development of oil palm plantations is a major driver of terrestrial forest and peat swamp deforestation in Malaysia and Indonesia (19, 38). That only a limited number of local or anecdotal case studies have identified oil palm cultivation as a potential driver of mangrove loss (39, 40) is surprising, given that our study highlights the large scale of oil palm production in former mangrove forests, particularly in Malaysia and Sumatra and Borneo in Indonesia. This is in keeping with the status of these countries as the top palm oil producers in the region; together they produce 85% of the world’s palm oil (41). Palm oil production is encouraged by governments in Southeast Asia to enable energy security and economic development, with most plantations run by larger private enterprises or by smallholders who sell to large private enterprises (40, 42). The responsibility for intertidal habitats, such as mangroves, commonly falls between marine and terrestrial government agencies, which can lead to neglect of monitoring and management (43). Thus, in the past, conversion of mangrove forest to oil palm plantations might have been unnoticed or underreported, because oil palm expansion is generally considered a terrestrial issue (19, 38), and because plantations that replace mangrove forests may look similar to those that replace terrestrial and freshwater peat swamp forests.

Palm oil production in Indonesia is expected to continue to increase by almost 30% above 2012 levels by 2019 (44), owing to increasing global demand for foodstuffs and national targets to ensure energy security (41). It is likely that a large proportion of Indonesia’s future oil palm expansion may occur in Papua. Papua has already granted large areas of terrestrial oil palm concessions (45), with a recent report showing an increasing rate of concession granting in the region (46). In May 2015, Indonesian President Joko Widodo announced the development of 1.5 million ha of new agricultural land in Papua within the next 3 y, as part of the Merauke Integrated Food and Energy Estate. This mixed agricultural development project is designed to increase food and energy security and stimulate economic growth in Papua (47, 48). Although our analysis showed a low deforestation rate in the mangrove-rich Indonesian province of Papua between 2000 and 2012 (Fig. 1), developments such as the Merauke project will bring substantial environmental and social impacts in the future (48).

Conclusions

Mangrove forests—a tropical coastal ecosystem on which millions of people depend—continued to be lost in Southeast Asia at an average rate of 0.18% per year between 2000 and 2012. Across Southeast Asia, mangrove forests are converted to alternative land uses to provide commodities, but the motivating factors for this conversion vary according to location and target commodity. In Myanmar, rice production is considered critical for national food security (35), whereas in Indonesia, Thailand, and the Philippines aquaculture is commonly presented as a means to develop the economy (27, 30). Palm oil production in Malaysia, Indonesia, and Thailand is promoted to enhance the economy and improve national energy security (41). Land use changes are performed by different demographic groups of people in different circumstances; in Myanmar, rice is farmed mainly by smallholders (36), whereas oil palm and aquaculture operations are commonly owned or managed by larger corporations (30, 40). This study provides quantitative data on the land uses that replace mangrove forests, at a high spatial resolution and annual frequency. This detailed information is required for decision makers to implement appropriate, evidence-based conservation. Thus, policy interventions must be targeted to address national and subnational variations in the drivers of mangrove loss.

Methods

Our analysis builds on two high-quality existing datasets: the global forest change dataset provided by Hansen et al. (24), which maps global deforestation annually between 2000 and 2012 at a detailed spatial resolution (pixel size = 0.09 ha), and the global distribution of mangrove forests in 2000/01 provided by Giri et al. (8). We performed a supervised land use classification of satellite imagery for each >0.5-ha deforested mangrove patch in Southeast Asia.

We cross-referenced the global forest change dataset (24) with the global distribution of mangrove forests in 2000/01 (8), and with the boundaries of the ASEAN states and Timor-Leste, to map the distribution of deforested mangrove pixels in Southeast Asia. We identified continuous patches of mangroves (i.e., pixels continuously connected in at least one of eight directions) that were deforested in the same year and that were larger than 0.5 ha, a total of 45,540 patches. Landsat satellite imagery was extracted for each deforested patch from the preprocessed 2012 Landsat image (24). We also calculated four other geographical predictor variables for each deforested patch: the Normalized Difference Vegetation Index in the surrounding 25 ha, the distance from the nearest major road at the center of the patch, and the United Nations Food and Agriculture Organization indices of climatic suitability for oil palm cultivation and rice agriculture (Table S1).

Table S1.

Sources of additional patch-level predictor variables

Dataset	Notes	Source
Surrounding Normalized Difference Vegetation Index	NDVI derived from Landsat 7 imagery in the surrounding area. Median from 3 y (2012, 2013, and 2014) (0.5-km resolution)	Google Earth Engine (https://earthengine.google.com).
Distance from nearest road	Euclidean distance raster derived from vector map of all major roads	Digital Chart of the World, obtained via DIVA-GIS (www.diva-gis.org)
Suitability index for oil palm cultivation	Crop suitability index for intermediate input level rain-fed oil palm (5-arcmin resolution)	UN Food and Agriculture Organization Global Agro-Ecological Zones (gaez.fao.org)
Suitability index for rice cultivation	Crop suitability index for intermediate input level rain-fed wetland rice (5-arcmin resolution)	UN Food and Agriculture Organization Global Agro-Ecological Zones (gaez.fao.org)

We used Google Earth Pro to view 3,091 deforested patches, and inferred the land use of 1,500 of these patches, which (i) were deforested before the date of the most recent imagery available and (ii) were clearly visible in the available imagery. We categorized the deforested patches into one of eight land use classes: aquaculture, rice field, oil palm plantation, urban, mangrove regrowth, terrestrial forest, coastal erosion, and recent mangrove deforestation with no observable replacement land use (Fig. S1). The majority of patches were sampled at random (n = 1,008), with additional targeted sampling done to increase the representation of less common categories (Fig. S2 and Table S2).

Fig. S1.

Examples of the eight coastal land cover and land use types that served as reference categories to classify deforested mangrove patches. The eight land uses were: (A) aquaculture ponds, (B) recent deforestation, (C) oil palm plantation, (D) rice paddy, (E) coastal erosion, (F) urban development, (G) mangrove forest, and (H) terrestrial forest. Variable scales.

Fig. S2.

Locations of the 1,500 deforested patches that compose the dataset used for model training and validation. Random points are indicated as black points, and those that were also selected to more evenly sample the range of habitat types are indicated as red crosses.

Table S2.

Number of deforested patches in the training dataset identified in each category

Land use category	Number of deforested patches
Aquaculture	374
Deforested	157
Oil palm	218
Rice field	218
Coastal erosion	83
Urban	144
Mangrove	182
Terrestrial cover (non-oil palm)	124

The training dataset of 1,500 deforested patches was used within a series of random forest classification models. Our classification problem was slightly different from the majority described in the remote sensing literature, which commonly classify at the level of single pixels within an image. In contrast, each area of interest in our study was a patch of variable size that contained a number of pixels. To prevent pseudoreplication caused by sampling multiple pixels from each patch, and to keep large deforested patches from having a disproportionate influence on the classification model, we sampled one pixel from each patch at random each time the classification was performed. This classification process was repeated 12 times, because the mean size of the deforested patches was 12 pixels. Each classification model was used to classify the 45,540 deforested mangrove patches, and the most commonly assigned category was assumed to be 2012 land use.

To assess the accuracy of the classification procedure, we carried out 100 cross-validations by randomly splitting the training dataset. In each cross-validation, we used 80% of our data to train a model, and compared the predictions from the model for the remaining 20% against their actual classification. Only randomly selected training data were used to test the classification accuracy. The median Cohen κ value of the 100 bootstrap models was 0.62, and the median accuracy of the whole classification was 68% (Table S3). To assess the potential impacts of systematic errors on the conclusions of the study, we calculated error-corrected estimates of the areal coverage of each land use and SEs for these estimates following recommended guidelines (49) for the overall and annual percentages of mangrove converted to each alternative land use (Fig. S3 and Table S4). More detailed descriptions of the methods and accuracy assessment are provided in SI Methods.

Table S3.

Confusion matrix showing the number of test data classified in each land use category (average of the 100 bootstrap runs)

	Reference
Map	Aquaculture	Deforestation	Oil palm	Rice	Erosion	Urban	Mangrove	Terrestrial
Aquaculture	57	6	4	3	0	1	1	1
Deforested	5	13	1	1	0	0	2	1
Oil Palm	2	1	36	2	0	3	4	7
Rice	11	3	1	49	0	0	0	2
Erosion	0	1	0	0	3	1	0	1
Urban	3	0	2	1	1	14	0	0
Mangrove	2	3	1	0	0	1	25	2
Terrestrial	0	1	8	0	0	0	3	10

Fig. S3.

Error-corrected estimates of the areal coverage of each land use type in 2012. Error bars indicate 1 SE. Open circles indicate the uncorrected estimate of areal coverage made by the classification model.

Table S4.

Uncorrected classification estimates, error-corrected estimates, and SEs of the percentage coverage in 2012 of eight land uses, in patches of mangrove forest that were lost between 2000 and 2012

Land use category	Classification estimate	Corrected estimate	SE
Aquaculture	30	30	2.03
Deforested	6	9	1.56
Oil palm	16	15	1.57
Rice field	22	18	1.48
Coastal erosion	2	2	0.65
Urban	4	5	0.98
Mangrove	15	15	1.49
Terrestrial	4	7	1.35

SI Methods

Identification of Deforested Patches.

We used the global forest change dataset (24) as an index of annual deforestation between 2000 and 2012. This dataset is produced from Landsat satellite imagery, with a spatial resolution of 1 arc-second (∼30 × 30 m) per pixel (24). The dataset provides a map of global deforestation during the study period, and in all deforested pixels the year of deforestation is recorded. There have been criticisms that the global forest change dataset uses a broad definition of forest that includes plantations and other seminatural forests as well as natural forest cover (24); however, by cross-referencing the location of deforested pixels with the global distribution of mangroves in 2000 (8), we identified only the areas of mangrove that existed in 2000 and have since been deforested. It is possible that mangrove forest has established in new areas since 2000, but more recent distribution data are not available. Our results therefore indicate the rate of forest loss rather than the net change in mangrove area between 2000 and 2012.

We focused our study on the mangrove forests of Southeast Asia, a region containing at least 30% of global mangrove cover (8). We used the United Nations geoscheme to define the area of Southeast Asia, which includes Brunei, Cambodia, Indonesia, Malaysia, Myanmar, Philippines, Singapore, Thailand, Timor-Leste, and Vietnam. Laos was excluded from this analysis owing to its landlocked status. By cross-referencing the global forest change dataset with the 2000 mangrove cover dataset and the extent of Southeast Asia, we were able to map the distribution of mangrove deforestation in this region every year between 2000 and 2012.

To identify continuous patches of mangrove that were deforested simultaneously (during the same year), we first extracted the pixels that were deforested in each year. We then processed the pixels using the “clump” function in the R package “raster” (cran.r-project.org/web/packages/raster/index.html). This function detects continuous patches of pixels that are connected in any one of eight directions.

For further analysis we used only deforested patches larger than 0.5 ha (made up of more than five pixels), because we judged that smaller patches may result from classification error in the underlying global forest change dataset. The above process provided a dataset of 45,540 patches for analysis, with a median size of 1.08 ha and a mean size of 27.9 ha. The largest patch of mangrove deforested in any year was 374.2 ha.

Land Use Classification of the Deforested Patches.

Predictor Variable Data Collation.

We collated eight datasets that describe the geographical and environmental characteristics of the deforested patches. For each deforested plot, we extracted the pixels from a Landsat satellite image from 2012 (24), and recorded the values for four satellite image bands that are commonly used in remote sensing of vegetation: the red band (band 3), near-infrared band (band 4), and short-wave infrared bands (bands 5 and 7). These data were extracted from a preprocessed Landsat 7 image provided as part of the global forest change dataset (24), and were created by compositing the latest cloud-free imagery for each pixel (typically from 2013, but no earlier than 2010).

In addition, we calculated four patch-level predictor variables at the central point of each patch, the Normalized Difference Vegetation Index in the surrounding 25 ha, the distance from the nearest major road at the center of the patch, and the FAO suitability index for oil palm cultivation and rice agriculture (Table S1). These additional patch-level predictor variables provided contextual data on the location surrounding each deforested patch.

Identification of Current Land Uses via Google Earth Pro.

We collected a dataset of current land use in a subset of 1,500 deforested patches, to build up a training dataset for remote sensing. Land use was inferred from high-resolution Google Earth Pro images, which include DigitalGlobe, Astrium, and TerraMetrics images. Deforested mangrove patches were categorized into one of eight land use/land cover types: aquaculture, oil palm plantation, rice field, urban development, terrestrial forest (including possible rubber plantations and natural terrestrial regrowth), mangrove forest, recent deforestation with an unclear replacement land use (including apparent logged patches and areas of natural tree loss patches), and deforestation due to coastal erosion. Fig. S1 provides example imagery of these eight categories in Google Earth Pro. An additional land use type, salt extraction pond, was originally considered but was combined with aquaculture owing to the small number identified in the training sample (two ponds out of 1,500 patches). Our classification method assumed that each deforested patch had only one replacement land use.

A random subset of all deforested patches was viewed. To gain a representative sample of larger deforested patches, a random subset of deforested patches larger than 2 ha (22 pixels) was viewed as well. Further areas were then targeted to sample different habitat types more evenly; over two targeted iterations, deforested patches in specific areas were chosen based on our knowledge of the likely key drivers of deforestation. Patches were excluded from the training sample if they were not clearly visible (e.g., if covered in clouds or if the resolution of the available satellite imagery was too low). Some patches were also excluded if the date of deforestation was after the date of the satellite image capture, although some flexibility was allowed to account for possible variation in the true date of deforestation (deforestation may not be recorded in the correct year if the area was covered by cloud during the intervening period), and uncertainty about the true date of the high-resolution Google Earth images. A total of 3,091 deforested patches were viewed. The training dataset showed good spatial coverage across Southeast Asia, although there is a lack of detailed imagery for Indonesia, particularly Sumatra (Fig. S2).

Image Classification of Deforested Patches.

Random forests were used to build the classification model that predicted land use category based on the eight predictor variables described above. Random forests were constructed by applying the R package “randomForest” (cran.r-project.org/web/packages/randomForest/index.html) to the training dataset. In the remote sensing literature, it is common to treat each pixel in an image as an individual unit of analysis. However, our units of analysis were patches of pixels, and the number of pixels within each patch varied substantially. Furthermore, some of our predictor variables were measured at the patch scale, whereas satellite imagery was available for each pixel within each patch. To avoid the issue of pseudoreplication that would be caused by using multiple pixels within each patch, and to prevent large deforested patches from having a highly influential impact on the classification model, we randomly selected one pixel to represent the conditions within each patch. We repeated this random selection 12 times, because the mean size of the deforested patches was 12 pixels. We constructed separate classification models for each of the 12 samples. For each deforested patch, the predicted 2012 land use type was defined as the classification most commonly assigned by the 12 random forest models.

Accuracy Assessment.

We assessed the accuracy of the classification model outlined above using a bootstrap cross-validation technique. For each of 100 bootstrap replicates, the training dataset was split randomly into two groups; 80% of the dataset (1,200 samples) was used to train the random forest models, and the remaining 20% served as test data. The land use of each test patch was predicted using this training model, and the resulting predicted land uses were compared with the actual land uses as recorded from Google Earth Pro. Only the 1,008 training samples that had been chosen at random were used to test the model accuracy; all 492 of the nonrandomly chosen reference patches were included in the training subset each time.

The median accuracy from the 100 bootstrap random forest classification models was 68%. The median Cohen κ value was 0.62, which represents a “substantial” level of agreement between the classification model and sampled data (50). There was variation in classification accuracy between the eight land use categories (Table S3). The greatest inaccuracies were found in the classification of terrestrial forests (41% correct classification), which were misclassified as oil palm in 30% of the test samples and mangrove in 10% of the test samples (Table S3). Similarly, the deforested class had a relatively low accuracy (48%) and was misclassified as aquaculture (22%), rice (10%), or mangrove (10%) (Table S3). Inaccuracies in classifying these two categories are not surprising, given that the deforested class represents a transition between mangrove and other land use types, whereas terrestrial forest includes miscellaneous non-mangrove and non-oil palm types. The classification accuracy for the three classes of most interest in the present study—aquaculture, rice, and oil palm—was 74%, considerably higher than the overall accuracy. The accuracy levels reported in the present study are comparable to those reported in classifications at similar spatial scales and with multiple land classes, including Southeast Asian land cover classifications (51) and several global analyses (52), including one derived from Landsat data (53).

To assess whether classification inaccuracy made it difficult to draw conclusions from this study, we calculated error-corrected estimates of the total area of each replacement land use type, and SEs for these estimates, using the method proposed by Olofsson et al. (49) (Table S4). A comparison of the original classification estimates and the error-corrected estimates is provided in Fig. S3. In all cases, the uncorrected estimates were close to the corrected estimates, and the corrected and uncorrected estimates were highly similar for the aquaculture, oil palm, erosion, urban, and mangrove classes (Fig. S3). The uncorrected estimate of rice was slightly higher than the corrected estimate, whereas the uncorrected estimates for deforested and terrestrial were slightly lower than the corresponding corrected estimates (Fig. S3). The inaccuracies caused by classification errors do not affect the overall picture of mangrove deforestation in Southeast Asia between 2000 and 2012; aquaculture, rice agriculture, and oil palm were the biggest drivers of mangrove change. Thus, we used the uncorrected classification model to calculate the overall percentage coverage and the country-level percentage coverage of the land use classes reported in this paper. The temporal comparison of each land use type (Fig. 3) reports the error-corrected percentage coverage of the land use types and the corresponding SEs.