Spatiotemporal analysis of deforestation patterns and drivers reveals emergent threats to tropical forest landscapes

As deforestation breaches into new tropical frontiers, proactive conservation strategies require a trifecta of information on where deforestation is accelerating (emergent), how drivers of deforestation vary spatiotemporally, and where to focus limited conservation resources in protecting the most integral yet threatened forested landscapes. Here we introduce Emergent Threat Analysis, a process integrating Emerging Hot Spot Analysis of deforestation, visual classification of deforestation outcomes over time, and spatial quantification of contemporary forest condition. We applied Emergent Threat Analysis to tropical Southeast Asia, a global epicentre of biodiversity threatened by deforestation. We found that emergent hot spots (EHS)—a subset of hot spots characterized by strong, recent, and clustered patterns of deforestation—accounted for 26.1% of total forest loss from 1992 to 2018, with deforestation within EHS proceeding at 2.5 times the regional rate of gross loss. Oil palm and rubber plantation expansion were the principal drivers of deforestation within EHS of insular and mainland SE Asia, respectively. Over the study period, oil palm shifted in importance from Sumatra and Sarawak to Papua and Kalimantan, whereas rubber became prominent in Cambodia and Tanintharyi from 2006 to 2015. As of 2019, more than 170 000 km2 of SE Asia’s remaining forest occurred within EHS, of which 21.7% was protected. High and medium-integrity forest constituted 19.2% and 49.1% of remaining EHS forest, respectively, but of these, 35.0% of high-integrity and 23.9% of medium-integrity EHS forest were protected. Because we anticipate that tree plantation expansion will continue to drive deforestation in SE Asia, significantly heightened protection is needed to secure the long-term preservation of high and medium-integrity forest, especially in highly contested forest frontier regions. Finally, as a flexible, integrated process, Emergent Threat Analysis is applicable to deforestation fronts across the global tropics.


Introduction
Deforestation is the most pervasive, imminent, existential threat to tropical forests worldwide; it results in a myriad of negative environmental outcomes including biodiversity loss, ecosystem service degradation, greenhouse gas emissions, and climate change (Gibbs and Herold 2007, Baccini et al 2012, Lawrence and Vandecar 2015, Giam 2017. In response, national policies along with international coalitions, treaties, and campaigns have evolved with the objectives of halting and reversing deforestation trends (IUCN, UNEP, andWCMC 2021, NYDF Assessment Partners 2021). The success of these campaigns is contingent in part, on the capacity to measure and track progress towards achieving deforestation reduction goals, which highlights an increasing need to leverage geospatial approaches to efficiently and effectively monitor, evaluate, and report deforestation trends at multiple scales. Even more pressing is the need to know where deforestation threats to tropical forest frontiers are accelerating, what the drivers of deforestation are and how they vary across space and time, and where research, policy, and conservation interventions should proactively target to mitigate those new threats.
Numerous studies have quantitatively analysed the spatiotemporal variation in deforestation patterns across the tropics (Hansen et al 2010, 2013 to name a few); however, owing to the historical nature of geospatial data, they are typically retrospective in scope. While useful in deepening our understanding of landscape change processes, reliance on these analyses can render conservation science, policy, and intervention reactive to rapidly emerging trends. Significant efforts have been made to make geospatial data contemporaneous with deforestation occurrence (Hansen et al 2016, Reiche et al 2021; nevertheless, tools are urgently needed to identify emerging frontiers of deforestation so that proactive policy and intervention strategies can be developed. One such tool is Emerging Hot Spot Analysis (Harris et al 2017), a method that integrates spatial statistics and trend analysis to identify areas of recent and accelerating deforestation concern (which we refer to as 'emergent hot spots').
Far from being a homogeneous process, deforestation results from complex interactions of proximate and underlying drivers that vary considerably across spatial, temporal, and institutional scales (Geist and Lambin 2002, Bürgi et al 2005, Hersperger et al 2010, Hosonuma et al 2012, Lim et al 2017, De Alban et al 2020, Verma et al 2021. This is particularly true for dynamic deforestation environments such as emergent hot spots (EHS), which are likely to be subjected to rapidly changing deforestation drivers Coomes 2016, De Alban et al 2019). To respond to these dynamic threats, precise spatiotemporal attribution of proximate deforestation drivers is needed, which can be further linked to underlying drivers of deforestation and translated into targeted policy responses or interventions (Finer et al 2018). Indeed, although land cover data sets have greatly improved in temporal resolution, they remain unable to offer the same level of land use and land cover discrimination as very high-resolution (VHR) satellite imagery, which can be visually classified with high accuracy and efficiency, and at increasingly reduced barriers to availability and accessibility (Schepaschenko et al 2019).
Besides information on emergent trends and drivers of deforestation, conservation intervention must also strategically prioritize efforts and resources to where they are most needed. For example, the expansion of protected areas has beenand is expected to remain-the gold standard for conservation Phillips 2006, Maxwell et al 2020). Moreover, conservation organizations require enhanced information on where integral, yet threatened landscapes are, to justify directing limited conservation funding. To effectively safeguard biodiversity in landscapes facing emergent patterns and drivers of deforestation, conservation deployment requires a spatially informed understanding on contemporary forest health/condition, and its contribution to ecosystem function (Dinerstein et al 2017). This can be combined with information on existing levels of protection (Grantham et al 2020), which can then guide conservation planning to focus on the most threatened, high-value landscapes in a more proactive, rather than reactive, manner.
Here we introduce the Emergent Threat Analysis, which we define as the combined process of quantifying EHS of deforestation, precisely attributing the spatiotemporal drivers of deforestation, and identifying the most threatened high-conservation-value forests within those EHS. Emergent Threat Analysis differs from previous approaches to studying deforestation by virtue of its novel combination of methods that have all been utilized independently (Harris et al 2017, Grantham et al 2020, but which have yet to be fully integrated with the purpose of using dense, historical geospatial information to indicate priority regions for proactive and enhanced research, policy, and conservation attention.
Emergent Threat Analysis is critically needed in Southeast Asia, a region where globally important forests for biodiversity and carbon sequestration (Myers et al 2000, Sodhi et al 2004, Mitchard 2018 continue to face an emergent deforestation threat (Feng et al 2021, Hoang andKanemoto 2021); the region is also where the highest rates of tropical deforestation have occurred in the past four decades (Miettinen et al 2011, Hansen et al 2013, Stibig et al 2014. Deforestation in SE Asia is driven primarily by agricultural expansion (Lepers et al 2005, Rudel et al 2009, Stibig et al 2014, with commoditydriven deforestation in the region outpacing all others worldwide from 2000 to 2015 (Curtis et al 2018). Indeed, SE Asia now accounts for the vast majority of global production of palm oil and natural rubber (Kenney-Lazar and Ishikawa 2019), two commodities that have been central in shaping regional and global agrarian economies. In addition to agriculture, deforestation in SE Asia is attributable to a wide range of spatiotemporally variable drivers that include logging, mining, aquaculture, and urbanization (Richards and Friess 2016, Hughes 2017, Lim et al 2017, De Alban et al 2020, Jayathilake et al 2020.
In this study we apply Emergent Threat Analysis across the entire region of SE Asia. First, we identify and quantify EHS of deforestation, which we contextualize using annual, historical patterns of deforestation from 1992 to 2018. Second, we characterize how land cover outcomes of deforestation within EHS have changed over time, to implicate the spatiotemporal dynamics of deforestation drivers.
Third, we combined the EHS analysis with forest landscape integrity to highlight conservation priorities and opportunities in the most threatened forest frontiers.

Quantifying dynamics of forest cover and change
Our analyses used the 1992-2018 European Space Agency (ESA) 300-m annual land cover dataset. We used this data because its long temporal domain was favourable for long-term trend analyses of forest change dynamics (De Alban et al 2019, Mousivand and Arsanjani 2019), and because it allowed us to incorporate pre-2000 deforestation data, which represented a period of intense deforestation in many parts of SE Asia. We projected the ESA data to the Equal-Area Scalable Earth Grid-2.0 cylindrical equalarea projection, which maximally preserved areal attributes over large geographical extents (Brodzik et al 2012).
To quantify forest change, we reclassified the ESA data into binary maps of forest and non-forest. Using the Intergovernmental Panel on Climate Change categorization in the ESA classification scheme (ESA 2017), we aggregated the eight ESA forest categories (broadleaved evergreen, broadleaved deciduous, needleleaved evergreen, needleleaved deciduous, mixed broadleaved/needleleaved, mosaic tree/shrub, flooded freshwater/brackish, and flooded saline water) into a single 'Forest' category, and reclassified all others as 'Non-forest' . We then differenced pairs of binary forest/non-forest maps using Google Earth Engine to create 26 interannual forest change maps, and one for the entire 1992-2018 interval (i.e. differencing 1992 and 2018). The 1992-2018 forest change map was used to quantify overall forest change statistics over the 26-year period, while the interannual change maps were used to construct the annual time series of net and gross forest change, and to quantify the amount of loss that occurred within EHS.
We then tessellated the SE Asian region into 5-km circumradius hexagons to aggregate data on interannual gross forest loss (supplementary methods (SM) 1.2, 1.3, and figure S1 available online at stacks. iop.org/ERL/17/054046/mmedia). Aggregating pixel data into hexagons facilitated spatial analysis of forest cover and change. Further, the hexagons provided spatially equivalent analytical units as required for hot spot analysis (next section). Given the area of each hexagon (∼65 km 2 ), they are referred to as 'landscapes' in the results (section 3) .
To investigate the spatial relationships between initial forest cover and gross deforestation across SE Asia for the 26-year period, we created a bivariate classification of initial (1992) forest cover and 1992-2018 gross loss. Initial forest cover was calculated as the percentage of hexagon area while gross loss was calculated as the percentage of initial forest cover. Based on the respective data distributions for each metric, initial forest cover was binned into marginally (<30%), moderately (30%-72%), and heavily (>72%) forested hexagons, whereas gross loss area was binned into weak (<10%), moderate (10%-42%), and intense (>42%) deforestation (SM 1.1). Binning was performed in R using the classInt package (Bivand et al 2020).

Identifying EHS of deforestation
We identified and quantified EHS deforestation using the Emerging Hot Spot Analysis (Harris et al 2017, hereafter referred to as Hot Spot Analysis). Hot spots were identified by analysing the amount of forest loss in a hexagon neighbourhood in relation to the total amount of loss within a predetermined landscape, using Getis-Ord G i * to test for spatial clustering of deforestation within a hexagon neighbourhood, and the Mann-Kendall trend test to evaluate for temporal trends of deforestation (Mann 1945, Kendall and Gibbons 1990, Getis and Ord 1992, Harris et al 2017, ArcGIS 2021. Following Harris et al (2017), we analysed SE Asia by subregions in order to identify hot spots that were likely driven by and related to subregionspecific social, political, and economic contexts. We defined subregions using national-level boundaries from the Global Administrative Areas data, except for the following Indonesian island subregions: Sumatra, Kalimantan, Java and Nusa Tenggara, Sulawesi and Maluku, and Papua (adapted from Margono et al 2014). Singapore, Brunei, and Timor Leste were subsumed within their neighbouring regions owing to their small land areas. In total, there were 12 subregions for Hot Spot Analysis. For each subregion, we aggregated pixel-scale information from the forest change maps into the hexagonal grid, resulting in a spatially explicit, annual time series of subregional deforestation density from 1992 to 2018 (SM 1.4). We then performed Hot Spot Analysis for each subregion and masked out areas of forest/non-forest persistence (table S1, and SM 1.5).
Hot Spot Analysis categorizes deforestation patterns per hexagon into hot spots (eight types), cold spots (eight types), or without a pattern (ArcGIS 2021). As our emphasis was on areas of recent and accelerating deforestation, we focused on consecutive, intensifying, and persistent hot spots. Consecutive hot spots meant that deforestation trends developed over an uninterrupted series of time intervals (ArcGIS 2021). Persistent hot spots were identified when deforestation was consistently clustered over time (ArcGIS 2021). Intensifying hot spots were similar to persistent hot spots, but in addition, they were identified when the intensity of deforestation clustering increased over time (ArcGIS 2021). We consolidated consecutive, intensifying, and persistent hot spots into EHS because these three hot spot types exhibited (1) strong patterns of clustered deforestation over time, and (2) were identified in the final time step, thereby including an element of recency (table S2 and figure S2). The other hot spot types (i.e. diminishing, historical, oscillating, sporadic, and new) did not reflect strong, recent, and developing trends of clustered forest loss, and thus did not represent as much of a growing deforestation threat (SM 1.6). We then merged the 12 subregional Hot Spot Analysis outputs together to create the EHS map for SE Asia, which was used for subsequent analyses.

Classifying land cover outcomes of hot spot deforestation
We conducted a sample-based visual assessment of land cover changes associated with hot spot deforestation. Using a stratified random sampling approach, we selected interannual forest loss pixels within each EHS region, and grouped them into three-year intervals (e.g. 2000-2003, 2003-2006, etc; the 2015-2017 period only contained two interannual intervals). Distance between selected pixels was ⩾5 km to ensure independence. The initial sample count of 1146 was distributed across subregions in proportion to the total amount of interannual forest loss from 2000 to 2017; no samples were allocated in the EHS of Java, Nusa Tenggara, Timor Leste, and the Philippines, each of which accounted for ∼1.0% of total loss during this period (table S3). Samples were filtered to include those with sufficient, interpretable images to confidently ascertain both (1) the deforestation event (false positives were rejected) and (2) the resultant land cover following deforestation. After filtering, 571 samples were used in the visual assessment of land cover outcomes (figure S3).
We classified 14 deforestation land cover outcomes under general themes of tree plantations, farming systems, degraded landscapes, and other land ( figure S4). Tree plantations consisted of oil palm, rubber, pulpwood, cashew, and coconut. Farming systems included rice cultivation, mixed farming/homegardens, and shifting cultivation. Degraded landscapes consisted of shrub/grassland and bare ground that could be intermediate or permanent; a special category of this was degraded forest, where reductions in forest cover were observed over time without evidence of deforestation. Other land included built up areas, artificial water bodies, and mining.
We set five years post-deforestation as the window to classify land cover outcomes of deforestation because our preliminary assessment revealed that oil palm, rubber, and pulpwood plantations were identifiable within that time frame if the land was immediately cultivated. In this case, the immediate post-deforestation outcome was assumed to be the deforestation driver because the conversion was likely driven by an intended land use outcome. In some cases, however, the conversion of forest into a longterm land cover outcome may be delayed, inducing a transitory intermediate land cover in the process. To include these transitions, we also evaluated the land cover of all samples in 2021; however, we did not assess 2021 land cover for samples deforested in 2017 because the initial and final outcomes would have been synonymous.
Classification of land cover outcomes was performed using Collect Earth (Bey et al 2016), in which visual interpretation of VHR imagery in Google Earth was the primary analysis, supplemented when necessary by annual Landsat and Sentinel imagery in Google Earth Engine. Infrared false colour composites from Landsat and Sentinel offered contextual information regarding the evolution of land cover over time and allowed us to discriminate between rubber and pulpwood plantations, which have similar canopy morphologies. Data entry was achieved through a customized survey form in Open Foris Collect (https://openforis.org/tools/collect/). Land cover outcomes were quantified as a percentage of subregional samples in each three-year interval.

Evaluating contemporary forest integrity and protection
We leveraged the 2019 Forest Landscape Integrity Index (FLII) to quantify the integrity of naturally occurring or regenerating forest that remained in EHS. FLII delivered 300-m information on forest integrity that was calculated as a function of forest extent, human pressure, and landscape connectivity (Grantham et al 2020). Before integrating the FLII data, we used the 2018 ESA forest cover map as a mask to exclude tree plantations and urban greenery from that dataset (SM 1.7). We then discretized the continuous FLII data into high, medium, and low integrity, using the thresholds defined by Grantham et al (2020). High-integrity forest (FLII ⩾ 9.6) represented core areas that were largely unmodified and with ecosystems functioning at near-natural levels. Medium-integrity forest (FLII > 6.0 and < 9.6) had sustained higher degrees of modification, resulting in lower provisioning of ecosystem services. Lowintegrity forest (FLII ⩽ 6.0) referred to areas that were extensively modified, fragmented, and were expected to have weak contributions to ecosystem function. Given that the FLII categories carry significant ecological relevance, we were able to identify the areas of highest conservation priority in landscapes threatened by deforestation.
We then quantified protected area coverage within EHS and stratified this analysis by forest integrity to identify where current gaps in the protection of threatened high and medium-integrity forest existed. Protected area data were primarily sourced from the International Union for Conservation of Nature World Database on Protected Areas (IUCN, UNEP-WCMC 2020), and the Wildlife Conservation Society Myanmar. We dissolved the protected area data into a single layer to prevent double-counting (UNEP-WCMC 2015), and clipped it to each EHS region. By spatially joining protected area and forest integrity data within EHS, we were able to quantify the amount of protected and unprotected forest that was threatened with emergent deforestation.

Spatiotemporal dynamics of deforestation, 1992-2018
In 1992, forest cover in SE Asia was 2343 773 km 2 , of which 65.7% was in insular SE Asia (table 1). From 1992 to 2018, gross forest loss was 219 833 km 2 , balanced by a gain of 143 215 km 2 for a net loss of 76 618 km 2 (−0.13% yr −1 , table 1). Insular SE Asia accounted for 98.7% of total net loss and 65.2% of total forest change (gross losses + gross gains) (table 1). Thus, total forest change in insular and mainland SE Asia was proportional to the amount of 1992 forest cover, but net losses were almost all contained in insular SE Asia (figure S5). Cambodia was an outlier in mainland SE Asia, with annual net losses of −0.60%, comprising a unique combination of high gross loss and almost no gross gain (table 1).
Interannual forest change varied considerably across the subregions, exhibiting six temporal trajectories (figure 1). Consistent net losses occurred for more than 80% of the time period in Sumatra, Kalimantan, East Malaysia, and Sulawesi and Maluku. Attenuating losses, which indicated reduced deforestation rates over time, occurred in Cambodia, Vietnam, and Singapore. Thailand exhibited a period of consistent net forest loss until 2003 followed by consistent net forest gain, i.e. a 'forest transition' sensu (Mather 1992). Oscillating patterns, which occurred when forest change cycled between periods of net losses and gains were observed in Papua, Brunei, and Laos. West Malaysia exhibited a short period of net gain that was succeeded by an uninterrupted period of net loss that spanned 75% of the time series. Finally, net gains for more than 60% of the time period were recorded in Myanmar, Java and Nusa Tenggara, Philippines, and Timor Leste.
Through bivariate mapping of forest cover in 1992 and gross forest losses for 1992-2018 (figure 2), we found that heavily forested landscapes (i.e. hexagons with > 72.0% forest cover) contained 70.3% of total forest cover in 1992 and accounted for 48.1% of gross losses (table S4). Moderately forested landscapes (i.e. hexagons with 30.0%-72.0% forest cover) were targeted for deforestation, as evidenced by the fact that they contained 24.5% of initial forest cover but accounted for 42.4% of gross loss (table S4). Heavily and moderately forested landscapes that experienced intense deforestation-defined as ⩾ 42.0% reduction in initial cover-contained 7.0% of all SE Asia's forest in 1992 and accounted for 50.6% of gross loss from 1992 to 2018 (table S4). These were the areas where deforestation over the 26-year period was most concentrated; observed in Sumatra, Kalimantan, Sarawak, south Myanmar (i.e. Tanintharyi), Cambodia, and central Vietnam (figure 2).

EHS of deforestation
Gross forest loss occurred in a diverse mosaic of hot and cold spots that indicated a complex mélange of deforestation dynamics across SE Asia (figures S6 and S7). Interannual gross forest loss from 1992 to 2018 totalled 233 950 km 2 (table S5), of which EHS accounted for 61 140 km 2 , i.e. 26.1% (table 2). EHS deforestation led to a 24.0% gross reduction in 1992 forest cover, with gross losses occurring at a rate of 0.92% yr −1 , which was 2.5 times the regional rate of gross loss (0.36% yr −1 , table 1).

Land cover outcomes of deforestation within EHS
The most prevalent land cover outcome of EHS deforestation was tree plantations, which comprised 45.7% of initial and 62.2% of final outcomes (table 3). Tree plantations, mostly in the form of oil palm plantations, dominated in all insular EHS (53.2% → 67.4%, initial to final, table 3) except Sulawesi, which was dominated by conversions into shrub/grassland (50.0%, table S6). Over time, the importance of oil palm as a deforestation outcome declined in Sumatra and Sarawak, whereas it increasingly drove deforestation in Papua and Kalimantan (figure 3). In mainland SE Asia, tree plantation outcomes had an initial occurrence rate of 21.5%, but this increased to 45.2% by 2021 (table 3). Rubber plantations were the primary tree plantation type across mainland EHS (11.1% → 23.0%, table 3). They were most prominent in EHS of Cambodia and Tanintharyi, predominating from 2006 to 2015 (figure 3) when rubber prices were increasing (figure S8).  Degraded landscapes accounted for 44.0% of all initial outcomes and was predominately shrub/grassland in all EHS regions (except Papua, table S6), followed by bare ground and degraded forest (table 3). However, 49.0% of degraded landscapes were subsequently converted to another land cover (figure 4), reducing its importance to 27.7% (table 3). Such conversions from intermediate to final outcomes generally followed the subregional trends of direct deforestation, i.e. towards oil palm and rubber plantations (figure 4, tables S7 and S8). A notable exception was cashew, in which a majority of outcomes arose from six types of intermediate transitions in mainland EHS (table S8), reflecting its growing prominence amongst other land cover outcomes. Figure 2. Forest cover in 1992 and gross forest loss from 1992 to 2018, where each map unit is a 5-km circumradius hexagon (∼65 km 2 ). Y-axis breaks were defined using a head-tails (heavy-skew) method, while X-axis breaks were defined using a natural breaks (jenks) method to suit their respective data distributions. Of the farming-related outcomes, shifting cultivation made only a minor contribution to EHS deforestation (table 3), and was most prominent in southern Laos (table S6). Further, of all initial outcomes identified as shifting cultivation, 55.6% were converted into rubber (33.3%) and cashew (11.1%) plantations, and mixed farming/homegardens (11.1%), while the remainder were identified Table 3. Percentage of initial and final land cover outcomes of deforestation in insular and mainland SE Asia. Initial outcomes refer to the land cover at five years post-deforestation, while final outcomes refer the 2021 land cover. 'n' denotes the number of samples. as active swidden or shrub/grassland that arose from swidden abandonment or fallow.

Forest integrity and protection in EHS
As of 2018, EHS contained 9.0% of total remaining forest in SE Asia (table 2). Based on the FLII, 170 737 km 2 of forest remained within EHS, of which 64.4% occurred in insular SE Asia (table 4). Overall, 19.2% of remaining EHS forest in SE Asia exhibited high integrity, 49.1% medium integrity, and 31.7% low integrity; these percentages were similar for both insular and mainland EHS (table 4). High-integrity forests comprised <10.0% of forest area in most EHS regions (table 4). Notable exceptions of EHS with substantial proportions of highintegrity forest were Papua (47.8%, 15 117 km 2 ), Sulawesi (20.6%, 4636 km 2 ), southern Myanmar (36.0%, 4319 km 2 ) and western Thailand (50.2%, 4191 km 2 ) (the latter two encompass the Dawna Tenasserim landscape, figure 5). A majority of the remaining EHS were dominated by medium-integrity forest, such as in central Sumatra, southern Laos, and central Vietnam (table 4); only a small proportion of high-integrity forest remained in these areas (figure 5).
The largest area of unprotected high-integrity forest was 14 045 km 2 , in Papua's EHS where only 7.1% of high-integrity forest was protected (table 4). Other EHS with >1000 km 2 of unprotected high-integrity forest include Sulawesi, Laos, and Myanmar. Medium and low-integrity forest formed the bulk of unprotected forest (table 4), but the proportional contributions of these forest types varied considerably across EHS subregions (figure 6).

Discussion
Oil Heightened protection is thus needed to effectively mitigate against the loss of such forests, but these could be coupled with alternative strategies to develop oil palm on underutilized or degraded land (Goh et al 2018), or with appropriate oversight and enforcement pursuant to strict certification standards such as No Deforestation, No Peat, No Exploitation commitments (European Palm Oil Alliance 2021).
In the case of rubber plantations in mainland SE Asia, we show that rubber-driven deforestation was prominent in EHS of Tanintharyi and Cambodia, particularly after 2006 when rubber prices were increasing (figure S8). Given that there is a demonstrable link between rubber prices and deforestation (Grogan et al 2019), and that rubber is typically considered a long-term investment that smallholder farmers-who form the bulk of rubber production (Fox and Castella 2013)-are reluctant to convert, it is plausible that rubber will remain a major driver of deforestation, particularly when prices trend upward, as has been the case since early 2019 (figure S8). Based on our analysis, we expect future expansion to occur in EHS along the Tenasserim and Annamite mountain ranges, thus supporting previous studies that forewarned of rubber expansion towards marginal environments and higher elevations (Ahrends et al 2015, Zeng et al 2018. Considering that these regions still possess substantial forest cover, a resurgence of rubber-driven deforestation would present a double threat of habitat loss and localized warming (Zeng et al 2020) that would increasingly threaten biodiversity.
Our analyses revealed shifting cultivation as a minor driver of EHS deforestation, which contrasts with narratives identifying it as the primary cause of widespread deforestation across SE Asia (see Fox 2000. Where swidden occurred, we found that a majority were converted to tree plantations and mixed farming systems, reflecting the regionwide transition of traditional swidden to more permanent and intensified forms of agricultural production (Vongvisouk et al 2014, Suhardiman et al 2019. The low importance of shifting cultivation as a deforestation driver in this study may in part be because swidden fields tend to have a small individual footprint (i.e. smaller than an ESA 9-ha pixel), and do not characteristically exhibit spatiotemporal clustering.
Of all remaining forest in SE Asia, 9.0% were found in EHS; these are the areas most actively threatened with habitat loss, indicating a need for conservation prioritization. However, only 35.0% of high-integrity and 23.9% of medium-integrity EHS forest is protected; these values are lower than the global averages reported by Grantham et al (2020), where 52.3% of high-integrity and 30.3% of mediumintegrity forest are protected globally. Our results further demonstrate that this shortfall in existing protection of high and medium integrity forests is severely pronounced in EHS of insular SE Asia, where only ∼15% of both high and medium-integrity forest are presently protected, which contrasts starkly with EHS of mainland SE Asia where 69.9% of high-integrity and 37.8% of medium-integrity forest are protected. These findings therefore suggest-solely from a standpoint of spatial coverage-that the current protected area network is likely to be insufficiently poised to deal with frontier deforestation in the region, especially in insular SE Asia.
Despite only comprising 19% of all threatened forest, high-integrity forest should be prioritized due to their unrivalled importance in the provision of multifarious ecosystem services (Grantham et al 2020). Indeed, their loss represents a keystone environmental problem in that it would trigger disproportionately large and long-term consequences that could offset decades of conservation progress and climate action, particularly in areas containing vulnerable, yet irreplaceable carbon stocks (Noon et al 2021). By contrast, medium-integrity forest, which contribute substantially to ecosystem functionalbeit at reduced levels due to their semi-disturbed state (Grantham et al 2020)-constituted 49% of all threatened EHS forest. Therefore, conservation prioritization must carefully balance the trade-offs in securing safeguards for invaluable, yet less extensive high-integrity forest, versus administering immediate interventions for more expansive, yet relatively less protected medium-integrity forest. These findings underscore the importance of systematic conservation planning  to address emergent deforestation in the region, especially in highly contested forest frontiers where we anticipate tree-plantation expansion to continue driving deforestation.

Significance of Emergent Threat Analysis
Emergent Threat Analysis supports proactive conservation strategies through the integration of Emerging Hot Spot Analysis, visual classification of deforestation outcomes over time, and spatial quantification of contemporary forest condition and protection to identify increasingly threatened, high value forested landscapes. Individually, each method provides useful geospatial information on land cover and change; however, Harris et al (2017) noted the opportunity to combine metrics of forest loss clustering with contextual information, and Emergent Threat Analysis integrates and synergizes those information flows. Importantly, the underlying data sets are replaceable according to research needs and data availability. For example, the identification of EHS would remain effective with other forest loss datasets (e.g. Harris et al 2017, Pacheco et al 2021. Further, we use VHR imagery to visually discriminate between deforestation outcomes over time, but this could be substituted with increasingly finer scale imagery (e.g. Planet Team 2017), or augmented with automated classification techniques. Similarly, whereas we used the FLII because it delivered the most contemporary data relevant to the long-term preservation of forest connectivity and ecosystem function (Grantham et al 2020), other geospatial databases relating to the present state, quality, and/or value of extant forest could be used. Given this flexibility, Emergent Threat Analysis is highly relevant to a diverse set of stakeholders ranging from conservation scientists, practitioners, legislators, lobbyists, and donors, who, in congruence with the 2030 global sustainability agenda, share cooperative interests in targeting deforestation fronts across the global tropics (Pacheco et al 2021).

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files). The raw geospatial data can be obtained from the following sources: ESA annual land cover maps (https://maps.elie.ucl.ac.be/CCI/viewer/), Global Administrative Areas version 3.4 (https:// gadm.org/download_country.html), 2019 Forest Landscape Integrity Index (www.forestlandscapeinte grity.com/download-data), and protected areas (www.protectedplanet.net/en (version July 2021), and Wildlife Conservation Society Myanmar).