Conservation Status of Forest Types Vary Greatly in Myanmar - the Most Forested Country in the Indo-burma Biodiversity Hotspot

Diverse forests with distinct forest types, harbor exceptional biodiversity and provide many ecosystem goods and services, making some forest types more economically valuable and prone to exploitation than others. The high rates of deforestation in Southeast Asia endanger the existence of such vulnerable forest types. Myanmar, the region’s largest forest frontier provides a last opportunity to conserve these vulnerable forest types. However, the exact distribution and spatial extent of Myanmar’s forest types has not been well characterized. To address this research gap, we developed a national scale Forest Type map of Myanmar at 20m resolution, using moderate resolution, multi-sensor satellite images (Sentinel-1, Sentinel-2 and ALOS-PALSAR), extensive eld data, and a machine learning model (RandomForest). We mapped nine major forest types and developed a Conservation Status Score to evaluate the conservation status of the mapped forest types. Swamp, Mangrove, Dry Deciduous, Lowland Evergreen and Thorn forests were ranked as the ve least conserved forest types. We also identied the largest remaining patch for each of the ve least conserved forest types and determined their protection status to inform future forest conservation policy. In most cases, these patches lay outside protected areas indicating areas that may be prioritized for future conservation. order and identied the rst ve forest types.


Introduction
The potential natural distribution of different forest types is determined by geographic location (latitude and longitude), climate 1 and local environmental factors, such as elevation 2 , and soil 3 . Spatio-temporal variation in these factors have led to the evolution of a wide range of highly distinct forest types across the globe 4 . Often, these highly diverse forests with many distinct forest types form the core of global biodiversity hotspots and represent critical habitat to rare, endemic, and other plant and animal species 5 . In addition to biodiversity resources, these forest types critically in uence ecosystem goods and services, with some forest types being more widely exploited for economic purposes than others. Over time natural variation modulated by anthropogenic factors via resource use (e.g., rewood, building materials), forest conversion (e.g., conversion to grazing land), targeted management and conservation of forest resources 6,7 result in some forest types becoming increasingly rare, while others expand. In addition, forest types naturally restricted to very speci c ecosystems 8, 9 (e.g., Mangroves, Dry Deciduous forests) are more vulnerable to destruction and exploitation.
Southeast Asia is a major global deforestation hotspot 10,11 where rare and vulnerable forest types such as Indo-Malayan Lowland Sundaic forest, Dry Deciduous Dipterocarp are at risk 9,10 . Often these forest types are not wellrepresented in protected area networks or are located in ineffective protected areas 12 . Myanmar, the largest forest frontier in the region 9,13,14 is expected to have the remaining large patches of these invaluable forest types.
Myanmar lies in the Indo-Burma biodiversity hotspot 5 and has a wide range of forest types across the country, including Mangroves in Ayeyarwady delta, Evergreen forests in Tanintharyi, and Thorn forests in central Mandalay 15 .
Approximately, 70% of Myanmar's population reside in rural areas 16 and are dependent on forests for their livelihood 17,18 . Overexploitation of forest resources and conversion of forests to agriculture and urban development has resulted in high rates of deforestation in Myanmar 13,14,19 . Not surprisingly, in 2015, United Nations' Food and Agriculture Organization (UN-FAO) declared Myanmar as the country with third highest rate of deforestation in the world, after Brazil and Indonesia 20 . Although Myanmar has a fairly extensive network of protected areas it still has not achieved its declared goal of protecting 10% of the country 21 . It is challenging to assess if the existing network proportionally protect and represent Myanmar's diverse and critically important forest ecosystems, in part due to the lack of inventory information on forest type distribution. For example, there is only one relatively small, protected area that secures Dry Deciduous Diperocarp Forests (i.e., Chatthin Wildlife Sanctuary) and very few protected areas were designed to conserve critical Mangrove forests except Meinmahla Kyun Wildlife Sanctuary and Lampi Marine National Park. Only a few studies have attempted to assess the conservation status of speci c underrepresented forest types such as the Dry Deciduous Forest 9,22,23 or forests highly vulnerable to anthropogenic disturbances and climate change such as Mangroves 8 . At global and broader regional scales, very few detailed remote sensing maps exist that quantify and map the extent of different forest types. Such information is critically needed to inform forest management and biodiversity conservation.
Mapping forest types from satellite data is often confounded by technical di culties in separating high biomass vegetation types based on spectral characteristics alone [24][25][26] . This may be especially the case when the forest type is characterized by a complex species composition or when different forest types have dominant species with similar spectral properties. For example, the spectral signatures for natural tropical forests and mature rubber plantations at individual pixel level often are very similar 27,28 making it challenging to separate them based on only pixel-level remote sensing data at one point in time. Moreover, tropical forests tend to be highly heterogeneous in species composition, resulting in mixed spectral signals 26 . Mapping diverse forest types in the tropics would bene t from satellite images with ne spatial and spectral resolution due to vast information retrieved from the large number of spectral bands at high resolution 24 . However, while there have been a few attempts to map forest types at ner spatial scales, such as the Tanintharyi landscape in southern Myanmar 29 , country-wide mapping efforts are lacking. This is especially true for Myanmar where limited nancial and technical resources have hampered the development of a spatially explicit and reliable nationwide forest type map.
To advance sustainable forest management and conservation at the national level in Myanmar, it is necessary to develop country-wide, spatially explicit maps of different forest types. The current forest type map 30 used for broad reference by the Forest Department is more than 90 years old, is in coarse scale and outdated. Forest type classi cations in global land cover products like ESA Climate Change Initiative (CCI) Land Cover (LC) 31 provide some information about the national forest type, at a relatively coarse scale (e.g., ESA CCI LC product has 300m resolution).
The general forest types in the global (ESA CCI LC) and regional maps 32 also do not provide the detailed thematic classi cation required for effective forest inventory and conservation analyses at national level. The absence of recent and accurate forest type maps is a primary challenge for the Myanmar Forest Department 33 .
Recent advances in cloud computing technology such as Google Earth Engine (GEE) 34 , the availability of new, free, mid-resolution optical and radar satellite imagery like the Sentinels (Sentinel-2 and Sentinel-1), and machine learning methods provide an unprecedented opportunity to develop customized forest type maps suitable for national scale applications. GEE is a cloud-based platform which integrates access to large volumes of satellite imagery and ancillary datasets available from their data catalogue with geospatial analysis. It can process the results of any analysis with large data volume in a short time compared to traditional computers. Among the large volume of satellite images available through GEE, a combination of Sentinel-1 and 2 images has shown to improve the accuracy of forest type mapping in Myanmar compared to using Landsat-8 or Sentinel-2 alone 33 . To collect a large number of reference data on local forest types and their distributions, we collaborated with experts in Myanmar forests within local Non-Governmental Organizations (NGOs), Government and International Organizations. By combining state of art remote sensing technologies, satellite data, and extensive reference data, we developed the rst spatially explicit forest type map of Myanmar circa 2020, at 20m resolution.
We used this new map to determine the extent, distribution, and conservation status of the different forest types in Myanmar. We developed a Conservation Status Score, based on abundance, protection status, and representation of the forest type within protected areas, to identify the ve least conserved forest types and their largest remaining patches. The results from our study will be critical to inform biodiversity conservation planning and forest management in Myanmar and the broader Southeast Asian region. In addition, our methods could easily be transferred to expand forest type mapping into neighboring countries in the Mekong region.

Results
At the national level, we mapped 507,361 km 2 of forests (Table 1) including 475,403 km 2 of natural forest and 31,958 km 2 of plantation forest. Natural forests cover ~70% of Myanmar's total land area. The most abundant forest types are Upland Evergreen (38.17% of total forest area), Bamboo (22.65%), and Mixed Deciduous (17.86%) forests, which combined make up ~79% of the total forest area in Myanmar. All other forest types combined constitute less than a quarter (21%) of the total forest area. Many of these less common forest types (eg. Swamp and Thorn) are only found in small areas and are highly restricted in their geographic distribution. Forest types accounting for between 10% and 1% of the forested area (Table 1) include Lowland Evergreen with a total area of 33,552 km 2 (6.61%) followed by Dry Deciduous Forests including Indaing (30,  Among other classes, it is noteworthy to nd that Bamboo and Plantation occupy a large extent within protected areas (7,814 km 2 and 885 km 2 respectively).

Discussion
We developed the rst spatially explicit countrywide forest type map of Myanmar and mapped nine major forest types at 20m resolution. Our map represents a signi cant expansion beyond previous efforts in mapping forest types. These previous efforts a) were restricted on smaller areas, such as the Tanintharyi region 29 ; b) used single-class classi er to map a speci c forest types, such as Dry Deciduous forest 9,23 ; or c) mapped general land use and land cover with the inclusion of fewer and more general forest types [35][36][37] .
Our map provides a comprehensive, baseline estimate of major forest types within the country. This map may ll the existing knowledge gap created by the absence of national level forest type map. The estimates of forest type extent generated from our map may be used by agencies and stakeholders to: Identify vulnerable forest types for conservation and forest management purposes.
Monitor forest-type-speci c deforestation rate by using our estimate as a baseline extent.
Estimate and understand past deforestation trends by forest type using previous maps of deforestation.
Contribute to improving the estimates of emission factors and activity data for future Forest Reference Level Our map shows improved forest cover/type mapping in Central Myanmar. Previous studies tended to misclassify Dry Deciduous forests as other land use types 35,36 , or underestimate these forest types 9 . Our method of using region speci c tree cover thresholds facilitated improved mapping of the Dry Deciduous in Central Myanmar compared to existing products ( Figure 3). In our study we used a greater than 10% tree cover threshold identifying forested pixels in Central Myanmar. In addition, the 20m resolution of the Sentinel-2 imagery also helped us to include formerly excluded small fragmented Dry Deciduous forests thus improving our estimate of Dry Deciduous forest extent. To demonstrate the improvement of our methods over existing land cover products we selected a well-studied region of Dry Deciduous forests at Chatthin Wildlife Sanctuary (CWS) ( Forest. Both studies 35,36 fail to classify the dominant vegetation of CWS as forest and instead classify it as Shrubs/Grass or Woody, indicating that the vegetation is less than 5m tall which is not true according to ground conditions. Our forest type map can bene t both products by providing them with more accurate and detailed information on forest types. Historically, deforestation patterns across Myanmar are not uniform but concentrated in some regions and often on speci c forest types 9,13,14,23 . Different forest types have different economic and conservation values and are exposed to different levels of threat due to regional differences in the socio-economic, political factors driving deforestation 13,41 . For example, the Mixed Deciduous forests in Myanmar's Bago region used to be an important resource to support the hardwood industry in Myanmar, especially the globally renowned 'Burma Teak', but have been severely overexploited, resulting in signi cant forest degradation 42 . The dominant presence of Bamboo on the eastern side Bago Yoma, bears testimony to the degraded condition ( Figure 1) as the presence of Bamboo in Evergreen and Mixed Deciduous forest is a sign of forest degradation. For the same reason, the presence of Bamboo within protected areas (Table 1) is concerning. However, within Rakhine, Bamboo is considered a part of the natural ecosystem and is an integral part of the local re ecology 43,44 . The presence of Plantations in protected areas is another reason for concern as it is a sign of anthropogenic activity within protected areas leading to loss of habitat 45 .
Swamps, Mangroves, Dry Deciduous, Lowland Evergreen and Thorn forests were identi ed as the ve most vulnerable forest types in Myanmar (Table 1). All have low representation in Myanmar's protected area system. The largest patch of most of these forests lie mostly unprotected providing a last opportunity for conservation and improved representation of the vulnerable forest types within Myanmar's protected area system.
Swamp forests are the rarest and most vulnerable forest type in Myanmar. Very little is known about these forests though they provide essential ecosystem services and are critical in supporting local wetland biodiversity. They Mangroves are highly threatened at a global and regional scales 8 . In Asia, Mangroves have been historically felled for agricultural expansion, aquaculture development, and charcoal production 14,41,48   Thorn forest is a rare forest type in Myanmar covering only 0.54% (2721.86 km 2 ) of forest area. Much of the remaining Thorn forest only persisted because it is located in protected areas or zones that are of cultural importance -as is the case with the slopes of Mount Popa in central Myanmar. This explains why the remaining forest appears to be relatively well-protected (17.89 %). This number can be misleading without discussing the context that this forest type is extremely rare and has low representation in Myanmar's protected area network (1.15%). Although there have been localized efforts to map Thorn forests, to our knowledge our map provides a rst approximation of how much is left and where the remaining Thorn forests can be found. Efforts to increase the representation of Thorn forests may include protecting the largest patch of currently unprotected Thorn forest in Sagaing.
According to our national level accuracy assessment, Plantation was the least accurately mapped class (UA: 51.43%, PA: 45%). In Myanmar, the type of plantations is different in the north and south. The plantations in the north are generally more challenging to separate because of the relatively small scale and diverse types of agroforestry 57 that are present. Because that our eld teams were unable to collect large number of plantation points in Kachin and Shan States due to active armed con icts, the accuracy of plantation in the northern part of the country was lower than the national average. We caution the use of plantation maps alone in northern Myanmar as it is often underestimated and confused with Lowland Evergreen forests.
In the south, however, industrial scale oil palm and rubber plantations are widespread and account for majority of the plantation land cover type. They are relatively easy to identify based on their texture ( Figure 6a) and are easily separated from the surrounding forest types found such as Mangroves, Lowland and Upland Evergreen. We estimated the user's accuracy of plantation maps in the south, particularly in Tanintharyi region to gauge a regional classi cation accuracy in the best case scenario. We generated 150 random points within the Plantation layer in b) Using high-resolution images (<5m) from commercial small satellites like Planetscope, RapidEye, Qucikbird, Worldview etc. will help to improve the detection of the low tree cover and small fragmented forest types, like Dry Deciduous forests and Thorn forests due to more spatial details captured. Using high-resolution images (<5m) may also help to map a few other, relatively rare, and localized forest types. These include pine, oak, and rhododendron forests which occur at higher elevations, often along ridge tops and which are no included in our Upland Evergreen forest category. these forests are of high importance for biodiversity conservation and there should be increased efforts at mapping and validating maps for these forest types.
ii) Open-source ground collected data-We collected extensive ground truth data during this project with the help of our own eld team in Myanmar and through local (Myanmar Forest Department) and international (UN-FAO) collaborations. Despite our sincere efforts, some parts of the country were inaccessible due to con ict. Increasing the variation and volume of training/validation data is expected to improve the forest type classi cation. Creating openaccess data sharing mechanisms for ground collected forest type data across local and international will be key to facilitating development of more accurate forest type products not only in Myanmar but around the globe.
iii) More research into the lesser-known ecosystems such as Swamp and Thorn forests are needed.

Conclusion
In this study we developed Myanmar's rst spatially explicit national level forest type map using multi-sensor satellite imagery, ancillary datasets, extensive training data and machine learning methods. Our map provides a current, accurate status of forest type extent, distribution, and protection status. We also determined the conservation and protection status of all existing forest types and identi ed the ve most vulnerable forest types. Our maps will be able to inform future forest conservation policies, especially those related to greenhouse gas emissions, biodiversity conservation and sustainable forest management.

Study Area
Myanmar lies at the junction of South and Southeast Asia between 9°32'N to 28°31'N and 92°10'E and 101°11'E ( Figure 4).

Satellite Data Sources
Sentinel-2 collects multispectral satellite images consisting of 13 bands and was chosen because of its higher resolution (20m vs 30m) and greater number of spectral bands (13 vs 8) compared to Landsat-8. In our previous analysis of forest types for a selected areas of Myanmar using Sentinel-2 images signi cantly improved map accuracy over using Landsat-8 images due to ner spatial resolution and the unique contributions of the vegetation red edge bands 33 .
The Phased Array L-band Synthetic Aperture Radar (PALSAR) is a L-band radar system widely used in mapping forests 59 . The L band allows for better penetration of the forest canopy compared to C-band Sentinel-1 and consequently provides new and additional information on forest structure 60 . We used the 25m global PALSAR/PALSAR-2 mosaic developed by 59 .
Sentinel-1 collected C-band radar data. We used the VV and VH bands which have 10m resolution. Sentinel-1 radar data had ner resolution than PALSAR (10m vs 25m), and the addition of Sentine-1 C-band also brought signi cant accuracy improvement in forest type maps for central Myanmar 33 . We used both, L-band PALSAR and C-band Sentinel-1 images to maximize the structural information retrieved from the ground.

Ancillary Datasets
The Myanmar percent tree cover map of the year 2018 61 available at https://smithsonian. gshare.com/articles/dataset/Myanmar_percent_tree_cover_map_of_the_year_2018/12772490 provides information on percent tree cover value at each 30m Landsat pixel. The percent tree cover values range from 0-100%. The dataset is derived by calibrating tree-cover estimates from Landsat Vegetation Continuous Fields (VCF) tree cover layer against high-resolution estimates derived from drone imagery and other sources. The product was developed by a collaboration between Smithsonian Conservation Biology Institute and terraPulse. We used this percent tree cover map to develop a non-forest mask, allowing us to focus our mapping on forested pixels only.
The Global Forest Canopy Height dataset 62 is available at https://glad.umd.edu/dataset/gedi and provides an estimate of the forest canopy height at 30m resolution. It was developed by integrating lidar-based forest structural metrics from the Global Ecosystem Dynamics Investigation (GEDI) sensor and surface phenology information from Landsat time-series 62 . The variable forest canopy height was expected to be a predictor forest types by differentiating between tall (Evergreen) and short (Thorn) forests.
Previous research has shown that elevation, slope and aspect are a strong predictor of forest types 63 . To better assess and predict different forest types we relied on elevation, slope and aspect data which we obtained from 90m It was uploaded into GEE as a private asset. All images and datasets were accessed through GEE.

Forest Type Classi cation and Field Data Collection
We developed a countrywide forest type classi cation suitable for satellite-based forest type mapping for Myanmar (Table S1). This classi cation system was based on existing literature on Burmese forests and plants 65 , forest types To verify our forest type classi cation in the eld and to collect training data, we conducted eld work at 12 locations across Myanmar during the dry season in late 2018 and early 2019 ( Figure 4). Field work was conducted by three teams, consisting of three members each. Each team member had a background in forestry and/or botany and was well versed in identifying the different forest types found in Myanmar. The teams recorded latitude, longitude, elevation, forest type and photos of location acquired for 10 directions (North, North-East, East, South-East, South, South-West, West, North-West, Above, and Ground). The entries were recorded in customized forms in the Collect Mobile app (http://www.openforis.org/newwebsite/tools/collect-mobile.html) available in the Open Foris platform.
The collected data included 370 data points representing different forest types.
To validate our forest type map, we compiled a validation dataset by combining ground points collected by our team (127), with the ones collected by FAO (1329) (Figure 4). The validation dataset is independent of the training dataset and consisted of a total of 1456 ground points. Each of the ground points collected by our team had a speci c forest type attributed to it. The ground points collected by the FAO followed a hierarchical forest type classi cation. The hierarchical FAO forest types were reclassi ed to match our forest type de nition (

Methods
We developed an open-source method to map the broad forest types in Myanmar using freely available satellite images and machine learning algorithm in GEE. Forest type maps were developed for each level 1 administrative unit (State/Region/ Union Territory) and then mosaicked to form the national forest type map. The national forest type map developed is representative of the year 2020. The overall work ow ( Figure 5) of the forest type map development consists of the following steps:  (Table S2). Band ratios, Normalized Difference of bands, and texture metrics 66 were also calculated for each band in every monthly composite. Texture metrics were computed using a 3 X 3 window.
The PALSAR and Sentinel-1 images were converted to decibels and a Re ned Lee Filter was run to remove speckle in the images 67 . Since the PALSAR dataset was an annual dataset no monthly compositing was done. We used the HH and HV band of the PALSAR data. The Sentinel-1 images were available for the entire year, so it was split into monthly composites. We used the 10m VV and VH bands of the Sentinel-1 images. Texture metrics were computed from each radar band of both radar satellites in a 3 X 3 window.
Finally, all the satellite datasets were resampled and stacked to create a raster stack with 20m resolution.

Masking Non-Interest Areas
To focus our image classi cation on forest areas within administrative boundaries, we used masks to exclude areas that were not of interest. We developed three masks, a non-forest mask, a water mask, and an area mask to ensure that we select only the forested pixels within the level 1 administrative boundaries.
For the non-forest mask, we utilized two data sources, For the water mask we selected pixels which had NDVI less than 0 in any month of the dry season. These two masks combined allowed us to remove pixels with low biomass vegetation cover with clear trend in seasonality which may be confused for forests in the Myanmar percent tree cover map of the year 2018.
The third mask, an area mask, masked all pixels outside the level 1 administrative boundary.

Training Data Creation
Training data was collected for each distinct forest type by delineating polygons for homogeneous forest pixels based on eld data and overlaying high resolution images available on GEE, and Google Earth. Training polygons for each forest type were identi ed in high resolution and satellite images based on their characteristic texture and seasonality ( Figure 6a and 6b). For example, Bamboo has a characteristic thin star shaped canopy, mature oil palms have a thicker star shaped canopy, young oil palm, rubber, tea, coffee, or any other bush/shrub plantations are planted in rows leading to typical vegetation texture, evergreen forests have green cover and constantly high NDVI throughout the year, whereas the mixed and dry deciduous forests as well as thorns have progressively higher difference in NDVI between dry and wet seasons. Mangroves were found around coastal areas and near rivers and mostly inundated.
Special care was taken to include training polygons around the boundaries of each study area to ensure a smooth blending of forest types when the units are mosaiced at country level.

Model Parameterization and Selection of Important Variables
We used a random forest 68 algorithm to develop the administrative level forest type map. The random forest algorithm was run with the number of trees equal to 500 and the number of variables/split set to square root of the number of input variables.
The random forest algorithm was run twice. In the rst run we selected 30 pixels from each forest type class and used all the bands and metrics computed from satellite datasets to select the top 20-25 bands which contributed to our forest type classi cation based on their importance value, calculated from the random forest function.

Map Development
In the second run, we used the selected bands and all the training data created to develop the forest type map. Initially, a forest type map was created for each of the 15 Level 1 administrative units (States/Regions/Union Territory) and then the 15 units were mosaiced together to derive a countrywide map. This approach enabled us to adjust the tree cover threshold regionally to include low tree cover forested pixels in Central Dry Zone which would otherwise be missed if a single countrywide tree cover threshold is used to de ne forested pixels.

Accuracy Assessment
A national level accuracy assessment was conducted to calculate the user's, producer's, and overall accuracy.
Protected Area Dataset and Analysis.
To assess the conservation status for different forest types, we used a shape le for the latest existing protected areas in Myanmar 69 . The dataset included 43 existing protected areas. We considered three factors to determine the conservation status: 1. Abundance determined as the percent area of all forest areas covered by a forest type; 2. Protection Status, calculated as the percent of each forest types that is protected; 3. Conservation Representation, computed as the percent of all protected areas covered by this forest types; Finally, the Conservation Status Score for each forest type is calculated as the sum of the previous indices: a+b+c, representing the comprehensive degree of vulnerability. Since it is a sum of three indices in percentage, its value ranges from 0-300 and its unit is percentage.
Forests with low values for the Conservation Status Ranking are rare, less protected, and not well represented in Myanmar's protected area system, making them more vulnerable than forests with higher Conservation Status Ranking. To determine the ve most vulnerable forest types, we ranked the Conservation Status Score of all forest types in ascending order and identi ed the rst ve forest types.

Identifying the Largest Remaining Patches by Forest Type
We were also interested in providing maps of the largest remaining forest patch for the ve most vulnerable forest type. Such maps may provide guidance for the development of new protected areas. To identify the largest remaining patches of forest, we clumped connecting forested pixels together in ArcGIS 70 for each of three forest types using the regiongroup algorithm. Based on these clusters we then identi ed the largest remaining patch. For each of the vulnerable forest types identi ed, we determined the location, extent, and protection status of the largest remaining patch.

Data Availability
The forest type map developed and associated metadata are available on Figshare at https://smithsonian. gshare.com/articles/ gure/Myanmar_Forest_Type_Map_2020/16613818    Diagram showing the outline of the adopted methodology. Figure 6 a. Characteristic texture used to identify the different forest types and plantations.
b. Phenological information from Sentinel-2 used to identify the seasonal forests. False Color Composites for Dry (April) and Wet (November) season are displayed in band combinations B11, B8, B3 in Red, Green and Blue channels.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. SupplementaryInformationSR.docx