Trees outside forests are an underestimated resource in a country with low forest cover

Trees outside forests (TOF) are an underrepresented resource in forest poor nations. As a result of their frequent omission from national forest resource assessments and a lack of readily available very-high-resolution remotely sensed imagery, TOF status and characterization has until now, been unknown. Here, we assess the capacity of openly available 10 m ESA Sentinel constellation satellite imagery for mapping TOF extent at the national level in Bangladesh. In addition, we estimate canopy height for TOF using a TanDEM-X DEM. We map 2,233,578 ha of TOF in Bangladesh with a mean canopy height of 7.3 m. We map 31 and 53% more TOF than existing estimates of TOF and forest, respectively. We find TOF in Bangladesh is nationally fragmented as a consequence of agricultural activity, yet is capable of maintaining connectedness between remaining stands. Now, TOF accounting is feasible at the national scale using readily available datasets, enabling the mainstream inclusion of TOF in national forest resource assessments for other countries.

Trees outside forests (TOF) are an overlooked and undervalued resource that are not regularly included in national inventories. Subsequently, data on TOF extent and structure is often scarce, despite their increasing importance in a continually urbanizing world. The continued rate of urbanization worldwide is driving an increase in concern for the separation of people from the natural environment 1 and TOF will be heavily relied upon as surrogates for natural forests and the services that they provide, albeit to a more limited capacity than large intact forests 2 . This is pertinent where TOF may provide the only wood resources to local populations 3,4 , as in developing nations where two-thirds of fuel wood is derived from TOF 5 . Furthermore, at a national level TOF are capable of contributing significantly to biomass stocks 6 but are overlooked in rural or agricultural areas despite 40% of all global agricultural land containing a tree cover of more than 10% 7 . This has been observed in South Asia whereby production from TOF has not been fully realized, despite the potential of TOF at the national level to decrease the gap between timber demand and supply 8 . Including TOF in resource accounting is therefore paramount to achieving accurate data on wood volume stocks, and ensuring the provision of ecosystem services to rural populations.
The lack of inclusion of TOF in forest assessments has led to little knowledge of TOF across large areas, especially at the national reporting level. The United Nations Food and Agricultural Organization (UN FAO) introduced TOF into their Global Forest Resource Assessment (FRA) of 2000 9 . The FAO definition of TOF are trees that occur outside of the FAO definitions of forest and other wooded land (OWL) 10 , yet TOF do not definitively adhere to limitations on extent nor crown cover 11 . Subsequently in 2005 the FAO added a category of Other Land with Tree Cover (OLwTC) to the RFA to account for forests that meet the FAO forest definitions but are situated within a predominantly agricultural or urban setting 12 . Here we adopt this definition of TOF. This category was easier to implement in practice and enabled all tree resources to be included in national forest inventories, thus providing a pathway for the inclusion of TOF. This was bolstered by a range of UN initiatives that required the inclusion of TOF to meet adequate levels of reporting (e.g., UNFCCC) 13 . In addition, inventories conducted under the National Forest Monitoring and Assessment (NFMA) program of the UNFAO include all trees irrespective of land use 14 and this approach has also been adopted by a limited number of national forest inventories that include TOF in their reporting 15 . However, the monitoring of TOF is highly variable, both in terms of the level of inclusion of TOF and the quality of reporting. Bias exists in the inclusion of trees in one domain and not in others, such as in agricultural land but not in urban environments, further exacerbated by

Results
Tree extent. TOF accounts for an additional 31% more tree area than existing estimates of TOF 22 in Bangladesh. A total of 2,233,578 ha of TOF was mapped across the country (Fig. 1) accounting for 15.1% of the Bangladesh surface area. The area of TOF mapped is 53% larger than the existing mapped forest area of 1,464,000 ha 22 , highlighting the prominence of TOF in comparison to forest area in Bangladesh. The largest proportion of TOF was in the division of Dhaka (415,582 ha) with an almost equitable area of TOF in the division of Chittagong (399,331 ha). The area of TOF mapped per division is given in Table 1. Barisal contained the largest proportional TOF area (22.8%) while Sylhet contained the lowest proportional forest area (11.4%). TOF was mapped at the national scale with an accuracy of 91.5% (Supplementary Table S1), ranging from 89.3% in Dhaka to 93.7% Rangpur.
TOF within Bangladesh were characterized by numerous types of stands, varying in size and density across the full extent of the country (Fig. 1). In agricultural regions (Fig. 1C), the TOF were located along agricultural field boundaries and among nearby local settlements. Here the TOF occurred in narrow lines which connected www.nature.com/scientificreports/ at nodes to other stands of TOF. All settlements contained trees, particularly rural settlements with lesser tree extent mapped in large urban cities. Figure 1A demonstrates these small homestead trees that occur as isolated stands in geometric patterns that surround individual dwellings. These are not interconnected but occur as small single stands, separated by agricultural fields. Rivers were also commonly tree lined often accounting for the largest stands of TOF. Here the TOF were mapped as narrow stands, but as long tributaries connected to other similar TOF patches, forming large continuous connected extents at the landscape level. These had a larger width in relation to their length, defined as the largest axis of the TOF patch, than other stands of TOF. This is visible in Fig. 1B whereby the TOF occurs as narrow stands but in less geometric patterns across the landscape. The majority (65.6%) of the TOF stands were below 1 ha in size, accounting for 9% of TOF area, with an additional one quarter (25.6%) below 5 ha (15.4% of TOF area). Almost all of the stands patches (99.7%) were below 100 ha in size, accounting for 62.3% of TOF area. TOF stands over 1000 ha (0.3% of stands) contained 37.7% of TOF area (Table 2). TOF is heavily fragmented in terms of the number of stands, but maintains connectedness as a small number of the patches constitute a large proportion of the TOF area. This connectedness is reflected in Fig. 1B and C where the stands remain connected at the landscape level despite being fragmented at the local level. These stands, while covering an area large enough to qualify as forest, remain classified as TOF due their setting among agricultural and urban land uses. Furthermore, these stands are connected by as little as a single pixel, thus do not necessarily have a continuous canopy. Examples of these patch sizes are provided in supplementary Figure S1.   32 when converted from canopy cover to extent (see methods). The largest difference between the maps occurred in Dhaka where 286,862 ha more were mapped at the 10 m scale. The smallest discrepancy was 38,939 ha in Barisal which exemplifies the large differences in tree extent mapped at these two different scales. The TOF map presented is of all vegetation of a height greater than 2.19 m, thus we applied a 5m height threshold as used by the global forest cover dataset 32 . Using this threshold we map a greater area of TOF in all divisions, with the exception of Sylhet. The area per division in each map is given in Table 3, alongside the area of TOF greater than 5 m. Similarly, the TOF map was compared with the 50 m German Aerospace Agency (DLR) Forest/Non-Forest (FNF) map 33 which mapped a greater quantity of forest in five of the eight divisions. The largest difference occurred in Khulna where 123,028 ha more forest was mapped in the FNF dataset, although small differences were mapped in Rajshai and Sylhet where the FNF quantified 20,956 ha and 29,380 ha less, respectively. The differences between these datasets is given in Fig. 2, which shows the discrepancies and similarities between the three maps. Our TOF map resolves stands in the FNF map in greater detail and includes TOF largely omitted from the global tree cover map. Our estimate of TOF extent is 532,178 ha (31.3%) greater than the previous remotely sensed estimate of TOF forest cover area by Potapov 22 , although restrictions on the distribution of this dataset prevent a more thorough quantitative analysis.  Table 4. The Chittagong contained the tallest maximum canopy height stand at 20.5 m while the shortest maximum canopy height tree stand was in Rajshahi (15.2 m). The distributions of the tree heights by division is given in (Fig. 3). There was no relationship between height and spatial distribution at the national scale with a range of heights observed across the landscape, although localized relationships were more evident. This was more prominent at the stand level, whereby larger stands contained taller trees, with smaller trees at the outer edge of the stand, increasing in height towards the stand center (Fig. 4C). TOF that were isolated within homestead settings contained shorter trees (Fig. 4A) while TOF that were more continuous (Fig. 4B) contained taller trees. Canopy heights varied across the landscape with adjacent pixels often displaying a contrast in heights, indicative of heterogeneous canopies. The average tree height increased when only trees above 5 m were accounted for, as expected when the large number of small trees were removed. These average tree canopy heights above 5 m are also provided in Table 4. www.nature.com/scientificreports/ and TOF map (Bottom). Spatial resolution drives the differences between the maps, with most detail captured by the TOF map. The TOF and DLR FNF map 33 are most similar, perhaps due to the use of small wavelength radar while the global forest cover map does not capture small forest stands. The differences and similarities between the maps reflect the modes of data used and the metric that was mapped (cover or extent). Maps made with QGIS 2.8 (https:// www. qgis. org/ en/ site/ index. html).

Discussion
We provide among the first country scale canopy height maps 34 and the first TOF canopy height map for Bangladesh, a country with a highly fragmented and distributed wood resource. At 10 m spatial resolution, we provide the most recent and detailed map of TOF extent for Bangladesh alongside associated canopy height at a resolution of 30 m.   www.nature.com/scientificreports/ Our map of TOF increases previous estimates by 328,541 ha (31.3%) over that of Potapov 22 , 816,941 (67.3%) over that of the FAO FRA 35 and 1,281,034 ha (57.4%) over the global forest cover map 32 . The DLR FNF 33 estimate was 10% larger than our mapped extent, which was the smallest discrepancy from our TOF map. This small difference may reflect the use of radar imagery, which is sensitive to tree structure, used in both methods and the coarser pixel resolution whereby an increase in backscatter from a small stand could be the dominant scatterer even at 50 m resolution. There was a larger difference with the other approaches that used optical remotely sensed imagery alone, but these estimates were closer to our TOF extent than the FRA estimate. While different methods and definitions of forest were used in the creation of each map and should be considered when comparing these datasets, this also highlights that although definitions are important for broadly classifying forests, they are a dynamic and continuous factor. The consequence of this is a risk of excluding large quantities of trees, particularly where TOF is a large proportion of a nations wood resource.
The influence of anthropogenic activity upon the forests of Bangladesh and the rise of agriculture which replaced it 36 was reflected in the area of TOF within the eight divisions of Bangladesh. The large area of TOF in each division reflects the widespread loss of forest and conversion to agriculture in Bangladesh, particularly outside of elevated regions, leaving large quantities of TOF remaining. The total area of TOF was equitable to 0.014 ha per person, in contrast to the US where the FAO definition of Other Land with Tree Cover (OLwTC) equates to 0.08 ha of forest per person. This requires TOF in Bangladesh to provide ecosystem services to a larger number of people from a smaller area covered by trees. As TOF were the only trees in some divisions, this provides the only resource available to local populations. These TOF were observed to consist of many types which the Bangladesh government has promoted 20 through agroforesty, homegardens, farm forestry, social forestry and urban forestry, among others 37 . Some of these are identifiable in the tree extent map, particularly those of the homegardens and the urban forestry. These TOF are critical as they are able to serve as proxies for forest, allowing the rural and landless poor to procure subsistence products such as fuel wood, fodder and timber. In this way, TOF contribute substantially to the rural economy through the provision of products which reduces the reliance on purchasing power, while also offering employment 38,39 . In fact, economic drivers have been primarily responsible for the documented increase of TOF in Bangladesh although landowners are also able to appreciate the secondary environmental benefits that TOF provide, which can often be entwined with economic factors 40 . The variation in the stands that TOF form reflects the range of ways in which TOF is utilized and subsequently manifests as a heterogeneous TOF type at the national level.
There was a large variation in TOF stand size, which was a function of the relationship between the size of the stands and their connectedness. Over 90% of the stands were <5 ha in size, accounting for 24.4% of total TOF, reflecting the disparate nature of TOF, promoted by homestead and social forestry that are not well connected and occur as isolated patches. However, the majority of the TOF area (76%) occurred in stands greater than 5 ha when stands joined by at least one pixel were mapped, which suggests a connectedness between remaining TOF despite heavy anthropogenic management. Their location in agricultural and urban settings qualify these stands as TOF but this connectedness allows ecological continuities between forest patches, over large fragmented stands at the landscape and national level 41 . Bangladesh TOF is host to significant species diversity, particularly in insect, avian and tree species despite the amount of area it occupies 42 , thus the connectedness of TOF within Bangladesh serves an important ecological function despite accounting for 15.1% of the land area. Additional losses would have a disproportionately negative impact as the remaining TOF serves an ever important ecosystem function. As the area demarcated for natural forest in Bangladesh is low and almost all limited to specific locations, TOF is a keystone in ensuring ecological richness and diversity 40 across the country. The ability of TOF in Bangladesh to meet these demands, is not guaranteed into the future without the continuation of the commitment of local level parties and grassroots movements to promote participatory social forestry and increase TOF 16,18 .
The disparity between our TOF map and other mapped extents is a consequence of two controlling factors. The first is definitions of forest and TOF which can be susceptible to interpretation and thus affect the categorization of tree type. We interpret the majority of trees within the Bangladesh deltaic plain as TOF due to the heavy presence of agricultural and urban land uses, which have fragmented the remaining woody resource. We therefore align our definition of TOF with that of the FAO 12 definition of Other Land with Tree Cover. In contrast, the Potatpov 22 map uses a combination of tree cover (>50%) and user assigned labels within a sampling approach to classify forest type. The Hansen 32 map used in our comparison includes all pixels with >0% canopy cover and the DLR FNF map is a binary tree map with no categorization into forest type 33 . Each map is justified in its choice of definition but highlights that maps are not reliably comparable due to an lack of consistent definitions and the comparison of forest with non-forest pixels. Secondly, we use a high spatial resolution of 10 m, increasing the resolution over other remotely sensed methods 33 by as much as 25 times. This enables the detection of small stands that are below that of coarser resolution imagery, but which are common in TOF. This combined with differences in definitions drive discrepancies between mapped products and while we report different metrics than existing maps, we do not claim a holistic superiority over them.
The average tree height varies by 1.7 m between divisions although the distribution of the TOF heights exhibit marked differences. Maximum tree heights differed by 5.3 m and the divisions of Barisal, Dhaka and Sylhet had an increased quantity of small trees compared to the other divisions. This variation may reflect the different ways in which TOF are used across the country and the difference between rural, urbanized and agricultural regions, whereby tree stands may be affected by use intensity or allowed to grow in order to provide other services such as shade 43 . TOF have been observed to be used as wind breaks and can be grown in mixed stands which form a semi-permenable barrier over a range of heights. This may drive the variation in tree heights as part of efforts to reduce soil erosion and desertification in more arid regions 37 . Average tree height increased within each division when a 5 m minimum threshold was used, as is stated within FAO FRA guidelines for forest. This was accompanied by a reduction in TOF area, suggesting that much of the TOF is below 5 m in stature. This reflects the heavy dependence of the population on their TOF resource, as shorter woody vegetation emerge in the www.nature.com/scientificreports/ absence of forestry. The 5 m threshold aligns this work with previous estimates 35 but results in the underestimation of trees which may be responsible for the large discrepancies between the TOF map and existing datasets. Therefore, a 5 m threshold may not be suitable for accounting for woody resources which can be mapped with high-resolution (10 m) imagery and should be reconsidered with the emergence of high resolution extent mapping. The temporal difference between our tree map (2018) and elevation data (2011-2013) is a potential source of error, as tree extent undoubtedly changed during this period. However, between 2000-2014 an overall tree gain of 4.3% was observed across Bangladesh 22 , thus we do not expect this to be a substantial source of error. Furthermore, the use of averaged tree heights in 1 ha plots, as outlined in the methods, was designed to reduce the impact of anomalous values on our tree height map. To derive a tree height map, a bare earth surface model was generated to subtract from the TDX DEM. We provide a labor intensive but accurate method of creating bare earth elevation models in regions with modest changes in elevation and slope. This method is applicable to other regions, particularly coastal zones and agricultural areas which are often characterized by small changes in elevation and sufficient locations from which to choose ground control points (GCPs). This provides a means of generating digital surface models (DSMs) from openly available elevation data such as SRTM, for vegetation height mapping across broad geographical scales. Given the often disparate distribution of TOF, maps of extent and canopy height can provide information for resource accounting at the national scale once combined with field collected data. Field surveys are time intensive and expensive, therefore remote sensing provides a means of directly measuring extent and structure across landscapes. Our maps are directly applicable for use in tandem with a Forest Resource Assessment (FRA) or National Forest Inventory (NFI) by capturing the spatially explicit distribution of TOF and reducing the uncertainty in these smaller reporting units where high fragmentation and inaccessibility can provide barriers to comprehensive reporting 2 . This would increase the capability for resource monitoring, particularly where TOF are the only wood resource. Prior to this, there was no knowledge on the structure of TOF which extends across almost the entirety of the Bangladesh. Knowledge on the distribution of height will be pertinent to maximizing efforts to manage this living biomass and the subsequent distribution of the ecosystem services that it can provide. This has important environmental and economic implications in terms of carbon storage and timber, respectively, and provides additional uses at the local level [37][38][39]43 . We envisage that our height map may be utilized as an input into national biomass accounting, when combined with additional field data, which is relevant to carbon accrediting schemes such as REDD+. Given the large proportion of tree area composed of TOF, this equates to a large percentage of the national carbon budget within living biomass, yet the proportion of carbon fluxes from changes in TOF extent are not yet known. This may overlook the importance of the inclusion of TOF and sparse forest types in national accounting and emissions estimates, even when they compose significant proportions of aboveground biomass. Our maps can help ensure this and with the recent availability of tree height estimates from spaceborne LiDAR missions (e.g. GEDI, ICESat-2), we anticipate that TOF canopy height will become commonplace when combined with accurate estimates of tree extent.
Given the fragmented distribution of TOF, spaceborne sensors with increased spatial resolution will be instrumental in resource monitoring of TOF. The 10 m pixel resolution facilitated the mapping of small stands of trees, over openly available data traditionally used for large scale land cover mapping (e.g., Landsat) 32 . The Sentinel-1 imagery was an ideal companion to Sentinel-2 as the combination of the radar imagery, indicative of structure, with the optical imagery for estimating land cover "greenness", provided a novel partnership for land cover mapping that has not previously been witnessed as part of the same satellite constellation. The availability of sufficiently high-resolution high-cadence imagery, particularly when accessible as analysis-ready-data in cloud computing environments (e.g., Google Earth Engine), position them as a solution to prior limitations of routine TOF mapping 2 .
TOF accounts for 31% and 52% more TOF and forest, respectively, than presently estimated 22 in Bangladesh, accounting for a significant proportion of the woody resource. This discepancy is driven largely by differences in the definition of TOF and the higher resolution dataset used in this study. We demonstrate that this information can now be readily supplied for resource accounting through the use of openly available high-resolution spaceborne imagery and easily accessible cloud computing environments. Furthermore, we demonstrate a method for achieving canopy height estimates from existing elevation data, which can be used in combination with TOF maps to provide comprehensive data on TOF extent and structure to inform national forest inventories/resource assessments at the national scale. We highlight that in countries such as Bangladesh, a focus on forest definitions detracts from achieving the most informative products for resource accounting, particularly for those that require management at fine scales in order to most effectively benefit populations at the local level. Each is a C-band Synthetic Aperture Radar (SAR) collecting dual polarization images of Earth's surface, with a repeat orbit every 6 days at a spatial resolution (pixel size) of 10 m. Radar instruments, such as those aboard Sentinel-1, are not inhibited by cloud cover nor atmospheric conditions, enabling data to be collected consistently over tropical regions during both day and night. Radar instruments are active sensors and transmit a pulse of energy towards Earth's surface and measure the response, termed backscatter. The quantity of backscatter returned to the sensor is a function of the target properties such as shape, size and water content and a function of the transmitted wavelength. Large, rough and sensor facing targets return a greater quantity of backscatter than small or smooth targets or those oriented away from the sensor. Sentinel-1 imagery is available within Google's Earth Engine environment where it is radiometrically calibrated, terrain corrected and in log scale (dB) units.

Sentinel-2.
Sentinel-2 is a satellite constellation composed of Sentinel-2A and Sentinel-2B. Each carries a passive multispectral instrument with 13 visible, near-infrared and shortwave infrared bands with spatial resolutions of 10, 20 and 60 m. Sentinel-2 acquires images over the same location on Earth's surface every 5 days in high latitudes, with a field of view of 290 km. The range of multispectral bands measure the quantity of light that is absorbed or reflected from different land cover types. Sentinel-2 imagery can therefore relay information on the water content, photosynthetic cover and bio-physical properties of land cover types. As with Sentinel-1, Sentinel-2 imagery is available within Google's Earth Engine environment.
TanDEM-X DEM. TerraSAR-X is an Earth observing x-band SAR instrument operated by the German Aerospcace Agency (DLR), launched in 2007. Since 2010 TerraSAR-X was accompanied by an almost identical second satellite. The two satellites fly in a helix formation, enabling the generation of interferograms with no temporal decorrelation and adverse impacts from atmospheric disturbances. The interferogram enables a digital elevation model (DEM) to be generated with a spatial resolution of 12 m and a relative vertical accuracy of 2 m. TanDEM-X is available from DLR through a science proposal.
Mapping tree extent. The mapping of tree extent was carried out in Google's Earth Engine environment, allowing instant access to imagery and algorithms to create cloud free optical composites and radiometrically calibrated and terrain corrected radar mosaics. To minimize cloud cover over Bangladesh, Sentinel-2 optical scenes were acquired during the dry season between February 1st 2018 and March 31st 2018, utilizing 824 Sentinel-2 10,000 km 2 tiles. For each image in the collection, the NDVI was calculated. This was used within the Earth Engine Quality Mosaic function to create a 'greenest' image composite, whereby a single image is generated using the pixel location of the most photosynthetic pixel from each image in the collection. From this www.nature.com/scientificreports/ composite, a single NDVI image was generated to determine photosynthetic pixels whilst avoiding differences in atmospheric conditions and subsequent reflectance. Thresholds on NDVI were used to segregate photosynthetic from non-photosynthetic land cover using manually defined thresholds (0.3< NDVI<0.9). The result of this was a binary map of photosynthetic land cover. Thresholds were chosen with constant reference to the imagery and underlying VHR Google Earth imagery which is readily achieved using the GEE interface. Sentinel-1 imagery was acquired for the whole of 2018, from January 1st to December 31st, totaling 771 images from both ascending and descending passes. Although the radar is not inhibited by illumination nor atmospheric conditions the data can be degraded by image speckle. To reduce this the imagery was averaged over a large temporal range to improve the quality of the imagery, based upon the assumption that trees in Bangladesh are a static land cover type relative to other land cover types (e.g., agriculture) and the backscatter will remain consistent. Both ascending and descending scenes were acquired forming two collections. Both the VH and VV polarized bands were selected from interferometric wide (IW) swath mode imagery. Separate ascending and descending images were generated by averaging the images in each collection. Prior to averaging the imagery was converted to Sigma Nought from the Earth Engine provided decibel (dB) units. Scan lines of very low backscatter between swaths in the average descending image were filled with values from the average ascending image. Each polarization was converted to decibel (dB) yielding two backscatter images. Manually defined VH backscatter thresholds of −9 to −14 dB, were used to segregate land cover with structure from those without. The VH polarization is indicative of volumetric scattering, such as that from a tree canopy, which denotes structure. The VV polarization was used to reduce false positives by removing pixels that were outside of the backscatter thresholds from −2 to −8 dB. Thresholds were chosen with constant reference to the radar imagery and underlying VHR Google Earth imagery within the GEE interface.
The intersection of the radar and optical binary maps combined pixels identified as photosynthetic with pixels identified as having structure, synonymous with trees. The use of decision thresholds were preferred over supervised classification algorithms. As atmospherically corrected data were not available, only normalized optical image indices in combination with two radar bands could be used, which limited the data layers available. Also, given the large geographical area, a large but unknown number of classes would need to be trained, increasing potential misclassification. Our approach therefore is simplified yet robust based upon our accuracy assessment and is in line with other radar derived remote sensing derived forest maps 46 . Adjoining pixels were clumped into objects and those composed of less than 25 pixels (0.25 ha) were omitted, in order to reduce the detection of small false positives and noise. The combination of these two sensors yielded a high resolution country wide map of all trees in Bangladesh with a combined crown cover greater than 0.25 ha. TOF were separated from forests using a manually drawn mask with reference to the land use in very high resolution Google Earth imagery. These forested regions were characterized by their setting outside of agricultural/urban settings and with large continuous canopy cover. The land area omitted is visible in supplementary Figures S2 and S5. The thresholds chosen were considered to be conservative, in order to reduce the false classification of trees in the radar imagery. This reduced the noise in our data but also led to the underclassification of some pixels. This was most common among more continuous stands of TOF whereby heterogeneous surfaces caused destructive interference and the reduction of backscatter in radar imagery. Plantations were not readily separable from other forests in the imagery and were considered a form of TOF. Patch size was determined by clumping TOF pixels that were connected by at least one pixel edge and calculating the area of each.
Digital terrain and canopy height models. TanDEM-X (TDX) elevation data was acquired 2011-2013 and covers almost the entirety of Bangladesh. In order to attain the extent required, the highest resolution (12 m) data was supplemented with coarser resolution (30 m) data to maintain wall to wall consistency. The TDX measures surface height so it was not possible to directly measure canopy height from the ground surface as the terrain elevation was unknown. Areas of bare ground were sampled with 8200 points, using the TDX and Google Earth imagery as reference. A Geospatial Data Abstraction Library (GDAL 47 ) inverse power to distance interpolation was used to estimate the ground elevation and a sample of 400 independent points over bare surfaces were used to calculate an RMSE of 1.2 m. A map of the 8200 GCPs and 400 validation points is given in supplementary Figure S4 alongside a contour interval map of the interpolated surface in supplementary Figure S5. The interpolated surface was subtracted from the TDX DEM and intersected with the tree extent map to generate a canopy height map for the extent of the TDX imagery. Areas of continuous forest and elevated terrain were omitted as closed canopies prevented the accurate extraction of ground points and the interpolated surface could not be adequately fit in areas of rapidly changing slope. These regions contained predominantly continuous forests and not TOF. The DTM is provided in supplementary Figure S2.
Canopy height model calibration. The canopy height model was dependent upon the accuracy of the TDX DEM and the interpolated surface. Derived from radar, the elevation value is dependent upon the penetration of the radar wavelength into the canopy and was susceptible to changes with natural variations in the physical structure of the canopy. To calibrate and validate this, the TDX canopy height model (CHM) was compared to canopy surfaces captured in digital surface models (DSMs) derived from high resolution spaceborne imagery (HRSI). These HRSI DSMs were derived from stereoscopic image pairs collected during along-track acquisitions by the Worldview-2 satellite. This imagery has been demonstrated to successfully model canopy surfaces and derive canopy heights comparable with lidar [48][49][50] . WorldView stereo imagery was divided into calibration and validation datasets, using a 2013 dataset for calibration and a 2016 dataset for validation. The TDX canopy height model was subset to the WorldView calibration extent and both datasets were reprojected and resampled to 10 m and the TOF tree extent map masked out the ground values in both DEMs. The average elevation value from the TDX CHM and WorldView stereo imagery was extracted using 9499 1 ha plots. A 1 ha plot was chosen www.nature.com/scientificreports/ to reduce errors associated with geolocation or classification error. The relationship between the two elevation datasets assessed the performance of the TDX CHM (MAE: 3.38 m) and provided coefficients to improve the TDX CHM (Fig. 6A). This was applied to the canopy height model and the process was repeated with an independent WorldView Stereo image and 3024 1 ha plots, yielding a new relationship (Fig. 6B) and reduced MAE of 1.41 m. The validation dataset contained a small number (12) of erroneous values that were greater than twice the 99th percentile of the difference between the DEMs and were excluded. We set a minimum tree height> 2.19 m as derived from the canopy height calibration model, removing any erroneously classified TOF pixels at or below 0 m prior to CHM calibration, thus we confidently measure tree height above 2.19 m. The maximum tree height was limited to the 99th percentile for each division, in order to eliminate erroneous canopy heights. A map of the location of the 1 ha calibration and validation points are available in supplementary Figure S3.
Comparison with existing maps. Existing high-resolution maps of forest extent in Bangladesh were not available and the Potapov et al. 22 map was not openly available. One of the most widely used tree cover maps is that of Hansen et al. 32 , that maps global tree cover at 30 m using Landsat data. This dataset maps tree cover per 30 m pixel and is therefore not a direct comparison with the extent map presented here. In order to make these maps more comparable, the tree cover map was converted to a binary mask where all tree cover pixels over 0% were converted to binary tree pixels. The Hansen et al. 32 year 2018 baseline was combined with all gain pixels for Bangladesh up to the year 2012 to generate a maximum tree extent map for the period. Similarly, the DLR Tan-DEM-X Forest/Non-Forest map 33 was downloaded from DLR and was also compared with our TOF map. This dataset uses TanDEM-X interferometric SAR (InSAR) to classify forests globally at 50 m. This map, although at a coarser resolution to the global forest cover dataset, generates a binary map of forest extent that can be compared to our TOF map. Both datasets were either already available or uploaded to the Google Earth Engine, masked to match the TOF bounds and the area of forest in each division was calculated.
TOF accuracy assessment. The accuracy of the TOF map was assessed with reference to Microsoft Bing Airborne Imagery, available in QGIS. A custom plugin tool was used to manually verify 500 validation points for two classes (TOF/Non-TOF) per division. This created a total of 8000 manually validated points. Each point was buffered by 10 m to account for possible geolocation error or differences in the resolution of the TOF map and reference imagery. The 95% confidence interval on the area of each division was calculated following suggested best practices 51 . All points were combined to calculate the overall accuracy of the TOF map at the national level. The national confusion matrix is available in supplementary Table 1.