A 50-m forest cover map in Southeast Asia from ALOS/PALSAR and its application on forest fragmentation assessment.

Southeast Asia experienced higher rates of deforestation than other continents in the 1990s and still was a hotspot of forest change in the 2000s. Biodiversity conservation planning and accurate estimation of forest carbon fluxes and pools need more accurate information about forest area, spatial distribution and fragmentation. However, the recent forest maps of Southeast Asia were generated from optical images at spatial resolutions of several hundreds of meters, and they do not capture well the exceptionally complex and dynamic environments in Southeast Asia. The forest area estimates from those maps vary substantially, ranging from 1.73×10(6) km(2) (GlobCover) to 2.69×10(6) km(2) (MCD12Q1) in 2009; and their uncertainty is constrained by frequent cloud cover and coarse spatial resolution. Recently, cloud-free imagery from the Phased Array Type L-band Synthetic Aperture Radar (PALSAR) onboard the Advanced Land Observing Satellite (ALOS) became available. We used the PALSAR 50-m orthorectified mosaic imagery in 2009 to generate a forest cover map of Southeast Asia at 50-m spatial resolution. The validation, using ground-reference data collected from the Geo-Referenced Field Photo Library and high-resolution images in Google Earth, showed that our forest map has a reasonably high accuracy (producer's accuracy 86% and user's accuracy 93%). The PALSAR-based forest area estimates in 2009 are significantly correlated with those from GlobCover and MCD12Q1 at national and subnational scales but differ in some regions at the pixel scale due to different spatial resolutions, forest definitions, and algorithms. The resultant 50-m forest map was used to quantify forest fragmentation and it revealed substantial details of forest fragmentation. This new 50-m map of tropical forests could serve as a baseline map for forest resource inventory, deforestation monitoring, reducing emissions from deforestation and forest degradation (REDD+) implementation, and biodiversity.

Introduction

Dramatic changes in forests, especially tropical forests, have significant impacts on regional climate, water and carbon cycles as well as biodiversity [1], [2]. Southeast Asia consists of 11 countries (Cambodia, Laos, Myanmar, Thailand, Vietnam, Malaysia, Brunei, Indonesia, Philippines, Singapore, and East Timor) and its population increased rapidly from approximately 359 million in 1980 to 593 million in 2010 [3]. To meet the rising demand for food, fiber, water, and housing, substantial land use and land cover changes, especially deforestation, have taken place in the region. As the third largest area of tropical rainforests in the world following the Amazon and Congo Basin [4], [5], Southeast Asia experienced more dramatic deforestation than any other continent in annual rate in the 1990s [6]. This region was also a hotspot of forest cover change from 2000 to 2010, e.g., Margono et al. [7]. For example, Indonesia and Myanmar were listed among the top ten countries with the largest annual net loss of forests while Vietnam had a large annual net gain of forest area from 2000 to 2010 [8]. National forestry policies differ between these countries, which further contributes to different forest change patterns and asymmetric forest transitions affected by the wood product trades among them [9]. Although government and scientists have made efforts to reinforce forest protection in Southeast Asia, forest conversion to plantations (e.g. oil palm) has continued to increase, which leads to fragmentation and affects biodiversity and carbon sequestration in peatlands [10]–[12]. To support regional sustainable development, including forest management, carbon emission estimation, habitat planning, and biodiversity conservation, it is critical that accurate and updated information on forest area, extent, fragmentation and change is developed [13], [14].

Several efforts have been carried out to map forest extent and change in Southeast Asia using optical remotely sensed data from the Advanced Very High Resolution Radiometer (AVHRR) [15]–[17], SPOT-Vegetation [4], [18], Moderate Resolution Imaging Spectroradiometer (MODIS) [14], [19], [20], and Medium Resolution Image Spectrometer (MERIS) [21]. However, the area estimates of forest cover in Southeast Asia from the above-mentioned studies differ substantially [4], [15]–[18], [22]. For example, the MODIS-based land cover product (MCD12Q1) estimated a forest area of 2.69×106 km2 in Southeast Asia in 2009, the MERIS-based GlobCover land cover product estimated a forest area of 1.73×106 km2 in 2009, and the FAO Forest Resources Assessment (FRA) 2010 reported a forest area of 2.14×106 km2. All these three data products are widely used today for forest resource survey, climate simulation and biodiversity conservation [23]. The large discrepancy among these forest maps may be attributed to (1) fragmented forests or mixed pixels at moderate spatial resolutions (several hundreds of meters to 1-km), (2) frequent cloud cover [23]–[25], and (3) individual definitions and algorithms for forests. Although Landsat TM/ETM+ imagery (30-m spatial resolution) is available and has been widely used to map forests [14], [19], [22], [26], frequent cloud cover in the humid tropical zone makes it difficult to obtain cloud-free Landsat images over the entirety of Southeast Asia during a certain period. Consequently, the sample-based approach was used to select cloud-free Landsat images and map forest cover change [6], [14], [27]–[29]. This sampling approach does provide statistically-inferred information on forest cover change at national and continental scales, but forest management and decision-making for forest resource planning and biodiversity conservation requires location-specific, detailed, and updated annual maps of forests. Therefore, annual continental-scale forest maps of Southeast Asia at higher spatial resolutions (<100 m) are critical and urgently needed.

Synthetic aperture radar (SAR) with fine spatial resolution provides cloud-free imagery and is an alternate source for tropical forest mapping [30]–[32]. A long radar wavelength (e.g. L-band SAR) has an improved capability to delineate high-biomass forest than shorter wavelengths (e.g., C-band SAR) because of its greater penetration capability through the tree canopy [33]. Early in the 1990s, a single HH polarization dataset from the Japanese Earth Resources Satellite (JERS-1) was used for forest mapping (e.g. clear-cut) [34], [35]; however, continental forest maps based on JERS-1 imagery have not yet been developed. The Phased Array Type L-band Synthetic Aperture Radar (PALSAR), onboard the Advanced Land Observing Satellite (ALOS) launched by the Japan Aerospace Exploration Agency (JAXA) in January of 2006, provides an enhanced capacity for forest mapping and deforestation detection [31]–[33], [36]–[43]. JAXA has generated the first 10-m global forest/non-forest map by using ALOS/PALSAR data [44], [45], and the Woods Hole Research Center in Massachusetts, USA, is developing a pan-tropical forest cover map for subsequent deforestation and forest degradation monitoring by using ALOS/PALSAR data [46]. However, both of these are still unavailable to the public at this time.

The objective of this study was twofold: (1) to generate a map of tropical forest cover in Southeast Asia in 2009 at 50-m spatial resolution; and (2) to evaluate the fragmentation of forests in the region, based on the new and improved forest cover map. The forest map was generated by using the publicly available ALOS/PALSAR 50-m orthorectified mosaic imagery and the decision tree method reported in a previous study [47]. We evaluated the resultant map with validation sample data derived from the ground truth field photos at the Global Geo-referenced Field Photo Library (http://www.eomf.ou.edu/photos) and Google Earth. It was also compared to two regional/global land cover maps and to the FAO FRA 2010 statistics to supplement the validation. Based on this first continental-scale forest map at 50-m spatial resolution in Southeast Asia, forest fragmentation was investigated by using the fragmentation model presented by Riitters et al. [48], [49]. We aimed to better understand the fragmentation of forests in the region, which may help improve habitat planning and biodiversity conservation.

Data and Methods

PALSAR 50-m Orthorectified Mosaic Product and pre-processing

As part of the ALOS Kyoto and Carbon Initiative Project, JAXA released the PALSAR 50-m Orthorectified Mosaic Imagery Product in 2007, 2008, and 2009 for many parts of the world, including Southeast Asia. In this study we used the PALSAR 50-m Orthorectified Mosaic Product derived from images collected from June to October in 2009, and it has HH (horizontally transmitted and horizontally received) and HV (horizontally transmitted and vertically received) polarizations. The datasets are freely available to the public at the ALOS Kyoto and Carbon Initiative official website (http://www.eorc.jaxa.jp/ALOS/en/). The original PALSAR data with the observational mode of Fine Beam Dual (FBD) polarization has an off-nadir angle of 34.3 degrees, a range resolution of 14–88 m, and they have been geometrically rectified and mosaicked [50]–[52]. The Digital Number (DN) values (amplitude values) of these images were converted into the normalized radar cross section in decibel (σ, with the unit of dB) according to the following formula [36],where DN is the original digital number value in HH or HV polarization, and CF is the absolute calibration factor and equal to −83.

We generated two additional images from HH and HV images: (1) the band ratio (HH/HV) image and (2) the band difference (HH–HV) image. Both the difference image and band ratio image have proven valuable for land cover classification [53], [54] as they provide additional information of different land cover types. For example, the difference image was used to separate palm plantations and other trees [10]. The false color composite map (Fig. 1) shows the separability capability of the PALSAR image in four main land cover types.

Figure 1

The false color composite of PALSAR 50-m orthorectified mosaic imagery (R/G/B  =  HH/HV/HH-HV) in Southeast Asia in 2009.

Country names were labeled as Myanmar (A), Thailand (B), Laos (C), Vietnam (D), Cambodia (E), Malaysia (F), Brunei (G), Indonesia (H), Philippines (I), Singapore (J), and East Timor (K). The inset graphs show forest, cropland, water body, and built-up land, respectively. The PALSAR 50-m mosaic data was unavailable in the West Papua and Papua regions.

Ground-based Points of Interest (POIs) data

Geo-referenced field photos available at the Global Geo-Referenced Field Photo Library (see http://www.eomf.ou.edu/photos/, available to the public) at the University of Oklahoma (Xiao et al., 2011) were used as ground-based points of interest (POIs) data. The field photos and associated land cover information were provided by citizen scientists and researchers and archived at the Field Photo Library. We acquired the field photos only at unrestricted locations where no special permission was required and avoided the national park and other protected areas of land or sea related to wildlife. We confirm that our field studies did not involve endangered or protected species or provide specific location information for such.

Researchers usually take photos to document their study sites and landscapes by using (a) GPS cameras, (b) smartphones, or (c) digital cameras plus handheld GPS receivers. The Field Photo Library is a community remote sensing and citizen science data portal for archiving, sharing, and exchanging these geo-tagged field photos and associated thematic databases of land cover types from the individual photos [55]. It hosts field photos contributed by our research team and other researchers around the world. Many field photos were interpreted and labeled by photo providers with the land cover types they represent, which results in a thematic database of land cover types, associated with the field photos. The geo-referenced field photos can be downloaded in different formats, compatible with Google Earth (kml or kmz), ArcGIS (shapefile), and attribute database software (attribute tables). The data portal also provides a MODIS data extraction function that allows people to retrieve time series MODIS data (e.g., surface reflectance, vegetation indices) for individual pixels with geo-referenced field photos (Fig. 2a), which enables rapid analysis of phenology at individual POIs. With the contributions from citizen scientists and other researchers, the Field Photo Library now hosts over 30, 000 photos for Southeast Asia.

Figure 2

The spatial distribution of geo-referenced field photos in the study area, as hosted in (A) the Global Geo-referenced Field Photo Library.

The circle size indicates the number of the field photos. The figure also shows the search options, selected photos with GPS locations, and the link to extract the MODIS time series data. More information can be found in the data portal (www.eomf.ou.edu/photos). (B) The Regions of Interest (ROIs) used for the algorithm training and results validation, which were acquired by referring to the field photos shown in Fig. 2A and Google Earth.

In this study, we first selected field photos that we collected during our field trips in Thailand, Vietnam and Indonesia; ∼7,000 geo-referenced field photos were collected in the past few years. Casio EX-H20 GBK Hybrid-GPS digital cameras were used in field trips to acquire GPS-referenced photos. All participants in the field trips used a standard protocol for taking photos in the field. For photos that describe high quality study sites, one takes photos from the center of the area of interest (one photo looking down and one photo looking up) and photos in each of four directions (N, E, W, and S) for a general description of the landscape, which results in a minimum 6 photos per site. For photos that describe only landscapes, photos were taken often from one or two directions (see the Field Photo Library website for detailed description of the field photo protocol). For the countries with limited field photos (Myanmar, Cambodia, Malaysia, and Philippines), ∼500 landscape photos with GPS information were collected from the Degree Confluence Project (confluence.org) and the Panoramio (www.panoramio.com). Together, a comprehensive coverage of geo-referenced field photos was achieved in all the countries except for Brunei, East Timor, and Singapore (Fig. 2a).

Regions of Interest (ROIs) for algorithm training and product validation

We combined the geo-referenced field photos (POIs) and high-resolution images in Google Earth to generate homogeneous land cover polygons (polygon sampling units). The procedure is composed of two steps: (1) overlay geo-referenced field photos with high-resolution images in Google Earth; and (2) digitize high-resolution images in Google Earth to generate polygons. Previous studies have shown that Google Earth is feasible for ROI digitization of land cover classifications [56]–[59]. Our previous studies also showed that integrating the Field Photo Library and Google Earth is reliable [47], [60]. During the digitalization process, the minimum size of a polygon is required to be at least 3 times larger than the minimum land size used in the forest cover definition. For example, FAO provides a classic forest definition with three components: 1) tree canopy cover >10%, 2) tree height >5 m, and 3) minimum land size 0.5 ha [61]. We used this FAO definition of forest cover and ensured that each polygon has a size of at least 3 times larger than 0.5 ha and tree canopy cover >10%.

The resultant polygon sample units, called the regions of interest (ROIs), were then used for both algorithm training and product validation, respectively. A total of 78 ROIs for four land cover types (forest, cropland, water body, and built-up land) were generated as the algorithm training ROIs (Fig. 2b), and they were the same as those used in our previous study in mainland Southeast Asia [47], including 25 forest ROIs (997 986 pixels), 32 cropland ROIs (160 916 pixels), 10 water ROIs (303 948 pixels), and 11 built-up land ROIs (26 970 pixels). As the PALSAR L-band backscatter reflects more about physical characteristics, we assume the same algorithm will work in the insular Southeast Asia as well. The initial aim of this study is to verify whether the algorithm developed in the Mainland Southeast Asia [47] is suitable in insular Southeast Asia.

We also developed a second set of ROIs for validation of the results in this study (Fig. 2b). A total of 1,233 ROIs (11.2×105 pixels) were collected, including 422 forest ROIs (458 488 pixels), 599 cropland ROIs (129 655 pixels), 70 water ROIs (445 538 pixels), and 142 other land cover ROIs (mainly built-up land, 88 019 pixels). The mean size of forest ROIs was bigger than those of cropland and built-up land, as the cropland and built-up landscapes are fragmented in Southeast Asia.

Land cover classification based on PALSAR data and decision tree algorithm

Signature analyses of PALSAR backscatter values were conducted based on the ROIs for algorithm training. We calculated the mean and frequency histograms of backscatter values for the four land-cover types. These land cover types have distinguishable backscatter characteristics. Water has the lowest HH and HV backscatter due to its depolarization effect and strong absorption of energy. According to our definitions in this study, forest has a height over 5 m and has higher HH and HV backscatter due to more canopy backscatter in tree trunks and leaf canopy. Croplands have lower HH and HV backscatter than forest. Built-up lands show very complex characteristics as buildings are complex and building orientations and corner reflectors exist [47].

The thresholds in decision tree rules were determined according to a statistical analysis of these ROIs. When we digitized the training ROIs, we tried to find the area with pure land covers, by assuming that the pixels within a 95% confidence interval of backscatter values are pure land cover types, and other pixels with the lowest and highest 2.5% backscatter values are likely mixed land cover types. The resultant decision tree rules and PALSAR data enable a simple and consistent approach to generate a continental scale forest cover map in Southeast Asia. For water body pixels, HH<−16 and HV<−24 threshold values were used. For forest pixels, we used 3.547]. The land cover map was generated by using the decision tree function in ENVI 5.0 software (Fig. 3), and then converted into a forest/non-forest map.

Figure 3

The resultant land cover map of Southeast Asia based on the PALSAR 50-m orthorectified mosaic data in 2009 and the decision tree algorithm.

Validation with ground truth data

A systematic and rigorous validation is always an essential concern in land cover classification, and is often carried out by comparing resultant maps with (a) a higher-resolution imagery or land cover product, and/or (b) ground truth samples from field trips. Field trip samples are the most straightforward validation references. Based on the validation ROIs mentioned in Section 2.4, validation of the PALSAR-based land cover map was conducted, and the confusion matrix was reported in Table 1, including user's accuracy, producer's accuracy, and overall accuracy.

Table 1

The confusion matrix between PALSAR-based land cover classification in Southeast Asia and Regions of Interest (ROIs) by referring to field photos and Google Earth.

	Class	Ground truth (GT) samples (pixels)				Total classified pixels	User Acc. (%)
		Forest	Cropland	Water	Others
Classification	Forest	392800	2383	91	29036	424310	93%
	Cropland	9837	102247	10434	1001	123519	83%
	Water	25	18090	434735		452850	96%
	Others	55826	6935	278	57982	121021	48%
Total GT pixels		458488	129655	445538	88019
Prod. Acc. (%)		86%	79%	98%	66%

Comparison between PALSAR-based forest map and other products

In addition to the above-mentioned validation with ground truth data, we also compared the PALSAR-based forest map with the two land cover products at three levels (national, subnational, and pixel levels): (1) GlobCover 2009 [21] and (2) MCD12Q1 2009 [19]. The FAO FRA 2010 statistics [8] were also used to do comparisons with the other three land cover products at the national level as it has no statistical data at subnational or smaller scales. Subnational boundary data were obtained from the GADM database of Global Administrative Areas (http://www.gadm.org/home). The pixel level comparison was conducted by aggregating these spatial forest datasets (PALSAR, MCD12Q1, and GlobCover) into forest fractional maps with 1.5-km×1.5-km resolution (1.5-km is the lowest common multiple of 50-m, 300-m, and 500-m).

As different land cover products have different classification schemes, we merged forest-related classes before the comparisons. GlobCover has 22 land cover types, and seven forest-dominated types were combined into one forest layer: closed to open broadleaved evergreen or semi-deciduous forest, closed broadleaved deciduous forest, open broadleaved deciduous forest/woodland, closed needle-leaved evergreen forest, open needle-leaved deciduous or evergreen forest, closed to open mixed broadleaved and needle-leaved forest, mosaic forest or shrubland/grassland. MCD12Q1 has several land cover classification schemes, and we used the classification scheme from the International Geosphere Biosphere Programme (IGBP) with 17 land cover types. Five forest types (evergreen needleleaf forest, evergreen broadleaf forest, deciduous needleleaf forest, deciduous broadleaf forest, and mixed forest) were combined into one forest layer. Land cover categories in FRA 2010 include forest, other wooded land, other land and inland water bodies [8]; only the forest category was used in this study. In addition, the GlobCover and MCD12Q1 datasets used different forest definitions. The GlobCover dataset uses 15% tree canopy cover and >5 m tree height, while the MCD12Q1 dataset uses 60% tree canopy cover and >2 m tree height [19], [21].

Forest fragmentation analysis

Forest fragmentation has substantial impacts on animal and plant habitat quality and biodiversity [48], [49]. One forest fragmentation model was recently developed and it considers two indicators: forest area density (P) and forest connectivity (P), within a certain “window” or “landscape” [48]. It has been widely used to assess the forest fragmentation by using various satellite-derived forest maps [62], [63]. We employed this forest fragmentation model to conduct a forest fragmentation analysis. The forest fragmentation was calculated based on a forest/non-forest binary map (forest = 1 and non-forest = 0). The two indicators were calculated with the following equations, where P is the proportion of forest pixels in a certain window (e.g., 9×9), and is calculated by dividing forest pixels (N) in a certain window by the total number of pixels (N); P is the forest connectivity, and is calculated by dividing the pixel pair number that includes at least one forest pixel (D) by the pixel pair number that includes two forest pixels in cardinal directions (D).

The resulting forest fragmentation map is grouped into six fragmentation categories. The fragmentation categories and their criteria are: (1) Patch, if P <0.4; (2) Transitional, if 0.4