- Top of page
- LITERATURE CITED
- Supporting Information
Quantifying forest change in the tropics is important because of the role these forests play in the conservation of biodiversity and the global carbon cycle. One of the world's largest remaining areas of tropical forest is located in Papua New Guinea. Here we show that change in its extent and condition has occurred to a greater extent than previously recorded. We assessed deforestation and forest degradation in Papua New Guinea by comparing a land-cover map from 1972 with a land-cover map created from nationwide high-resolution satellite imagery recorded since 2002. In 2002 there were 28,251,967 ha of tropical rain forest. Between 1972 and 2002, a net 15 percent of Papua New Guinea's tropical forests were cleared and 8.8 percent were degraded through logging. The drivers of forest change have been concentrated within the accessible forest estate where a net 36 percent were degraded or deforested through both forestry and nonforestry processes. Since 1972, 13 percent of upper montane forests have also been lost. We estimate that over the period 1990–2002, overall rates of change generally increased and varied between 0.8 and 1.8 percent/yr, while rates in commercially accessible forest have been far higher—having varied between 1.1 and 3.4 percent/yr. These rates are far higher than those reported by the FAO over the same period. We conclude that rapid and substantial forest change has occurred in Papua New Guinea, with the major drivers being logging in the lowland forests and subsistence agriculture throughout the country with comparatively minor contributions from forest fires, plantation establishment, and mining.
Sopos long kisim gutpela save long senis i kamak long tropics em i wanpela bik pela samting long wanem, bikpela bus em wanpela hap we wok konsevason na carbon cycle bai inap kirapim gutpela wok. Insait long olgeta hap long world, PNG em wanpela hap we bikpela bus em i stap yet. Insait long dispela wok mipela soim olsem bikpela senis em i kamap long insait long bikpela bus na long hamas bikpela bus yumi gat. Nogat wanpela kain wok painimaut emi painim dispela senis bipo. Mipela lukluk gut long we olgeta bikpela bus i raus na we bus i kisim bagarap insait long, yia 1972 i kamap inap long yia 2002. Long yia 1972 mipela i usim map ol i kolim land cover map na long yia 2002 mipela lukluk long olgeta PNG high-resolution satellite imagery. Long yia 2002, 28,251,967 hectares bikpela bus i stap insait long Papua New Guinea. Long namel long 1972 igo inap long 2002, Papua New Guinea i lusim 15 percent long algeta bipela bus belong en. Insait long dispela 15 percent, 8.8 percent em i kamap bikos ol lain i katim diwai long salim. As bilong senisim bikela bus emi stap long ples we igat bikpela diwai long katim. Insait long dispela hap yumi lusim 36 percent, sampela we yumi inap long salim, tasol narapela emi bikos yumi rausim bus long wokim gaden or narapela kainkain pasin yumi wokim. Long 1972 i kamap inap long yia 2002, yumi lusim 13 percent long bikpela bus raonim ol bikpela maunten. Mipela painim olsem, long yia 1990 igo inap long yia 2002, long algeta kantri kain senis i wok long kamap bikpla. Senis istap insait long 0.8 igo inap long 1.8 percent long wan wan yia, tasol insait long wan wan liklik hap some pela i kisim bikpela senis, na ol narapela ino tumas. Long ol hap igat gutpela diwai long katim, senis i stat long 1.1 percent igo inap 3.4 percent. Dispela senis em i winim estimates we ol lain FAO i bin tokaut long em bipo. Long dispela wok painimaut, mipela iken tok olsem, as bilong dispela bikpela senis emi kamap long wanem ol i rausim na bagarapim bikpela bus. Dispela asua i kamap taim yumi rausim planti diwai tumas long salim na sampela taim yumi katim bus long wokim garden. Sampela taim bikpela paia tu i save kukim bikpela bus.
Tropical forests are undergoing wide-scale deforestation and degradation, documented by the use of remote sensing data (Asner et al. 2005, Defries et al. 2007). Papua New Guinea (PNG) contains approximately half of the third largest extant area of tropical rain forest in the world, and is one of the most biodiverse and ecologically distinct forested regions (Brooks et al. 2006), highly important for both biodiversity conservation and carbon capture. Subsistence agriculture, forestry, fire, plantation development, and mining have all driven deforestation in PNG (McAlpine & Freyne 2001, Haberle 2007). Forest degradation has also occurred, largely as the result of conversion of primary forest into secondary forest by commercial logging. An intact canopy may regenerate within a few years after logging (Steininger 1996, Nepstad et al. 1999). However, beneath the canopy, logging results in reduced biomass, damage to other vegetation and soils, and increased vulnerability to both fire and subsequent conversion to grassland, scrub, or agricultural land that may persist for decades (Holdsworth & Uhl 1997, Nepstad et al. 1999, Asner et al. 2005, Defries et al. 2007). This damage requires logged forest to be treated separately from unlogged forest both from an ecological and carbon perspective (Foley et al. 2007).
The State of PNG is comprised of the eastern half of the island of New Guinea (termed here ‘the Mainland region’), the islands of New Ireland, New Britain, and Bougainville, and many smaller islands (the ‘Islands region’). The mainland of PNG possesses a rugged central mountain range reaching to more than 4500 m elevation that is flanked to the north and south by a comparatively flat lowland region. In contrast to other tropical regions, the majority of PNG's population still practice subsistence agriculture. Customary ownership of land is protected under the nation's Constitution (National Statistical Office 2000). The primary land-uses in PNG are subsistence agriculture and commercial forestry, although mining and commercial agriculture are also significant (McAlpine & Freyne 2001). In contrast to many other tropical regions, these land-use practices are still largely spatially segregated—the customary land-tenure system has to date largely prevented migration, settlement and subsequent clearance of logged forests (International Tropical Timber Organisation 2007). Due to the rugged and mountainous terrain, large areas of forest are relatively inaccessible to commercial forestry, hence PNG's timber resources are mostly confined to the coastal lowlands and offshore islands.
In the last decade the population of PNG has increased dramatically, as have timber exports from commercial forestry operations and oil palm exports from agricultural plantations (Gresham 1982, Filer 1997, Hunt 2002, FAO 2007; PNG Oil Palm Research Association, pers. comm.). Additionally there have been large forest fires, especially during the 1997–1998 El Niño event (Haberle et al. 2001). Prior to our study there had been no recent nationwide high-resolution assessment of PNG's forest extent and condition, the rate at which these forests are being cleared, or the impact of three decades of commercial export-driven logging. Despite this lack of data, the Food and Agriculture Organization (FAO) reported a low deforestation rate for 2000–2005 of 0.5 percent/yr (FAO 2005a). This low rate was suggested to be the result of an increased intensity of subsistence farming in response to population growth (McAlpine & Freyne 2001).
Our study maps recent (2002) forest extent and condition, and examines long-term forest change in PNG's forests overall, as well as in the extent of commercially accessible forest. We were able to conduct such a long-term high-resolution forest change analysis due to the availability of an Australian Army vegetation classification of very high-resolution aerial photography from the 1970s covering all of PNG (Coulthard-Clark 2000). Our primary objective in the study was to create and validate high-resolution land-cover maps to serve as baselines from which future change could be measured. Our second objective was to examine the extent and causes of deforestation and degradation across PNG's three distinct provincial regions: the islands, the highlands, and mainland lowlands (Fig. 1). For each discrete area of deforestation or degradation, we identified the main driver of forest change—either forestry, subsistence agriculture, plantation development, forest fires or mining activities, in PNG's overall forest extent, as well as in its commercially accessible forest extent. Our third objective was to estimate recent rates of deforestation and degradation in both commercially accessible and overall forest extent through the use of a driver-specific model, and to examine the relationship between population and subsistence-related clearance.
Figure 1. Provincial boundaries in PNG. Each province is shown colored by the broad region in which it is situated, the Islands, the Highlands, or the Lowland coastal regions. Provincial abbreviations: BOU = Bougainville; CEN = Central; EHY = Eastern Highlands; ENB = East New Britain; ENG = Enga; ESK = East Sepik; GUL = Gulf; MAD = Madang; MAN = Manus; MIL = Milne Bay; MOR = Morobe; NIR = New Ireland; ORO = Oro; SHY = Southern Highlands; SIM = Chimbu; WES = Western; WHY = Western Highlands; WNB = West New Britain; WSK = West Sepik.
Download figure to PowerPoint
- Top of page
- LITERATURE CITED
- Supporting Information
PNG is located in the South Pacific, north of Australia and east of Indonesia. Annual mean rainfall varies from 1000 mm to > 8000 mm, and mean annual temperatures range from < 8°C to > 27°C in the lowlands, varying largely with elevation (McAlpine et al. 1983). The rugged and mountainous terrain prevailing over much of PNG means that regular acquisition of wall-to-wall cloud-free imagery is not possible, and automated classification of imagery is problematic (Colby & Keating 1998). Thus, although classification of sequential high-resolution satellite images can provide an accurate assessment of deforestation and degradation (Asner et al. 2005, Defries et al. 2007), the nature of the PNG landscape requires the adoption of novel techniques to map forest area and detect forest change.
A ‘forest’ is here defined as natural woody vegetation that has a contiguous tree canopy and a canopy height > 5 m. These forests are also known as tropical rain forest and are distinguished from open-sclerophyll forest, woodland, savannah, mangrove, and swamp vegetation. We confined our change analysis to rain forest.
Creation of 1972 land-cover map.— The 1972 land-cover map was digitized from color separations of Australian Army (T601) 1:100,000 scale vegetation maps. These maps were created by the Australian Army using visual classification and manual delineation of polygon class boundaries of vegetation types discernible in very high-resolution (1–2 m) stereo aerial photography taken largely between 1972 and 1975, but predominantly (68%) in 1972–1973 (Coulthard-Clark 2000). The 23 percent of aerial photographs that were recorded prior to 1970 were predominantly located in remote and sparsely populated parts of Western province where little forest change has occurred. We have termed the resultant land-cover map a ‘1972’ coverage. The vegetation classes used in this exercise were intact forest, mangroves, scrub, grassland, and water. For the change analysis, land-cover classes were combined into two categories: forest (rain forest) and nonforest (non-rain forest). All forests in the 1972 map were assumed to be primary forests as levels of commercial logging were comparatively low prior to this time (Hammermaster & Saunders 1995, McAlpine & Freyne 2001).
Creation of 2002 land-cover map.— To create a new land-cover map of PNG, we acquired multiband imagery from the Landsat Enhanced Thematic Mapper Plus (ETM+) and Systeme Pour l’Observation de la Terre (SPOT) 4 and SPOT 5 sensors. Due to almost perpetual cloud-cover, we were unable to obtain suitable cloud-free 30-m resolution (15 m panchromatic) Landsat ETM+ imagery for the whole of PNG. In areas not covered by Landsat, we acquired SPOT 4 or SPOT 5 (20- and 10-m resolution, respectively) instead. About 18 percent of PNG was classified from imagery recorded in 2000–2001, 62 percent in 2002, and 20 percent in 2003–2007: accordingly we term the land-cover map ‘2002.’ The location and date of capture of each image used is shown in Figure S1 and Table S1. Images recorded prior to 2002 were in general located in remote areas unlikely to have undergone substantial change. Images collected after 2002 were located across various parts of the country, some parts where forest change would be expected and some where no change would be expected.
In total we used 61 Landsat ETM+, 59 SPOT 4 and 5 images, and seven Landsat TM images to obtain cloud-free coverage for the whole of PNG. Each satellite image was orthorectified using 10–15 ground control points derived from the 1972 land-cover maps (Coulthard-Clark 2000) and a 90-m Digital Elevation Model (DEM) obtained from the Shuttle Radar Topography Mission (SRTM, Farr et al. 2007). The average root mean square error for the orthorectified imagery was 25–30 m for the Landsat and 15–20 m for the SPOT 4 and 5 imagery.
We applied a Tasseled Cap and Brovey transformation (Kauth & Thomas 1976) to our 2002 Landsat ETM+ imagery, and red/green/blue/infrared color enhancement to our SPOT 4 and 5 imagery. We used the object recognition software ‘eCognition’ (Definiens 2005) to automatically segment the satellite imagery into spatially continuous and spectrally homogeneous regions consistent with land-cover features, and to vectorize them into individual polygons. The process of image segmentation and vectorization is summarized in Figure 2.
Each polygon was classified using expert visual interpretation (Lu et al. 2004). Decision rules used to define land-cover classes are outlined in Table S2. Basic land-cover classes followed the classification system of Paijmans (1976), and were: tropical rain forest (referred to here as ‘forest’), swamp forest, dry evergreen forest, mangroves, scrub, herbaceous swamp, nonvegetation, water, and grassland/savannah. Within the change analysis, land-cover classes were grouped into two categories: forest (rain forest) and nonforest (non-rain forest—dry evergreen forest, swamp forest, mangroves, scrub, herbaceous swamp, nonvegetation, water and grassland). An illustrated example of the classification process is contained in Figure S2.
Change detection.— The 1972 land-cover map was superimposed onto the 2002 map using a geographic information system, and areas of forest loss and gain that have occurred since 1972 were identified and automatically vectorized. The change detection process is depicted in Figure S3. Our resulting estimates of forest change are net forest losses because gain was subtracted from forest loss. Gross area converted (loss) and area reforested (gain) for each province are contained in Table S3.
Assessment of drivers of forest change.— We mapped recent logging-related deforestation and degradation using the visual interpretation of logging roads, snig tracks, and canopy gaps to delineate a timber extraction radius (the area beyond a timber road or snig track in which timber has been extracted) for all commercial logging activity. Snig tracks, also known as skid trails, are temporary trails created by dragging felled trees or logs to a logging road. Forest within our timber extraction radius was designated as ‘degraded’ and clearances were designated as ‘deforested’ due to logging. In these areas the timber extraction radius was at maximum 500 m, but most often 100–300 m. Areas within logging concessions outside of this radius were assumed not to have been logged.
In those areas where logging ceased more than a decade ago and where logging roads were themselves not visible, data from previous helicopter surveys of logging activity (Hammermaster & Saunders 1995) were used to demarcate logged areas. This use of the helicopter survey data only occurred in relatively few locations as in the majority of locations older roads are still visible because of recent use unrelated to logging, or because those areas that were logged in the 1980s have since been cleared.
Forest clearance as a consequence of subsistence agriculture frequently involves burning that commonly spreads a distance into adjacent vegetation. This clearance was designated as being subsistence-related. There were, however, some extensive fires in primary forest not associated with subsistence agriculture. In many locations, large areas of burned forest could be recognized by their distinct spectral response (Nepstad et al. 1999) coupled with reports from field workers. The identification of ‘burned’ forests was assisted in numerous locations by the presence of fire ‘thermal hotspots’ derived from the Moderate Resolution Imaging Spectroradiometer product (MODIS 14) that showed the location of recent fires (Giglio et al. 2003). In other locations, areas of fire-related loss were recognized due to their large distance (> 10 km) from zones of subsistence or commercial use, their distinct pattern, or in the case of montane forest loss, their location on the tops of mountain ranges, usually adjacent to existing grassland. We defined the upper montane zone in our land-cover maps as the region > 2800 m elevation, consistent with the definition of the ‘Tropical Montane’ life zone in PNG (McAlpine et al. 1983). The region > 2800 m elevation was delineated using a 90-m SRTM DEM (Farr et al. 2007).
Areas of deforestation that occurred within currently established plantations were deemed to have been cleared to create the plantation. Areas of forest replaced by the open pits of mining operations and associated infrastructure, as well as the tailings-related dieback downstream from the Ok Tedi mine were deemed to have been cleared due to mining. This assessment of mining-related deforestation does not include increased agriculture that may be associated with the mining projects (Dambacher et al. 2007).
Subsistence agriculture was usually discernible in the imagery. Areas of forest loss adjacent to either subsistence gardens or rural villages were deemed to be subsistence-related. Areas of deforestation that could not be attributed to other drivers of forest change were also assumed to have occurred through subsistence activities.
Assessment of commercially accessible forests.— Areas of polygonal karst, slopes too steep for mechanized logging, and Wildlife Management Areas were delineated as physically ‘inaccessible.’ Forests growing on polygonal karst in our 1972 and 2002 land-cover map were identified using a landform map (Loffler 1977) and our 2002 satellite imagery. Slopes too steep to log were identified using a slope surface derived from a 90-m resolution DEM. We created the DEM from Shuttle Radar Topography Mission (SRTM) data (Farr et al. 2007), corrected for the effects of radar shadow in steep valleys using a DEM created from 1:100,000 and 1:250,000 contour maps of PNG (Coulthard-Clark 2000). The resulting DEM was calibrated against a high-resolution (15-m) LIDAR DEM acquired in 2007 over a 60 × 60 km area centered on the Port Moresby region. We found that ca 95 percent of the logged areas in our 2002 land-cover map occurred on slopes < 25°. We therefore used 25° as the limit for commercial logging access in PNG. This concurs with the assessment in the Wet Tropics of North Queensland that forests on slopes greater than 28° had a zero probability of being logged (Vanclay 1994).
Forested areas too small to support a commercial logging operation (50,000 ha) (World Bank & GoPNG 2001) were delineated as economically ‘inaccessible.’ The World Bank determined that commercially viable forestry operations in PNG required an annual timber extraction volume of 70,000 m3 over 35 yr (World Bank & GoPNG 2001). We conservatively applied a smaller area (50,000 ha) as the lower limit for commercial viability since many concessions are only exploited for 10–15 yr and many logging companies operate under marginal economic conditions (Overseas Development Institute 2006).
Rates of deforestation and degradation.— Change rates are commonly calculated by dividing the measured area of forest loss as a percentage of forest area evenly over each year in the time period of the study (Puyravaud 2003). This method assumes constant decline over the time period. As our analysis was long term, the assumption of constant decline was implausible. This is because timber exports, oil palm exports, and population, associated with forestry, plantation, and subsistence clearance, increased substantially in the latter half of our 30-yr time period (Fig. S4). We therefore calculated change rates by apportioning the measured net area of deforestation and degradation to each year over the time period by modeling annual fire, forestry, plantation, mining and subsistence-related clearance.
We apportioned the total area of forestry-related deforestation and degradation to each year according to the volume of timber exported from logging concessions in that year (Gresham 1982, Filer 1997, Hunt 2002). The total area of plantation-related deforestation was apportioned on an annual basis according to the volume of oil palm exported 4 years later (PNG Oil Palm Research Association, pers. comm.). We lagged oil palm exports by 4 years to account for the delay between forest clearance for planting and harvest. The total area cleared due to subsistence agriculture was apportioned to individual years using the annual population estimates between 1972 and 2002 (FAO 2007). We apportioned the fire-related clearance to individual years based on the occurrence, intensity, and duration of El Niño events between 1972 and 2002 (Yue 2001). Mining-related clearance was apportioned evenly to each of the 30 years due to lack of data on the periodicity of mine-related deforestation. This is unlikely to impact our estimates of overall rates of forest change as the area cleared for mining is relatively small. We then estimated the annual rate of deforestation and degradation in each year as the percentage of the forest area from the previous year, which had been cleared or degraded by the following year (full methods are available in Appendix S1).
Validation of the 2002 land-cover map.— Validation of our 2002 classification was conducted using two low elevation aerial photographic surveys (resolution: 0.1–1 m) of 431 locations in West New Britain and Madang provinces during 2004 and 2008. All land-cover types in the 2002 classification were present in these surveys. The flight paths of the surveys are shown in Figure S5. In total 431 vertical photos were captured, mostly at ca 1000 m above the ground, and their location was recorded using a global positioning system (GPS). The area covered by each image was approximately 1 km2. Each image was manually assigned to one of the classes used in the classification process. Where a photograph contained more than one class, the image was subdivided into nine equal rectangular areas and the central rectangle was classified. The classification was then compared to the classification of the satellite imagery over the same location.
Accuracy assessment of 1972 vegetation classification and change detection.— Assessing the accuracy of change detection procedures is difficult (Lu et al. 2004). This is especially so in PNG where additional recent imagery is difficult to acquire, and where suitably accurate data sets from the 1970s, independent of our own 1972 map, do not exist. Consequently we assessed change detection accuracy using the area of forest gained between 1972 and 2002. The 1972 land-cover map provided a reliable indication of actual vegetation cover across the entire nation due to its very high resolution and the extensive georeferencing and field verification undertaken during its production by the Australian Defence Force (Coulthard-Clark 2000). The change assessment, comparing our 1972 and 2002 land-cover maps, was bidirectional, estimating both loss and gain in forest cover. Forest gain occurred either through the conversion of nonforest to forest cover, or through error. Grassland could not become intact forest within a period of < 30 yr (Walker 1966, Gillison 1969, Duncan & Duncan 2000, Haberle 2007), so forest gain between 1972 and 2002 consisted of the regeneration of scrub to forest or was an artifact of error in the comparison. Forest gain therefore provided an upper limit on error in the change assessment.
- Top of page
- LITERATURE CITED
- Supporting Information
TABLE S1. Date, type, location and area covered by satellite images used to produce the 2002 land-cover maps.
TABLE S2. Characteristics used to define land-cover classes.
TABLE S3. Gross forest area converted and reforested in each province.
TABLE S4. Area and change in commercially accessible forests 1972ï¿½2002.
FIGURE S1. (A) Location and date of the Landsat and SPOT imagery used to produce the '2002' land-cover map; (B) Location of the Landsat and SPOT imagery used to produce the '2002' land-cover map.
FIGURE S2. An illustrated example of the classification process.
FIGURE S3. Change detection process.
FIGURE S4. (A) Population, (B) timber, and (C) oil palm exports in PNG between 1971 and 2002.
FIGURE S5. Flight paths for the low altitude aerial survey used in the validation of the classification.
FIGURE S6. An example of upper montane forest loss on Mt Kubor (3969 m), Western Highlands Province.
APPENDIX S1. Supplementary methods, including details of the change detection process, and the calculation of change rates.
APPENDIX S2. Discussion of the FAO change rate for PNG.
Please note: Wiley-Blackwell are not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.
Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.