Missed carbon emissions from forests: comparing countries’ estimates submitted to UNFCCC to biophysical estimates

In the UNFCCC Submissions, the countries selected which of the five REDD+ activities were to be undertaken (‘scope’), defined what constitutes a forest in terms of minimum canopy cover, tree height and area size, and established whether there are land uses that include trees that are not considered as forests (e.g. plantations) (table 1).

All countries within our selection, except Malaysia, included ‘reducing deforestation’ in their scope for REDD+. Malaysia elected instead to consider ‘sustainable management of forests’ only, which generally refers to the adaptation of methods to reduce the impact of timber harvesting practices (GOFC-GOLD 2013). Another noteworthy difference between the countries is that some included plantations in their forest definition. Indonesia excluded all types of plantations, while Cambodia and Sri Lanka excluded rubber plantations. Forest plantations were included by all but Indonesia.

Within these scope of activities and forest definitions, forest area change (referred as ‘Activity Data’ in the UNFCCC submissions) was estimated. The Activity Data includes the amount of forest area change or deforestation during the historical period selected (‘reference period’), which is used as a benchmark for assessing countries’ performance in implementing the selected REDD+ activities. The reference periods among the seven countries varied from eight to 22 years with the earliest starting year of 1990 and the latest of 2006 (table 2). The number of actual data points in the reference period also had a wide range, from 2 to 23. If a country believes that the historical rate does not reflect the likely changes in the future, adjustments can be made with justifications (Government of Papua New Guinea 2017, Ministry of Agriculture and Rural Development, Vietnam 2016).

In generating Activity Data, the commonly used method is wall-to-wall mapping and detecting changes by comparing classified maps (e.g. Cambodia, Indonesia, and Vietnam) (table 2). This method, however, can lead to substantial errors because each map inevitably contains some errors, which will be compounded when comparing two maps to detect changes (FAO 2018). In correcting the effects of classification errors, two countries (e.g. Nepal and Sri Lanka) used a stratified area estimation approach, which distributes a sample of reference data in a stratified manner based on the classes. The disadvantage of this method is that statistically derived area estimates may no longer match with the areas on maps (FAO 2018). Papua New Guinea is the only country that used a systematic sampling method, which is more transparent, as samples are distributed in a non-stratified manner, but it requires a large number of samples to achieve reliable results (FAO 2018). The highest overall accuracy rates in mapping were reported by Indonesia (98%) and Vietnam (95%) and the lowest by Cambodia (74%) and Sri Lanka (75%). Forest gain data had much lower accuracy rates with 68% by Nepal and 9% by Sri Lanka (Ministry of Forests and Soil Conservation, Nepal 2017, Sri Lanka UN-REDD Programme 2017).

Based on the Activity Data, the reference level is calculated with emission factors in tonnes of CO₂ equivalent per hectare per year (‘forest reference emission levels’), or net emissions (‘forest reference levels’, which include removals). From this, payments can be calculated if future monitoring suggests a positive deviation from the reference levels. Therefore, excluding certain activities or the way in which forests are defined affects the reference level, and future carbon payments significantly. For example, a decline or increase of plantations in Indonesia will not affect their performance in emission reductions, but the loss of natural forests will matter greatly; while in Malaysia loss of any forests other than their target production forests is not relevant to potential carbon payments.

Indonesia selected only deforestation in the scope of REDD+, therefore we estimated the forest cover by calculating the forest gain from the GFC v1.4 dataset using the national definition for minimum change area (table 3). Indonesia’s loss for 2010 was estimated by using the average of 2009 and 2011 loss data, as deforestation data were not provided for 2010. Papua New Guinea reported there was no forest gain during the reference period. The loss data for Papua New Guinea were directly estimated from the figure 7.4: Deforestation occurred in PNG 2000–2013 (PNGFA Collect Earth Assessment) (Government of Papua New Guinea 2017). Sri Lanka reported forest loss and gain between 2000 and 2010, but chose not to report the forest areas estimated in constructing the reference level. Therefore, we used the 2010 forest area reported in the Forest Resource Assessment 2015, and applied loss and gain data from the UNFCCC Submission to estimate the forest area for 2000 Forest Resources Assessment Programme (FAO) 2015, Sri Lanka UN-REDD Programme 2017).

A few countries used different minimum area size from their forest definition when detecting the changes: Cambodia used a minimum mapping unit (MMU) of 25 ha for 2006/2010 and 5 ha for 2014 and Nepal used a 2.25 ha MMU in detecting changes in country-defined REDD+ forests, while both countries used 0.5 ha for the minimum forest area (table 3). For Indonesia, Papua New Guinea, and Sri Lanka, we assumed that the countries used the same minimum forest area size to detect changes, which is 0.5 ha for all except Papua New Guinea (1 ha). It should also be noted that the minimum tree height for Papua New Guinea was 3 m, while others were 5 m.

Forest areas for the year 2000 (2006 for Cambodia) were calculated from treecover pixels, which were required to satisfy the minimum canopy cover requirement, and be connected to other pixels with sufficient canopy cover so as to form a patch of forest larger than the minimum area size (figure 1). The contiguity constraint was applied with a country-specific pixel area calculation, and with pixels connected diagonally (queen’s move) included as a single patch of forest.

Zoom In Zoom Out Reset image size

For each year thereafter we recalculated forest areas based on the loss of previously forested pixels (figure 1). Loss was recorded in cases where pixels that previously met each country’s forest definition were identified as a forest loss for each year from 2001–2016. Treecover pixels were still required to meet the minimum forest area or change area condition, thus loss was also recorded in locations where forests became fragmented to the extent that a forest patch was too small to meet this requirement. In these cases, an area of forest loss was only counted where the contiguous area over which a forest disturbance was recorded was larger than the minimum change area specification.

Increases in tree cover are reported by GFC as a total for the period 2001–2012. We therefore calculated the total forest area gain for this period and allocated it uniformly over the measurement period. In a similar manner to losses, forest area increases were subject to minimum forest area as well as a minimum change area requirements, and gains were included in cases where forest patches that previously didn’t meet the minimum area requirement increased in size to meet the minimum area size. In cases where the GFC dataset reported a gain at a location that was already recorded as tree cover, pixel areas were not included as part of the gain area. Where both losses and gains were reported at the same location, the gains were assumed to have occurred following loss, so pixels were included in both gain and loss area accordingly (figure 1).

Country-defined REDD+ forest area is less than biophysical forest area in most countries (figure 3, Nepal and Cambodia are the only exceptions). This is because a biophysical forest definition (based on minimum tree cover percentage, height, and area size) will include trees in non-forest land use areas, such as plantations, agricultural land or settlement areas with trees (table 1, figure 3). This could explain the difference of loss in Indonesia for example, where the proportional difference between the rates of loss broadly corresponds to the differences in the area of forests compared (figures 2 and 3). However, some countries show unexpected results: Sri Lanka has more forest loss in country-defined REDD+ forests than the changes in biophysical forests; and Nepal and Cambodia have larger areas in country-defined REDD+ forests than biophysical forests would predict. In all cases, however, this is likely due to differences in mapping methodology, for which see below.

Zoom In Zoom Out Reset image size

No mapping methods are free from errors (table 2) (Olofsson et al 2013). The GFC dataset’s overall accuracy using the direct detection method are 99.6% and 99.7% for loss and gain respectively, while the countries selected different mapping methods and the resulting overall accuracy varied significantly from 74% to 98%. For example, Nepal and Sri Lanka used a stratified area estimation method and achieved relatively low accuracy rates (table 2). Especially for Sri Lanka, the accuracy rates for loss and gain were 79% and 9% respectively (UNFCCC 2018). Cambodia’s biophysical forest area in 2006 was estimated with tree cover in year 2000, adjusted with gain and loss data using a 5 ha MMU, while the country-defined REDD+ forests were based on wall-to-wall mapping.

The changes in biophysical forest areas using the GFC dataset are blind to the process of change: it is looking at the net change in forest cover over the period, with forest as defined nationally based on canopy cover and minimum area size. However, the dataset includes processes that would under national definitions not be deforestation or reforestation. For example, both the clearance and growth of trees within plantation areas are included in the GFC-based biophysical forests, but not in country-defined REDD+ forests. This likely explains the large difference in change data in Vietnam (figure 2), whose UNFCCC submissions show net forest gain over the 2000–2010 period, when plantations in the country were expanding, while the change in biophysical forests show a large loss as deforestation continued and trees in plantations were harvested (figure 2). This is partially exacerbated by Vietnam’s decision to include plantations with tree crop shorter than 5 m, increasing the rate at which forest gain appears to occur. Malaysia shows a similar difference, with its reported net gain largely due to the exclusion of deforestation in the scope of REDD+ activities, and limiting it solely to production forests, which have increased in area over that decade.

One might assume that the differences in change data caused by the harvesting and replanting of plantations would stabilise with time: if the area harvested each year is the same as the area of plantation that reaches the required canopy cover and height threshold, then the impact of plantations on the net change in biophysical forests will be zero, matching country-defined REDD+ forests data. However, as trees are long lived, even in tropical plantations, and planting tends to happen in spurts of a few years related to national programmes and incentives, it may be that such an annual balance of planting and harvesting never occurs. This is further complicated because detecting forest gain in satellite data is much more challenging than the abrupt change in forest loss: therefore the GFC dataset includes only a single layer for gain, stating that an area became forested at some point in the range 2000–2012, meaning our gain data is smoothed compared to the annual loss data; further the gain from the GFC data only detects gains as occurring when trees reach a 50% minimum canopy cover, higher than the thresholds for loss. All plantations will reach this threshold long before harvest, so again this will not ultimately change the net number, but it may be another reason for differences between GFC-derived change in forest areas and national figures.