Comparison of Forest cover and net change area estimates for selected case studies in the Congo Basin region
Sannier, Christophe1; Fichet, Louis-Vincent1; Ramminger, Gernot2; Ott, Hannes2
1SIRS, FRANCE; 2GAF, GERMANY
Deforestation is currently known to account for about 15% of global greenhouse gas (GHG) emissions. Therefore, a significant decrease in deforestation can have a direct positive impact on reducing GHG emissions. The Reduction of Emissions from Deforestation and Degradation (REDD) initiative aims to compensate countries that take specific actions to reduce deforestation. The Congo Basin represents the second largest forest area in the world after the Amazon. Deforestation in Congo Basin countries is generally expected to be low. The assessment of forest cover and forest cover change area is essential for REDD+ to determine what is referred to as activity data in the Intergovernmental Panel on Climate Change (IPCC) 2006 guidelines on the Agriculture, Forestry and Other Land Use (AFOLU) sector. A probability sample combined with an appropriate response design can provide forest cover and net change area estimates and their associated uncertainty in the form of confidence interval at a set probability threshold as required in the IPCC 2006 guidelines and for reporting to the United Nations Framework Convention on Climate Change (UNFCCC). However, wall to wall mapping is often required by countries to provide an exhaustive assessment of their forest resources and as input to land use plans for management purposes. Such an approach was tested and implemented in three countries of the Congo Basin regions as part of the European Space Agency (ESA) Global Monitoring for the Environment and Security Service Element (GSE) on Forest Monitoring (FM) REDD project for Gabon and the European Commission (EC) seventh Framework Programme on research and development (FP7) REDDAF project for Cameroon and Central African Republic (CAR). Both projects are coordinated by GAF AG, Munich Germany. One country, Gabon, was covered completely, whilst for Cameroon the Centre region was covered and for the CAR the Ombella Mpoko, Lobaye, Bangui and Sangha Mbaere provinces were covered representing an area of approximately 70,000 kmē for each of the two countries. A similar methodology was adopted to produce Forest cover maps for the three study area to provide the highest level of accuracy and overcome the issue of heavy cloud cover prevalent in the Congo basin region. As an example in Gabon, up to 9 Landsat images were required to cover the area represented by a single scene. Forest cover and cover change maps were produced based on a combination of automated object based classification and manual enhancements to ensure the highest possible level of accuracy. The initial forest cover map was produced for the year 2000 in Gabon and 2010 for Cameroon and CAR. Forest change maps were constructed using the initial forest cover map as a stratification layer. The forest and non-forest strata were re-classified based on the 1990 and 2010 imagery for Gabon (1990 and 2000 for Cameroon and CAR). This was done to ensure that artifacts due to slight differences in the geometry of the same objects for each period were not introduced in the change detection process potentially leading to false changes. The forest cover maps for the remaining two periods result from the union of the existing forest cover map and the corresponding forest cover change map. A 1% sampling fraction for each study area was adopted as this was expected to lead to a coefficient of variation of forest cover area estimates below 10%. This was achieved by selecting 2 by 2 km square Primary Sampling Units (PSUs) randomly from a 20 by 20 km grid overlaid over the study area. This combination of a systematic and random approach ensures that the entire study area is covered and avoids the drawback of a pure systematic approach. This resulted in a total of 665, 175 and 173 PSUs respectively for Gabon, Cameroon and CAR. All the PSUs were visually interpreted using all available imagery and ancillary data by a team of photo-interpreters independent from the production team. Secondary Sampling Units (SSUs) were extracted from each PSU based on a random selection of 50 points. Paired observations were extracted from the forest cover map and the reference data for SSUs to construct error or confusion matrices. Overall accuracies above 90% were achieved and omission and commission errors are mostly below 10% for all three study areas. Area estimates were derived from the PSUs in a so-called direct expansion estimate: This can provide a rapid and reliable means to derive forest cover area estimates based on a sample approach when the production of wall to wall forest cover maps can take some time particularly in countries where cloud cover is significant. However, the direct expansion area estimate can suffer from sampling error resulting in relatively large confidence interval. On the other hand, wall to wall mapping exercise does not suffer from sampling but from potential misclassification error. A Model Assisted Regression (MAR) estimator is based on the combination of both reference and map data and the resulting area estimate is more accurate than either the direct expansion or the area measured from the forest cover map alone. In addition, as for the direct expansion, it provides an estimation of the precision of the estimate which is not available from the map statistics alone. The results obtained confirm the low level of deforestation expected in Congo basin countries. It also confirms the high level of forest cover in Gabon with more than 88% of the country covered by forest covering an area of just over 23.5 million hectares. In Cameroon and CAR, forest represents about 72% of the area of the regions selected with a total of over 5 million hectares. Coefficient of variation at 95% confidence interval for the direct expansion method represent just over 2% for Gabon at national level and just below 6% for Cameroon and CAR at regional level. This is reduced by a factor of 6 to 8 for Gabon at national level and 3 to 4 for Cameroon and CAR at regional level when applying the MAR estimator with coefficient of variation below 0.3% at national level and between 1.4 and 1.8% at regional level thus reducing significantly the level of uncertainty for forest cover area estimates. Deforestation rates are generally low, with less than 0.4% between 1990 and 2000 in Gabon. In CAR, the deforestation rate is about 1.5% between 199 0 and 2000 and 0.8% between 2000 and 2010. However, the deforestation in Gabon is not statistically different from 0 between 2000 and 2010. The same is observed for the Centre region of Cameroon. This is because the changes detected are very small and as a result the coefficients of variations of change estimates are larger. This can be resolved by applying an appropriate stratification of the study area aimed at better targeting areas of potential change. In conclusion, it is worth noting that out of the three study area, two show that there is no significant level of forest change in recent years.