The first determination needed for Article 3.3-identification of ARD land-requires classification of land as forest or non-forest for at least two points in time or an activity reporting mechanism to identify lands that have been subject to ARD activities. Regardless of the method chosen to identify where ARD activities have occurred, some uncertainty is expected to occur. The uncertainty involved in activity reporting is primarily caused by reporting errors or biases (underreporting or overreporting). In the land-use/land-cover classification case, for example, a remote-sensing identification of forest/non-forest areas may include some inaccuracy in classification, which is also a function of the type of orbital system used. Mantovani and Setzer (1997) compared the number of deforestation polygons using the Advanced Very High Resolution Radiometer (AVHRR) sensor (1.1-km resolution) with the number from the yearly deforestation project conducted by Instituto Nacional de Pesquisas Espaciais (INPE) using the Thematic Mapper (TM) sensor (30-m resolution). The comparative results indicated that only 49-57 percent of the total cases were correctly identified with the low-resolution sensor. This difference in ability to detect deforestation appears to be related to the size of the sites considered. As observed in the TM images, deforestation polygons less than 3.1 km2 in size were usually not detected in the AVHRR imagery.
In addition, depending on the season in which the imagery is taken, forested areas can be confused with other vegetated areas in classifications. Areas in which forest cover may be sparse or very young may meet certain definitions of forest yet may appear in the image to be in a non-forest condition. Here, using definitions of forest that are based on low canopy cover thresholds or that provide for classification of early growth stages likely will result in more misclassifications (of forest as non-forest) than definitions that require high percentage cover. Ground-based inventories are more accurate in this respect than inventories that are based on remotely sensed data.
Different definitions for forest, afforestation, reforestation, and deforestation will lead to different levels of difficulty of data collection and different levels of uncertainty. Definitions that are based on objective and readily measurable variables will tend to yield lower uncertainties. For example, a forest definition that is based completely on the proportion of land covered by woody species (e.g., Land Cover scenario) will lead to more accurate estimates than a definition that involves purpose or history of land use or functional mechanisms used for establishing forests (e.g., Land Use scenario).
The uncertainty of stock change estimates will also vary with the magnitude of the stock in terms of spatial extent and density. Detecting large relative changes in a small stock is easier than detecting small relative changes in a large stock. This fact implies that countries with higher carbon density in forests will have a greater uncertainty for a given magnitude of stock change than countries with lower forest carbon density. Furthermore, changing the threshold of forest cover in a definition may influence the uncertainty in detecting ARD activities.
Because ARD activities are change processes, definitions and methods have to be consistent on successive occasions. In the framework of the Kyoto Protocol, it is convenient to group the scenarios presented in Section 3.2 in two broad classes:
If a new data collection system has to be implemented, obtaining the necessary retrospective information will be difficult regardless of the scenario applied because, at a minimum, data are required on the status of lands in 1990. Use of archived remotely sensed data, inventory plots that may have been established for another purpose, and/or retrospective models may be the only available alternatives. Evaluating the uncertainty or verifying such estimates would be impossible.
Uncertainties in determining stock changes from ARD activities will be sensitive to various components of the data collection process. A crucial component is the sampling scheme used for detection of ARD activities, which will determine the minimum detectable change in forest condition. Difficulties arise if the sampling resolution is inconsistent with the assessment unit sizes in the definition adopted for forest. If a ground-based forest inventory is used to sample ARD activities, the sampling intensity-expressed as hectares of land represented by each plot-will give an indication of the ability of the system to detect ARD activities on small land areas. For assessments that are based on remotely sensed data, the spatial resolution of the sensor used will be crucial in determining the minimum detectable change.
Consistency in sampling schemes between assessments is also important: Erroneous conclusions can be reached if sampling intensity (or sensor resolution) changes substantially between assessments. For example, suppose that 1-km resolution data were used for a first assessment, and a 1-km2 area was identified as forest. At a second assessment, 100-m resolution data might detect non-forest areas within the original 1-km2 area, leading to an interpretation of deforestation. That 100-m opening may have existed within the 1-km2 area during the first assessment, however, in which case the determination of deforestation would be in error.
Timing of measurements will also affect the ability to detect stock changes from ARD activities. Forest inventories can be time-consuming, multi-year processes, particularly in large areas. In the United States and many European countries, for instance, the national forest inventory cycle is about 10 years. With this type of measurement timing, it is impractical to expect accurate, verifiable estimates of stock changes during a 5-year commitment period and for a small but comprehensive subset of the inventory area. Some nations are beginning to use annual inventory systems, but the transition from periodic national inventories to annual inventories is costly and difficult. In some cases, new sampling protocols (timing and intensity of samples) may need to be adopted if national inventory data are to be useful for the carbon accounting required under Article 3.3.
Although the focus here has been on implications for the first commitment period, there are considerations for subsequent commitment periods. Errors in classification of forest/non-forest can have effects in subsequent commitment periods, either in the determination of areas requiring stock estimates for Article 3.3 or in assessments of stock changes between the first commitment period and later commitment periods. For example, consider a parcel that was forested in 1990 and was subsequently misclassified as non-forest after 1990. This misclassification could lead to an erroneous classification of deforestation and carbon debits for the assessment of the first commitment period. The error depends on the definitions applied and the method used to determine carbon stock changes. If the method involves some form of carbon stock measurement, the misclassification of ARD land will have no impact on the reported carbon stock change because there will be none in the misclassified area. A subsequent (correct) classification of the same area as forested would produce a second faulty determination-this time of reforestation and creditable carbon stock increase. Similarly, errors in estimates of stock changes can have persistent effects after the first commitment period. No matter what sampling scheme is employed, these types of errors will occur. The key issue is the identification of systematic problems that will generate biased estimates of ARD activities and random errors that offset one another, having little effect on ARD activities determinations.
Other reports in this collection