more complete (Morton et al., 2008; van der Werf et al., 2009b). In contrast, afforestation causes a small long-term CO2 sink because woody biomass and soil organic matter build up slowly on the site for decades to centuries. Forested land that is harvested (e.g., forest degradation) and then allowed to regenerate produces a large and short-lived CO2 source, followed by a small and long-lived CO2 sink.
Table 3.1 summarizes the current levels of uncertainty for estimates of forest area, carbon stocks, and AFOLU emissions by remote sensing and the improvements in accuracy that could be accomplished using remote sensing and ground monitoring within several years. The most important messages in the table are the following:
The area of land that is deforested within a country can currently be assessed with an uncertainty of less than 50 percent over a period of 5 years (e.g., Hansen et al., 2008), but this could be improved to 10-25 percent annually by expanding efforts to analyze satellite imagery.
Uncertainty in remote sensing-based estimates of annual carbon fluxes from deforestation, reforestation, and forest degradation is now high (25-100 percent), but could be reduced to 10-25 percent by integrating remote sensing observations with new ground observations of forest biomass and peatland soil organic matter and biogeochemical models for spatial estimates.
The area of flooded soils that emit CH4 (rice, wetlands) can be determined each year with a relatively low uncertainty (~10 percent), although uncertainty in emissions is high (50-100 percent). Ground survey data on specific management practices would provide the greatest reduction in uncertainty.
Uncertainty in annual N2O emissions from managed soils is high (about 50 percent) for the best current inventory methods, and even higher (>100 percent) for developing countries. More flux measurements for different nitrogen management practices and agronomic systems, coupled with improvements in process-based models, would be the most effective action to reduce uncertainties. More information on trends in fertilizer consumption would provide insight about changing emissions when combined with information on management practices, crop production, and weather.
The next five sections describe the methods and studies behind these conclusions.
Although a variety of satellite and aircraft sensors have been used to map land cover and land use (e.g., see Table 3.2), Landsat is widely used and offers several advantages for regional- and country-level greenhouse gas inventory applications. First, the pixel resolution (30 m) is high enough to distinguish most plant cover characteristics, while providing a sufficiently large area of coverage per scene to allow regular global coverage to be practical. Second, the 16-day repeat cycle allows the seasonal information needed for classification purposes, such as greening and browning of grasslands, to be obtained in areas with little or no cloud cover. Third, Landsat data provide a time series of observations beginning in 1984, although the older data have some significant spatial gaps. A global map of land cover in 1990, when coverage is most complete, could serve as a baseline from which to identify subsequent changes (see Figure 3.1 for a Landsat map of the United States using data from the early to mid-1990s).
Forests can be distinguished from nonforests at accuracies of 80-95 percent using Landsat-type imagery (Table 3.1; Lu et al., 2007; GOFC-GOLD, 2008). Accuracies can be validated with in situ observations or very high resolution aircraft or satellite data at a statistically defensible subsample of locations (e.g., systematic or stratified sampling using Ikonos satellite data). The accuracy of remote sensing-derived land cover maps can be assessed in a number of ways, including comparisons with independent data collected in ground surveys. Errors are divided into those due to omission (exclusion of an area from a category in which it belongs) and commission (inclusion of an area in a category to which it does not belong), and categorized by land cover class and region. All approaches are based on statistical sampling. Land cover change also includes errors in geolocation and other factors.