Mapping global land or forest cover, even for a limited number of classes, is a difficult endeavor, as spectral data do not follow a single set of rules for pre-defined classes and problems associated with atmosphere and view/illumination geometry in large areas complicate matters still further. In addition, fine-scale land cover conditions, such as canopy density or forest fragmentation, often dictate spectral response to the sensor of coarse spatial resolution. For these problems, which are unique to large-area land cover mapping, this project relied on geographic stratification, mixture modeling, and other measures employed by other similar studies (e.g., Cihlar et al. 1996, Loveland et al. 1999, Malingreau et al. 1995, Roy et al. 1997, Zhu and Evans, 1994). The two global-scale maps and results of the validation effort produced for the FAO FRA2000 Program showed that this integrated approach worked reasonably well for the objectives of this study.
The FAO FRA2000 forest cover map, as validated by using the IGBP global validation data set, has an overall accuracy at 77 percent (standard error 2.4 percent) for the first four classes. Using the same validation data set, the forest and nonforest classification of the global forest cover map has an overall accuracy of 86 percent with a standard error of 2 percent. The relatively low per-class accuracy for open or fragmented forest may be partially attributed to the mixture model’s difficulty to faithfully and consistently map low canopy cover density mixed with large amounts of other land cover “background” such as bright soil and agriculture. On the other hand, the incompatibility of the existing reference information used in the validation probably had an adverse effect on accuracy estimates for the map. The reality that assessing accuracy for any large-area land cover map is necessary but costly led to the compromise in this study between cost and precision. To provide a sense of overall quality of the mapping effort to users, suitable existing reference data were used for three geographic areas: global, China, and the U.S. As a result, our validation is only relevant at these levels and reliability at scales finer than sub-continental is unknown.
Experiences learned from this project indicate that several key factors were responsible for the overall quality and accuracy of the global product; these factors include 1) knowledge and experience of analysts and availability of reference information, 2) quality of input AVHRR imagery, and 3) the flexible use of compositing, stratification, scaling, and modeling rules to account for data and vegetation variations. The modified spectral mixture analysis was necessary to produce an estimated percent forest canopy cover data set, based on which the first two classes of the FAO forest cover map were summarized. The other three classes were refined and adopted from the USGS global land cover database even though there is a 2-3 year difference of dates of input data. For the definitions of the three FAO classes and at the global scale, it may be safely rationalized that any changes between the three classes can be insignificant.
While the global forest cover map was developed primarily for use by the FAO FRA2000 program, there are potentially other uses for both digital maps. Particularly for global change studies, the two maps represent the distribution and conditions of large forest cover areas in the world as of 1995. The maps may also be valuable inputs into other climatic, ecological, and forest models that require large-area forest cover data. As an update to the USGS global land cover characterization database, the forest cover mapping effort enriches the diverse content of the database and makes it more useful to a wider range of applications.
Improvements in satellite sensor and imagery will be potentially the most important contribution to future global forest cover mapping. Accuracy, content, and timelines of global forest cover mapping can be significantly improved by using EOS (Earth Observation System of the National Aeronautical and Space Administration) era sensors such as MODIS and Landsat 7 that have better atmospheric, radiometric, and geometric qualities. Programs such as GOFC (Global Observation of Forest Cover) have been designed to take advantage of new sensors and their much improved mapping capabilities.
