4 Results and discussion

The result of the mapping project is a global forest cover map with the FRA2000 legend (figure 2), which portrays the spatial distribution of forested lands in the world as of mid the 1990’s. The modeled forest canopy density (fraction of square km) is an additional data set, an intermediate product as the result of our methodology. The primary function of this unique data set was to provide a quantification of forested land in various regions of the world, and to help define closed forest from open or fragmented forest, and open or fragmented forest from non-forested land. This global map may potentially be useful to other applications. As a quantified output, the data set may be used in large-scale (continental or global) climate models to test for different forest cover scenarios, or in other large-area biogeochemical, ecological, conservation models that require both quantitative and categorical forest cover descriptions. This data set, however, is not validated and is not required for the FRA2000 program.

The global forest cover map will be published as a poster map, in a global Robinson projection, by the FAO FRA2000 program to illustrate distribution and conditions of global forest resources surveyed under the 10-year program. The digital data set will be made available, similar to other USGS large-area land cover products, on the EDC webpage³. The forest cover data set is a new land cover layer in the USGS global land cover characteristics database and complements other thematic layers in the database in holistically describing the global land cover.

Area estimates are derived by classes and by major landmasses (Table 2). Although Antarctica was not contained in the input data, it is included in the table and on the global map as “other land cover” (barren, ice and snow). Affecting the area estimates in a relatively small way are small islands in the world and several tropical island chains (Hawaii, Micronesia and Polynesia). These islands were not included in the mapping effort due to the lack of quality image data consistent with the rest of the world. Compared to the missing islands, a few other factors inherent to the AVHRR source data probably exerted far greater impact on precision of the area estimates. These factors include image quality problems such as atmospheric contamination and view/illumination geometry, misclassification errors, issues related to spatial scale including resolution and area distortion. It is important to recognize that the area estimates in Table 2 (and mapped forest cover patterns) are imprecise, and they should be treated in the context of continental and global scales, instead of local significance.

Table 2. Area estimates (rounded to thousands square kilometers except for global percent) of FAO FRA2000 classes by global total and major landmasses. The five landmasses are Africa (AF), Australia and tropical Asia (AP), Europe and temperate/subtropical Asia (EA), North America (NA), and South America (SA). The area estimates are derived from continental Lambert equal-area projections.

⁴This figure includes total area of Antarctica given by the National Geographic world atlas, sixth edition.

The definitions of the FAO forest classes were developed based on ecological characteristics of broad regional forests in the world, rather than from any remote sensing studies. There exists a certain correspondence between forest canopy cover and satellite image data, as shown in this mapping exercise. However, such correlation between satellite spectral data and percent forest cover is not unique and variations are great in different geographic regions and vegetation types. Forest lands of different canopy cover are mapped based firstly on ecological knowledge of these forests. For example, northern boreal forests in subpolar zones or parts of the miombo woodland in Africa are known to have canopy densities below 40 percent. The use of models and geographic stratification only provided refinements to the canopy openness.

Our strategy generally assumes that forests tend to occupy a specific region in spectral space and that regionally specific compositing and modeling will ensure the separation of forest and non-forest. Unfortunately, even with an optimal strategy for enhancing discrimination, there are unavoidable ambiguities; these range from spectral ambiguities in land cover, to ambiguities resulting from atmospheric, illumination, and view angle effects. The former can include land cover that is physically similar to forest but doesn’t meet the forest height definition (e.g. Mediterranean shrubland and heathland of Scotland). Other land cover may appear similar to forest due to a mixed pixel effect. Non-woody wetlands, and other vegetation/water mixes, vegetation/dark soil mixes, and vegetation/burn mixes can have the spectral appearance of forest.

Similarly, cloud shadow, low sun angles, and topography are factors that can cause non-forest vegetation to look forested. However, most atmospheric and view angle effects have the tendency to make forests brighter, and thus more like non-forests; these include atmospheric effects (haze or subpixel clouds) that are especially prevalent in the tropics. Even year long compositing can fail to remove some of these problems, especially over some cloud forests. Similarly, backscatter viewing geometry can make forests look unforested, but compositing reduces these problems. Extreme cases, such as eastern North America, can be ameliorated with stratification, although view angle-specific corrections might be preferred.

As noted before, the global IGBP validation data set (primarily TM interpretation sites) was used as an independent reference data set to assess the accuracy of the global FAO forest cover map. To accomplish this goal, it was necessary to approximate the FAO classes with the IGBP legend. Table 3 shows the results of the accuracy assessment. The accuracy and associated standard error estimates were derived based on simple random sampling (Stehman and Czaplewski, 1998).

Table 3. The error matrix, user’s and producer’s percent accuracy estimates using the IGBP global validation data set. The IGBP legend was re-labeled to approximate the FAO legend (the first four classes). The overall accuracy and standard error estimates are 76.92% and 0.24%, respectively.

Not surprisingly, FAO classes 2 and 3 (open or fragmented forest, other wooded land) have low accuracy than class 1 (closed forest) and class 4 (other land cover). While closed forest and other land cover (e.g., cropland, grassland) are relatively easy to define, the exact boundaries in terms of canopy cover between open or fragmented forest and either the closed forest or other land cover are spectrally highly variable. Particularly at the lower forest canopy cover, the amount of open ground with bright soil tends to contribute to misclassification between open or fragmented forest and other land cover such as barren or shifting agriculture. Similarly for the other wooded land (class 3, primarily woody savanna and shrubland), where this type of woody cover ends and other land cover types (e.g., grassland, forestland or barren) begin is highly subjective. There is also the factor of incompatibility between the IGBP legend (basis of the IGBP global validation data set) and the FAO legend (see Appendix 1). For example, IGBP class 14 (cropland and natural vegetation mosaic) was re-coded to other land cover (FAO class 4), even though the original class clearly represented the situation where patchy forest land or shrub land could be present. The compromise of recoding is a potential source for artificially low accuracy, particularly for FAO classes 2 and 3.

Accuracy estimates, for the first four classes and for forest/nonforest aggregations, were also calculated for each of six continents: Africa (AF), Australia and New Zealand (AU), Asia (AS), Europe (EU), North America (NA), and South America (SA). Sample sizes for the continents varied between 13 (Australia and New Zealand) and 82 (North America). This variation largely reflected the diversity of IGBP land cover in the original random stratified sampling method (Scepan, 1999). Because of limited sample sizes for the continents, per-class accuracy and standard error estimates would be imprecise, and therefore were not calculated. Only global and continental overall accuracy estimates are given in Table 4.

Table 4. Sample size and overall accuracy of the first four FAO classes and forest/nonforest aggregations for global and six mapped continents: Africa (AF), Australia and New Zealand (AU), Asia (AS), Europe (EU), North America (NA), and South America (SA).

	Global	AF	AU	AS	EU	NA	SA
Sample size	312	69	13	68	43	82	37
Classes 1-4	76.92	68.12	92.31	82.35	72.09	79.27	78.38
Standard error	0.24	0.57	0.77	0.47	0.69	0.45	0.69
Forest/nonforest	85.58	76.47	92.31	85.29	86.05	86.59	97.30
Standard error	0.20	0.51	0.77	0.44	0.54	0.37	0.26

Notwithstanding the factor of limited sample sizes, accuracy estimates at the continental level clearly show that land cover homogeneity is an important factor in determining map accuracy. For forest and nonforest land cover classification, the FRA2000 map is most accurate in South America, largely due to the continuous, unbroken tropical moist forest in Amazon Basin; and it is least accurate in Africa, as Congo Basin rain forest is smaller than that of South America, and the large areas of miombo woodland in southern Africa are spatially and temporally highly variable. Reasonable accuracy results in the northern hemisphere perhaps reflect the fact that land cover is relatively established and knowledge of land cover in the regions can be easily obtained. For Australia and tropical Asia, the high accuracy may be the result of the small sample size (13 sample sites). Of course, regional land cover within the continents, and associated map accuracy, are highly variable spatially. An ideal and more complete assessment would be to obtain sufficient sample sizes to capture the regional variations.

China and the U.S. are two special cases (Tables 5 and 6) in the validation because of the availability of wall-to-wall land cover mapping with 30-meter Landsat data in the two countries. Perceptively, the all-pixel comparisons should be statistically more rigorous than sampling approaches for the same mapping areas. However, any gain in statistical precision may be discounted by the facts that the fine-resolution land cover maps have their own mapping errors⁴, and that the two countries probably represent the best case scenarios in the mapping with ample ancillary information and researcher familiarities. On the other hand, none of the three reference data sets were developed for the mapping effort, resulting in a mismatch with the FAO legend, despite the use of “cross-walk” to make these reference data sets as comparable as possible with the FAO legend. The legend mismatches should be included in consideration of overall quality of the FRA2000 map.

Figure 3 is a RGB (red, green, blue) color composite image for the conterminous U.S., with the FRA2000 forest canopy map displayed as red and MRLC-derived forest canopy map displayed as green and blue. The MRLC forest canopy map was derived by recoding the three forested classes (see Appendix 1 for MRLC classes) to forest and the rest of the classes to nonforest, and then aggregating 30-meter pixels to corresponding 1 km pixels. Areas of agreement between the two maps are shown in shades of gray (from black to white depending on canopy density value estimated). Areas of disagreement are shown in colors of red (canopy density estimates of the FRA2000 map are greater than that of the MRLC map) and cyan (vise versa). The eastern U.S. woody wetlands (which should have been included in the MRLC map as forested land cover) and the western chaparral (classified by MRLC as shrubland, but considered forested by the U.S. Forest Service) accounted for most of the FRA2000 overestimation (red shades). Conversely, the FRA2000 underestimation (cyan) occurs primarily in three situations: 1) the Great Basin (Nevada and Utah) dry woodland and shrubland, 2) patchy or riparian woodland in the Midwest agriculture region, and 3) southeastern U.S. fragmented forest lands. While the U.S. MRLC is a special case, the differences in Figure 3 probably represent typical errors or variations in the FRA2000 global forest canopy density map.

Table 5. Area estimates (rounded to thousand km²) and rate of forest cover for China and the conterminous U.S. FAO estimates are derived from 1-km AVHRR data, CAS and MRLC area estimates are derived from 30-meter Landsat data.

Table 6. Estimates comparing the FAO map to reference maps of China (CAS) and the conterminous U.S. (MRLC). Overall agreement for forest/nonforest classification is 86.12 and 83.01 percent, respectively for China and the U.S.

Certainly, the ultimate judgement of a map’s overall quality lies with its usefulness and effectiveness in meeting requirements of a variety of applications. User feedback can be an opportunity to examine potential areas of map error, whereas sensitivity of applications using the map as input may be a more objective indicator of the map’s accuracy. However, these holistic approaches are long term and difficult to measure in absolute numbers; traditional accuracy assessment is still necessary to provide a more immediate assessment to both producer and user. In this study, the use of existing reference data sets for accuracy assessment may not be the ideal approach, but it was a practical decision to obtain an accuracy estimation given time and budget constraints.

³ http://edc.usgs.gov

⁴ Accuracy assessment for either map product is not complete at this time.