2.1. Per-pixel mixture analysis

Our methodology is centered on an effort to break individual satellite pixels into constituent cover fractions of surface materials. To do so, we employ a general, probabilistic spectral mixture model for decomposing satellite spectral reflectance measurements into subpixel estimates of PV, NPV, and bare substrate covers (Figure 1). This model is based on an algorithm developed for forest, savanna, woodland, and shrubland ecosystems (Asner and Lobell 2000; Asner and Heidebrecht 2002; Asner et al. 2004a). It is fully automated and uses a Monte Carlo Unmixing (MCU) approach to derive uncertainty estimates of the subpixel cover fraction values. The Carnegie AutoMCU uses three spectral end-member “bundles” derived from field measurements and satellite imagery to decompose each image pixel using the following linear equation:

View Expanded

where ρ(λ)_e is the reflectance of each land-cover end-member (e) at wavelength λ, and ε is an error term. Solving for the subpixel cover fractions (C_e) requires that the observations [ρ(λ)_pixel in this case, Landsat ETM+ reflectance] contain enough information to solve a set of linear equations, each of the form in Equation (1) but at a different wavelength (λ).

Until recently, there were a limited number of spectral signatures of green and senescent vegetation and bare substrates for areas in the Amazon and cerrado. Our mixture modeling technique requires spectral reflectance bundles [ρ_pv(λ),ρ_npv(λ),ρ_bare(λ)] that encompass the common variation in canopy and soil properties. Asner (Asner 1998) and Asner et al. (Asner et al. 2003a; Asner et al. 2004a) collected these spectral data using full optical range field spectroradiometers (Analytical Spectral Devices, Inc., Boulder, Colorado) during field campaigns conducted from 1996 to 2000 (Figure 2). The spectral end-member database encompasses the common variation in materials found throughout Brazilian cerrado and Amazon ecosystems, with statistical variability well defined and deemed viable for end-member bundling (Asner et al. 2004a). The bare substrate spectra have been collected across a diverse range of soil types, surface organic matter levels, and moisture conditions. Spectral collections for NPV have included surface litter, senescent grasslands, deforestation residues (slash), and other dry-carbon constituents from a wide range of species and decomposition stages.

In contrast to the NPV and bare substrate spectra, which can be collected via ground-based spectroscopic measurements, the PV spectra of forest and savanna woody plant species require overhead viewing conditions. This is very difficult to achieve in woody canopies with heights ranging from 5 to 55 m. Spectral measurements of individual leaves, stacks of foliage, or partial canopies (e.g., branches) introduce major errors in spectral mixture models and cannot be used (Asner 1998). Therefore, we collected canopy spectra using the Earth Observing-1 (EO-1) Hyperion sensor. Hyperion is the first spaceborne imaging spectrometer for environmental applications (Ungar et al. 2003). We used Hyperion data from forest and woodland control sites established in 1999. The data were derived from more than 40 000 spectral observations made at 30-m spatial resolution (G. P. Asner 2005, unpublished manuscript), atmospherically corrected to apparent top-of-canopy reflectance using the Atmospheric Correction Now (ACORN) algorithm for hyperspectral data (ImSpec, Inc., Pasadena, California), and convolved to six Landsat ETM+ optical channels. These green vegetation spectra thus inherently included the variable effects of intra- and intercrown shadowing, which are prevalent in tropical forests (Gastellu-Etchegorry et al. 1999).

In Amazonia, shade fractions average 25% cover in humid tropical forests, but the variance is high with standard deviations of 12% or more (Asner and Warner 2003). In cerrado savannas, mean shade fractions range from 0% to 16%, with standard deviations up to 5% (Asner and Warner 2003). It is thus critically important to note that our PV results include shade, which varies substantially with forest structure. Using a separate shade end member is attractive (e.g., Souza and Barreto 2000), but doing so with multispectral Landsat data and such high shadow fraction variability often results in an underdetermined spectral and mathematical problem in linear mixture models. Imaging spectroscopy (hyperspectral) data are needed to solve this problem (Roberts et al. 1993). For now, we averted this issue by accepting the limitations of incorporating variable shade directly into our PV bundle derived from the EO-1 Hyperion sampling of woody plant canopies in Brazil. In the end, the total number of spectra retained in the end-member bundles for the AutoMCU mixture model was 252, 611, and 434 for PV, NPV, and bare substrate, respectively (Figure 2). These spectra represent the mean of more than 230 000 field and spaceborne spectrometer observations collected over a 5-yr period of study.

2.2. The Carnegie Landsat Analysis System

The AutoMCU algorithm has been successfully employed on subsets of Landsat imagery for studies of deforestation, pasture condition, and selective logging (Asner et al. 2003a; Asner et al. 2004a). The automated features of the approach make it conducive to larger-scale application, but to do so requires many additional methodological advances. We sought to achieve nearly complete automation of the AutoMCU approach by embedding it in a processing system designed to ingest large amounts of Landsat data, to convert the data to radiance and top-of-canopy reflectance via atmospheric correction, to mask water bodies and atmospheric perturbations such as clouds and haze, and to avoid areas containing cloud shadows. In going to a very large operational scale, such as for the entire Brazilian Legal Amazon and bordering cerrado regions, the system must also allow for normalization of images via rescaling and for image edge detection and mosaicking. These criteria resulted in the Carnegie Landsat Analysis System (CLAS) presented in Figure 3.

Raw Landsat ETM+ data are geocorrected, then sensor gains and offsets are used to convert from a digital number (DN) to exoatmospheric radiance. The radiance data are passed to a fully automated version of the 6S atmospheric radiative transfer model (Vermote et al. 1997). The 6S program is integrated into the CLAS processing stream and uses average monthly aerosol optical thickness (AOT) and water vapor (WV) values from the NASA MODIS sensor. Time stamping of MODIS AOT and WV with Landsat data is done on an automated basis (Figure 3).

A series of masks are designed to exclude clouds, water bodies, cloud shadows, and nonimage areas from the analysis. These masks are derived directly from the raw and calibrated Landsat imagery as well as from the AutoMCU submodel output of CLAS. Clouds are masked using the thermal channel (band 6) from the raw Landsat images (Figure 3). Following a series of tests, a thermal band threshold DN value of 125 was found to conservatively detect cloudy pixels over Amazonia. Water bodies are masked by finding pixels in the calibrated Landsat reflectance data in which bands 1–4 (blue to near infrared) displayed a negative slope. Only water displays such a negative spectral slope with increasing wavelength. Nonimage areas containing zero values are masked. Cloud shadows are masked using an empirically derived combination of the Landsat ETM+ reflectance values and the AutoMCU root-mean-square (rms) error calculations (from Figure 1).

Following the mask computation for each image, the AutoMCU submodel described above is run using the formulation and spectral end-member bundles shown in Figures 1–2. This code spectrally decomposes each image pixel into fractional cover estimates (0%–100% cover). Final steps include image rescaling and mosaicking. Rescaling of the data is necessary to equalize very small (< 5%) differences in fractional cover results that occur due to image calibration errors, residual atmospheric constituents (e.g., cirrus clouds and light smoke), and other random artifacts in the processing. This step produces a contiguous dataset that could then be readily mosaicked together for large-scale production, visualization, and study.

2.3. Proof of concept and capability

We tested CLAS on 197 Landsat ETM+ images compiled by the EarthSat Corporation and provided via the NASA Data Buy to the University of Maryland. These images were provided in orthogeorectified format as raw digital numbers at 28.5-m spatial resolution. The multiyear compositing of this mosaic results in a limited applicability to most basinwide studies of gross deforestation, selective logging, or agricultural expansion that require annual observations. However, this mosaic did provide an avenue to test our automated system and to study the structural typologies of Amazon and cerrado ecosystems, as discussed later.

The Landsat image dataset contained 10.87 × 10⁹ pixels, spanning the years 1999–2001 as shown in Figure 4. CLAS was used to process this enormous dataset on a Linux computer cluster utilizing distributed computational techniques on multiple processing nodes. Each image is linearly divided into smaller datasets, which are passed to each processing node along with the spectral unmixing code. Following computation on each node, the data are sent back to a central processing unit to reconstruct the final results in image format.

2.4. Sensitivity to atmospheric correction

A critical aspect of any large-scale, optical remote sensing study of vegetation lies in the assessment of potential sources of error. Atmospheric correction of optical satellite data has been a major issue in many studies of the humid Tropics, owing to the often high water vapor and aerosol (from biomass burning) concentrations found over these regions (Kaufman et al. 1998). To determine the effects of aerosols and water vapor on the fraction images produced by CLAS, the 6S atmospheric radiative transfer model was iteratively run with a range of AOT and water vapor values on a “reference image” (from Asner et al. 2004a), with all other atmospheric parameters remaining fixed. The reference image was corrected with the 6S code using a “true” AOT and water vapor values of 0.086 and 2.15 cm, respectively. Noise was then introduced using value ranges observed throughout the MODIS AOT and water vapor products from 2000 to 2002 over Amazônia. By running these artificially noisy scenes through the AutoMCU submodel and comparing the resulting PV, NPV, and bare substrate fraction images, the effects of variable amounts of aerosols and water vapor were compared to the reference image output.

2.5. Analysis of ecosystem typologies

A valuable use of the multiyear mosaic lies in defining the structural typology—as quantified via fractional cover in Landsat imagery—of the many ecosystem types found throughout Amazônia and in bordering areas of the cerrado. To our knowledge, such an effort has never been attempted using any high-resolution satellite imagery. Fractional canopy cover analysis provides a means to define an aspect of ecosystem structure that bears on studies of forest growth, degradation, land use, carbon cycling, and a wide range of ecological processes from flora to faunal dynamics and food web structure. Basic statistics and intercomparisons of the major vegetation types and ecosystems found throughout the region provide a biophysically based starting point against which to monitor changes in the future.

We used the vegetation-ecosystem map developed by the Brazilian mapping and statistics agency [(Instituto Brasileiro de Geografia e Estatística) IBGE 1993] to partition the fractional cover observations into recognizable classes (Figure 5). We compared the fractional material cover of PV (usually green, woody plant canopies), NPV (usually herbaceous and surface litter), and bare substrate (usually soils) between IBGE classes, by Brazilian state, and for the entire coverage area (see Figure 4). To develop the baseline ecosystem typologies without significant imprints of land use, we only compiled data with PV values greater than 70%. This threshold was experimentally derived based on analyses of images throughout the region and from Asner et al. (Asner et al. 2004a). Areas of deforestation, including pasture systems, agriculture, suburban, and urban land covers were thus largely removed by this analysis.

2.6. Case studies and forest gap fraction

To provide a more detailed understanding of the CLAS product, we focused on two LBA research areas where studies of forest structure, ecology, and biogeochemistry have taken place. One area is located on the Fazenda Cauaxi, in the Paragominas municipality in eastern Para State (Pereira et al. 2002). This area is a site of extensive selective logging, but it also contains several control forest areas useful for evaluation of the CLAS product. The second area is located in the Tapajos National Forest; it is an LBA core site with extensive measurements of forest structure and function (summarized by Keller et al. 2004). Tapajos contains a reduced-impact logging experiment and a large area of intact, mature forest. Together, these sites provided a means to examine details of the spectral mixture results, and to identify errors and limitations of the data.

Asner et al. (Asner et al. 2004a) developed a set of equations relating PV cover fraction derived via the AutoMCU algorithm to field-based measurements of forest canopy gap fraction. We extended these relationships to create regional mosaics of woody canopy gap fraction from the multiyear mosaic (Figure 4). We then calculated area-integrated canopy gap statistics for all IBGE vegetation classes (listed in Table 2).

3.1. Sensitivity to atmospheric aerosol and water vapor

The sensitivity of the CLAS process to atmospheric correction errors resulting from inaccurate estimates of aerosol and water vapor are shown in Figure 6. Errors in PV estimates reached a maximum of only 2.5% over dark forest vegetation when both water vapor and aerosol were very high (Figure 6a). NPV and soil estimates had maximum errors of 5% and 6.5%, respectively, over dark forest vegetation. All errors were smaller for pixels over bright clearings such as cattle pastures (Figure 6b). We would expect similarly low errors caused by poor atmospheric correction in brighter savanna vegetation types.

At typical values of atmospheric column water vapor (1.0–2.5) and aerosol optical thickness (0.1–0.3), as determined from MODIS satellite observations over Amazônia (data not shown), we expect our uncertainty in PV, NPV, and bare substrate to average less than 5% in most cases. Because the AutoMCU approach uses spectral end-member bundles (Figure 2) to propagate uncertainty in spectral signatures via Monte Carlo analysis [right side of Equation (1)], the resultant uncertainties in derived cover fractions tend to range from 1% to 8% (Asner et al. 2003a; Asner et al. 2004a). This level of uncertainty meets or exceeds that which is caused by insufficient atmospheric corrections (Figure 5). That is, the uncertainty in spectral end members on the right side of Equation (1) matches the uncertainty in pixel reflectance following atmospheric correction [left side of Equation (1)].

3.2. Regional case studies

Two regional focus areas were extracted from the multiyear mosaic results to facilitate detailed analysis of the spatial complexity of the CLAS product. The 142 km² Cauaxi region contains a variety of forest conditions, from intact mature forest (area 1) to conventionally logged forest (area 2) to cattle pasture (area 3) and forest in preparation for selective logging (area 4). The latter area is distinguished by newly developed logging roads and decks (Figure 7). A 2000 km² portion of the Tapajos National Forest is shown in Figure 8. This image contains a reduced-impact logging experiment (Keller et al. 2004) and vast areas of mature, intact forest on two major soil types: clay and sandy (as described by Silver et al. 2000).

The Cauaxi and Tapajos scenes provide a reference to better understand the spatial variance of fractional cover values among differing land-cover and soil types, presented here as a means to discover how the data express small but significant variations in forest structure. This understanding is critical to studies of basin-scale variations, as presented in subsequent sections of this paper. We also provide the separated PV, NPV, and bare substrate images for Cauaxi (in Figure 9), along with the standard deviation images generated by the AutoMCU submodel (Figure 1). The latter images express the uncertainty in fractional cover estimates on a per-pixel basis. It is notable that uncertainty in PV fractions is extremely low in most forested pixels (0%–3%), owing to the spectral separability of PV from either NPV or bare substrates. Per-pixel standard deviations increase to as much as 10% for NPV and bare covers (Figure 9), which reflects the much greater level of uncertainty involved in studies of NPV soil fractions using spectrally limited data from Landsat and similar sensors (see Asner and Heidebrecht 2002).

Analysis of 4 km² areas around each point in Figures 7 and 8 demonstrated the sensitivity of the fractional PV, NPV, and bare substrate cover to four common land-cover and edaphic scenarios found throughout Amazônia (Figure 10). In the Cauaxi image, mature tropical forest (area 1, Figure 7) had PV values of 92.0% ± 2.7% (standard deviation), NPV of 7.4% ± 6.4%, and bare substrate levels of 1.6% ± 4.3%. On an area-average basis, forest prepared for selective logging (area 2, Figure 7) had statistically similar values to that of intact forest (t tests by cover type; p > 0.20): PV = 92.1% ± 2.5%; NPV = 6.6% ± 6.1%, bare = 2.4% ± 4.3% (Figure 10). However, within the individual logging decks found in area 2, PV was 68.3% ± 11.8%, NPV was 24.3% ± 9.5%, and bare substrate was 7.9% ± 9.1%. All of these were statistically different from that of the intact forest area 1 (t tests; p < 0.01).

Following conventional logging in area 3, fractional surface cover of PV and NPV changed significantly from that of the intact forest area (t tests; p < 0.01; Figure 10). Area-integrated PV cover decreased to 86.4% ± 4.2% and NPV was elevated to 14.7% ± 8.3%, but bare substrate extent did not change (1.5% ± 4.8%). Individual pixels containing logging damage had much lower PV and higher NPV (surface debris and slash), as depicted in blue tones in area 3 of Figure 7. The fractional cover of materials substantially changed following deforestation for cattle pasture (area 4; Figure 7). In this large system of pastures, PV was 43.7% ± 3.4%, NPV was 24.4% ± 4.2%, and bare ground was 20.5% ± 2.6% (Figure 10).

As mentioned, the Tapajos image contains large areas dominated by sandy soils in low-lying zones and clay soils on uplands (Silver et al. 2000). These areas are visually distinct in Figure 8, with uplands having higher PV cover (92.2% ± 1.9%) than found on lowland terrain (89.3% ± 2.4%; t test; p < 0.05). Lowland, sandy soils areas of forest had slightly lower bare soil exposure and higher NPV, although these differences with the upland, clayey forest areas were not significant (t tests; p > 0.10).

Thus, the fractional cover of PV, NPV, and bare substrate highlighted the biophysical changes that occur with the land-use transitions from intact forest to selective logging to cattle pasture. In addition, differences in canopy opening (PV fraction) related to topo-edaphic conditions were evident. It is notable that small differences in these biophysical covers are clearly visible in the AutoMCU results. Our experience with the data indicates that the precision is very high, and that small differences in PV cover are important for defining forest types, land covers, and substrate controls over forest typologies, as will be presented in the following sections.

3.3. State-level and basinwide statistics

The IBGE vegetation map (Figure 5) provided a spatially explicit guide for developing topologies of fractional cover throughout the Amazon Basin and bordering cerrado regions. Thirty IBGE vegetation categories were identified for analysis (as shown in Table 2). The CLAS process identified 3.25 × 10⁹ pixels or 2.9 × 10⁶ km² of vegetation that was not masked for clouds, cloud shadow, and water bodies. Three IBGE classes—dense lowland tropical forest, dense submontane tropical forest, and open submontane tropical forest—accounted for more than 60% of the total area, or 1 768 000 km². However, tens to hundreds of thousands of square kilometers are covered by other vegetation types. In particular, the transition or contact areas between major vegetation classes accounts for more than 500 000 km² of Amazônia. IBGE vegetation classes vary by Brazilian state. Dense lowland tropical forests dominate the state of Amazonas, whereas open lowland (mostly bamboo) forest cover much of Acre. Submontane forests are most common in Roraima, Rondonia, Para, and Amapá, while transitional (contact) vegetation classes (e.g., dense tropical to seasonal deciduous forest) dominate in northern Mato Grosso (Table 2).

The mean fractional PV, NPV, and bare substrate covers differed significantly by vegetation class and, at times, by Brazilian state. Lowland and submontane tropical forests always had significantly higher PV cover than other vegetation types (t tests; p < 0.001). Fractional PV values were lowest in savanna regions of Roraima and Mato Grosso (Table 1). The most densely foliated ecosystems are found in Amapá and Para, with moderate NPV exposure and extremely low bare substrate fractions (Table 1). Bare substrate exposure was highest in alluvial tropical forests areas of Amapá and in open tropical forests with secondary vegetation found in Acre. The highest NPV fractions were always found in savanna systems. Seemingly small differences in fractional cover were often highly significant, even between tropical forest classes. For example, 1%–3% differences in PV fraction, which includes intercanopy shadowing, are readily observable in the CLAS results (e.g., Figures 7–8) and equate to forest gap fraction variations of 10%–25% (Asner and Warner 2003; Asner et al. 2004b). This issue is highlighted in the next section.

A close inspection of the results within each state indicated spatial gradients of surface fractional cover, both within and between forested and nonforested regions (Figures 11–17 show these state-by-state results). Areas masked by the cloud, cloud shadow, and water body detection steps by the CLAS processing are shown in white throughout these figures. In Acre, forests to the southwest along the Brazil–Peru border had higher levels of NPV and decreased PV values (Figure 11). These bamboo-dominated forests are known to have a more open crown structure with exposed woody stems and other nonphotosynthetic material. Western Amapá (Figure 12) contains some of the most densely vegetated forests in all of Amazônia, but to the east, the state transitions to open forest, dense savanna, and flooded forests with lower PV fractions (Table 1). These eastern vegetation types in Amapá have significantly higher NPV and bare substrate fractions leading toward the Atlantic Ocean.

Amazonas contains enormous areas of lowland tropical forest (Figure 13). Of the 1.7 × 10⁶ km² of land in this state, nearly one-third is dense lowland forest and an additional 10% is a more open lowland forest type according to IBGE (Figure 5). Our method was unable to detect statistically different PV, NPV, or bare substrate fractions between these two classes (Table 1). This lacking sensitivity results from our PV spectral end-member bundle including variations in shade, which alone accounts for variations in canopy openness among these densely foliated ecosystems. Imaging spectrometer (hyperspectral) observations are needed for more detailed determination of forest structure. Both forest types are densely foliated from the perspective of nadir-viewing satellite observations.

In contrast to Amazonas, the northern portion of Mato Grosso is dominated by lower stature forest, seasonal woodlands, and transitions to cerrado savanna vegetation (Figure 14). Much of the blue tone shown in Figure 14 indicates forest clearings, but a larger fraction of the land has high NPV levels due to herbaceous vegetation (Table 1). The Landsat ETM+ scenes were acquired primarily in the dry season (July–November) when cloud cover is low and herbaceous cover is senescent. Variability in NPV among pixels of the woodland and savanna vegetation classes in Mato Grosso was relatively high (7%–12%) in comparison to that of forested areas (3%–5%). This likely results from variations in regional climate, soils, and other factors controlling the availability of water for herbaceous vegetation growth, but also from greater uncertainty in measuring NPV and bare soil using multispectral observations.

In the state of Para, savanna and open woodland forests in the southeast transition into dense tropical forest in the central and northern regions (Figure 15). Extensive land use is found within this transition zone and along the TransAmazon Highway in the central portion of the state. Areas more heavily impacted by land use contain higher bare substrate fractions (red tones in Figure 15) than those with higher NPV cover (blue tones). Moreover, areas along the main stem of the Amazon River appear dominated by NPV and bare substrate fractions; these areas are alluvial plains with significant amounts of sedimentation (red tones) and herbaceous vegetation (NPV; blue tones).

The most obvious examples of the biophysical impacts of land-cover change are found in Rondonia (Figure 16). Areas once dominated by open lowland and submontane tropical forest (Figure 5) have been converted to pastoral and agricultural land uses. Very high NPV and bare substrate values are found in these clearings, with values typically ranging from 40% to 90%. Note the apparent differences in NPV and bare substrate conditions when crossing the boundary between Landsat images, such as shown in Figures 16a,b. These interscene differences highlight the challenges to large-scale automated processing of multispectral imagery from Landsat and similar sensors.

The state of Roraima contains a very wide range of fractional cover. Dense submontane forests to the west have PV, NPV, and bare substrate values averaging 90.7%, 9.2%, and 2.8%, respectively. Large areas of savanna bordering with Venezuela averaged PV, NPV, and bare substrate covers of 82.3%, 14.7%, and 5.2%, respectively. Considering the area that is averaged among these vegetation classes (Table 2), these are enormous gradients of materials with very different phenologies, biogeochemical cycles, and disturbance regimes. After masking deforested lands used to tabulate Table 1, we found Roraima to contain the largest variation in pre-land-use surface cover among all states of the Brazilian Amazon.

Integrating all results to the basin level, and masking deforestation (see section 2), we summarize the ecosystem typologies of the major vegetation types found throughout Amazônia and the bordering cerrado (Figure 18). Evergreen and deciduous tropical forests had the highest PV fractions, ranging from 84% to 89%. Bare substrate and NPV fractional cover ranged from 2% to 7% and 7% to 10%, respectively. The less common campinarana systems found throughout ∼100 000 km² of our study region develop around lowland depressions of nutrient-poor soils embedded in a matrix of tropical forest. The PV values in campinarana regions were very consistent at about 86%, but NPV varied substantially from ∼9% to 17% (Figure 18). Both the traditional forests and campinarana classes were typologically distinct from the transitional forest–savanna and wooded savanna lands bordering Amazônia. These systems had PV fractional cover values ranging from 74% to 85% depending on degree of woody cover, with concomitant differences in NPV and bare substrate (Figure 18).

Altogether these results suggest that the fractional PV, NPV, and bare substrate covers for different vegetation types varies sufficiently to warrant their separate treatment in future studies. Although the state-to-state differences in PV within most vegetation classes were often statistically different, the variation between IBGE classes was even more pronounced. These results provide an important baseline of ecosystem structural typologies for these regions and vegetation types against which future studies of forest and savanna disturbance should be compared using the same or similar method.

3.4. Woody canopy gap fraction

We estimated regional variations in woody canopy gap fraction using the CLAS results presented in Figures 11–17 along with PV gap equations derived from Asner et al. (Asner et al. 2004a). These equations were modified from their original formulation for terra firme forests to encompass a wider range of physiognomic conditions. The modifications included rescaling the equations and normalizing to values of fractional woody canopy cover reported for cerrado physiognomies (Sano and Pedrosa de Almeida 1998) and Amazon forest (Duivenvoorden 1996; Asner et al. 2004b). The equations are

View Expanded

The different denominators in Equation (2) arise from several factors. First, PV fraction is a planar metric, whereas canopy gap fraction is hemispherical in nature (Asner et al. 2004a; Asner et al. 2004b). Moreover, strong adjacency effects between satellite pixels result in a nonlinear component to vegetation mixture modeling with multispectral data (Richards 1994). This effect is maximum at forest gap values greater than ∼85%. Finally, the variable effects of intra- and intercrown shadow are not accounted for in this initial estimate. Shadow has a maximum impact on multispectral signatures of tropical forest when cover fractions are above ∼75% and on cerrado savanna observations when woody canopy cover ranges from 35% to 65% (Asner and Warner 2003). Nonetheless, these equations are a first approximation for converting our PV fractions from CLAS to estimates of woody canopy gap fraction. This effort was intended to provide an initial typology of woody canopy gap fraction for use in carbon, disturbance, and biogeochemical modeling studies (Keller et al. 2004).

Conversions of satellite-derived PV to woody canopy gap fraction highlighted the wide range of values found throughout Amazônia. The spatial variability and quality of the results are shown for Rondonia and Mato Grosso (in Figures 19a,b). Throughout the entire region, lowland tropical forests had the smallest gap fractions of 6.5% ± 1.9% (Figure 20), followed by different types of “dense forest” vegetation as classified by the IBGE (Figure 5). These dense forests had mean gap fractions ranging from 7.3% to 9.6%. Campinarana forests and woodlands (L classes; Figure 20) had more open canopies with gap fractions of approximately 14% and standard deviations averaging 4%. Although the gap fractions were not statistically unique within dense forest and campinarana classes, these two major groupings were significantly different (t tests; p < 0.001). Seasonal woodlands (F classes; Figure 20) had yet larger gap fractions of up to 21.9% ± 3.8%, and cerrado savanna classes had an enormous range of values, from 17.5% to 40.5% over thousands of square kilometers of area.

Our field measurements in five IBGE classes were near in value to those derived from these satellite estimates of woody canopy gap fraction (Figure 20). Only the field measurements from site 3 at the Tapajos National Forest (Figure From 8) were statistically different by a small margin [one-way analysis of variance (ANOVA) on ranks; p = 0.061]. Overall, the field data were well represented by this first-order model of canopy gap fraction throughout the Amazon and bordering cerrado region. Further studies in the field are needed to test and refine a canopy gap product for scientific use; however, the conceptual approach appears reasonable for this novel high-resolution mapping of a forest structural variable throughout Amazônia.