As an initial evaluation of the capability of the reconstructed SASS imagery to discriminate between various vegetation types, an experiment using measurements made over the extended Amazon basin was conducted. The study area portion of the image is shown in Fig. 8. This area was selected to avoid regions where available vegetation information may be suspect and to avoid areas of high local topographic relief, e.g., the Andes.
Using the UNESCO [15] small-scale land cover map published in 1978
for reference, vegetation in the central South American study area was divided
into forest, woodland, and grass-shrublands categories. The forest group
consisted primarily of extremely wet rainforest types, moist seasonal forest,
wet submontane forest, and related degraded forest formations. The subhumid
woodland complex consisted of several vegetation communities, associated
mosaics, and cultivated landscapes. The specific woodland types included chaco
and caatinga. The subhumid shrub-grasslands land cover class
included a variety of grass-woody species mixtures, pantanal and agriculture.
The statistics from the vertically-polarized image for each class and the
three major groupings (forest, woodland, and shrub-grassland) is presented
in Table 1.
In a supervised classification experiment utilizing quadratic discriminant
functions with training and withheld data sets, the vertically-polarized
image data was found adequately diverse to distinguish between humid forest,
woodland, and shrub-grassland with a classification accuracy rate of 88%(see
Table 2) [9]. Error in the classification was distributed logically
between categories with similar land cover.
Sobti et al. [13] suggest it may be more appropriate to classify
terrain types predicated on ``expected microwave response'' instead of a
priori land cover schemes. With this working hypothesis in mind and after
some initial experimentation and exploratory clustering, it became apparent that
between nine and twelve unique backscatter classes existed in the data set.
While somewhat greater statistical divergence could be achieved using
alternative divisions, a twelve-cluster solution was selected due to its greater
interpretability and ability to show gradations in the data. The average
statistical divergence [12] between these clusters was a very high 0.96
(on a scale between 0 and 1), and the average statistical divergence between
each cluster and its nearest neighboring cluster was a moderately high 0.75.
While the clusters were statistically unique, interpreting the clusters and
assigning the clusters unique labels was somewhat subjective. In the discussion
of the clusters below, the clusters have been ordered from highest to lowest
average . Table 3 shows the primary formations constituting each
cluster. For a formation to be considered primary, it had to either account for
10%of the pixels within the cluster or, alternatively, the cluster had
to subsume at least 10%of the formation.
The first four clusters can loosely be classified as tropical forest groupings. Generally these forests are found astride the equator in the central Amazon basin and northward into Venezuela, Guyana, and Surinam. Bounded on the east by the Pacific Ocean or a variety of coastal vegetation and agriculture, these forests stretch across the continent and extend up the eastern Andean slopes. The first (1) cluster consists of very moist forest, moist seasonal forest, wet submontane forest in northern Brazil (ending its wet season) and extremely moist forest. Cluster 2 consists of the same formations, but also includes some tropical evergreen seasonal lowland forest. Cluster 3 consists almost entirely of very moist forest and moist seasonal forest. It should also be noted that degraded forest formations also make-up a small fraction of these tropical forest clusters. Despite the introduction of some degraded woodland formation pixels, Cluster 4 is primarily a forest cluster, with 81%of its pixels originating from forest formations. In summary, given the trend through the remainder of the clusters, we interpret Clusters 1 through 4 to be tropical forest clusters with varying canopy densities, different canopy structures, or communities in different stages of seasonal growth or vigor.
With large mean decreases of .57 dB and .51 dB respectively, Clusters 5 and 6
appear to be transition clusters between the forest and woodland groups. While
18%of the tropical evergreen seasonal lowland forest formation is included in
Cluster 5, woodland formations account for 62%of its membership - primarily
degraded formations and caatinga. Cluster 6 is primarily wooded chaco, mixed with caatinga. While these two formations are widely
separated geographically, their combination is logical. First, in their typical
definitions, both the caatinga and chaco can accurately be
described as drought-deciduous lowland formations of woodland trees mixed with
shrubs and Cactaceae. Furthermore, both formations are found in regions with
nearly identical yearly precipitation characteristics, mean annual temperatures,
and mean annual numbers of dry months. Specifically, the chaco is found in
areas of Brazil bordering on the Rio Paraguay, and in the Chaco province of
Paraguay. These areas have a yearly precipitation between 500 to 1000 mm, mean
annual temperatures ranging between 20 and 25 degrees centigrade, and six to
seven dry months. The caatinga is found in areas of northeast Brazil, with
yearly precipitation between 500 and 1000 mm, mean annual temperatures between
23 and 29 degrees centigrade, and between six and eight dry months
[15]. Given the apparent trend in the clusters to reflect vegetation
formations of increasingly sparse tree canopy cover, Clusters 5 and 6 may
reflect chaco with tree and shrub spacing densities similar to some areas
of caatinga.
Cluster 7 is also has a mixture of caatinga and chaco, but includes
14%of the campos cerrados (N) pixels. This northern
formation of campos cerrados is found in central Brazil, and forms a zone
between the Amazonian forests to the east and north and the caatinga along
its western border. Physiognomy of the campos cerrados (N) varies from a
shrub savanna to woodland savanna of light to dense canopy densities. Gallery
forests and areas of grassland are also common.
Cluster 8 is primarily a campos cerrados (N) and chaco cluster, with
some caatinga, and pantanal. Interestingly, 45%of the
degraded caatinga formation is subsumed in this cluster, although it
accounts for only 5%of the total cluster membership. Cluster 9 is also
primarily a campos cerrados (S) cluster, but contains 10%pantanal too. This southern variety of campos cerrados is distinguished
from campos cerrados (N) by its shorter dry season and definite cool
season. The remaining 29%of the cluster pixels is accounted for by other
members of the woodland and shrub-grassland formations. While Cluster 10 is
also a campos cerrados (N&S) cluster, a significant number of grassland
formations account for a large percentage of the pixels, including pantanal, campos sujos,
campos limpos, and grassland with palms. Given
the moderate decrease in backscatter from the previous cluster (.47 dB) and the
entry of grassland constituents into the cluster, Cluster 10 represents a
transition between woodland-shrubland clusters to shrubland-grassland clusters.
Cluster 11 has a mean of .57 dB less than Cluster 10, and consists
primarily of agriculture and grasslands. It also include 42%of the campos
cerrados (S) pixels, 68%of the degraded subhumid campos cerrados
formation, and 34%of the grassland with palms. Cluster 12 has and average
of nearly a decibel lower than Cluster 11, and consists mostly of campos
sujos/limpos. Almost 20%of the grassland with palms formation is subsumed in
this cluster, with small percentages of campos cerrados (N&S).
In attempting to assess the meaning of each cluster, the limitations of using
the UNESCO map for ground truth, with all its cartographic generalization and
classification became apparent-it does not contain sufficient areal detail
within the mapped formations to account for the observed detail captured in the
clustering analysis. Based on 1) the pattern exhibited in the clustering, 2) the
nominal cluster constituents, 3) the general formation descriptions, and 4) the
consistency of the results through the cluster sequence, it seems the
coefficients are expressions of a large-scale physiognomic characteristic typical
of the formations rather than the species composition of each formation. For an
individual pixel, we hypothesize that this characteristic is canopy
density/cover integrated over the
16 km
area spanned by a single
pixel. This implies that the relative amounts of bare ground, grass, shrubbery,
and trees that vary between (and within) formations also will be reflected
in the backscatter coefficients. The phenophase (i.e., seasonal growth stage)
with its attendant differences in vigor, changing canopy structures, and canopy
moisture content will also influence the backscatter values. However, given the
information from the map, this hypothesis cannot be definitively tested.
Further research to investigate these hypotheses continues.