249
Following geographic registration, a rectangular
area 1747 colums (approximately 100 km) by 2659
rows (approximately 150 km) of overlap was
selected for direct comparison.
Average exoatmospheric reflectance values for each
of the Landsat MSS bandwidths in the areas of
scene overlap provided evidence of seme important
differences between the two dates (Table 1).
Differences in MSS Band 2 (Band 5 on 1973 data
set), the bandwidth where plant chlorophyll
strongly absorbs photosynthetically active
radiation, are significant and suggest that the
1987 data provide a view of a less hardy, perhaps
drier landscape. Increased backscatter from the
smoke contaminants in the atmosphere should not
selectively occur only in visible red wavelengths,
so this difference provides one piece of evidence
indicating area-wide change in the environmental
condition. Lower average reflectance in Landsat
MSS Band 4 (7) also is typical of the 1987 data
set. One interpretation of this difference is
that higher levels of water vapor were present in
the atmosphere at the time of the 1987 data
collection, because water vapor strongly absorbs
in these wavelengths.
Table 1. Average exoatmospheric reflectance values
for the two corresponding Landsat MSS scenes - for
each sub-site.
Year MSS 1(4) MSS 2(5) MSS 3(6) MSS 4(7)
SITE
1-A
1973
26.1
17.1
26.8
18.2
1987
22.2
18.7
26.6
12.2
SITE
1-B
1973
27.3
20.0
25.8
16.0
1987
24.0
22.2
27.3
11.8
SITE
1-C
1973
27.2
20.4
22.5
12.7
1987
24.2
23.4
21.4
8.2
SITE
1-D
1973
27.0
19.1
25.4
15.9
1987
23.6
21.7
24.5
10.4
SITE
1-E
1973
27.1
20.4
25.2
15.5
1987
24.0
23.7
23.3
9.4
SITE
1-F
1973
26.8
18.9
27.9
18.7
1987
23.9
22.5
25.4
10.7
The average exoatmospheric reflectance values
indicate that Site 1-A exhibits a relatively
slight increase in the visible red wavelengths
[MSS Band 2 (5)] during the fourteen year
interval. At all other sub-sites, this increase
is larger with the greatest changes occurring at
Sites 1-C, E, and F. All six sub-sites exhibit
lower reflectance in MSS Band 4 (near IR) in 1987.
Since these lower values are characteristic of the
sub-sites that have little smoke contamination,
the suggestion concerning increased water vapor
levels and atmospheric absorption of these
wavelengths remains a viable hypothesis.
METHODS
This investigation focussed on identification of
forest land cover and changes to those forests
over time. An unsupervised parametric
classification algorithm was employed whereby the
interpretation of the classified data was based on
an understanding of the spectral signatures
representative of forests, as well as grasslands.
A nonparametric classification algorithm was also
implemented and discussed in detail elsewhere
(Mausel et al., 1990).
A statistical clustering (STATCL) program
available in the ERDAS image processing software
package (ERDAS, 1989) was used to develop
statistics for grouping spectral signatures. A
Euclidean distance maximum likelihood classifier
(MAXCLAS) was used for final classification.
Given the diversity in cloud/smoke/haze
characteristics across the 1987 data set, two
subareas in cloud-free segments containing all
major features of interest were used to generate
statistical classes. Geographic distributions of
the numerous spectral clusters as well as analysis
of spectral signatures for the feature classes
aided in the selection of statistical classes used
in image classification.
Vegetation indexes are widely used in remote
sensing studies designed to assess spatial and
temporal variation in vegetative condition, and
the information content of different vegetation
indices is functionally equivalent (Perry and
Lautenschlager, 1984). Hence, only the normalized
difference vegetation index (NDVI) was evaluated
for this research project. The majority of
studies using Landsat MSS data and NDVI have
utilized Band 4 (near IR) in the ratio: NDVI =
(4—2)/(4+2). However, the influence of absorption
by atmospheric moisture on the 1987 Landsat MSS
data in Band 4 made the use of this equation
problematic. An alternative NDVI equation
involving Landsat MSS Band 3 (also near IR) was
used; this ratio is: NDVI = (3- 2)/(3+2).
Transformation of multiband, multitemporal data
sets into principal components (PC) provides an
additional method of change detection (Lee et al.,
1989; Conese et al., 1988; Fung and LeDrew, 1987).
PC analysis can identify categories of change, but
that change is in terms of overall spectral
change. The PC results were similar to those
derived from a combination of the parametric
classifier and the NDVI differencing routines as
reported elsewhere (Mausel et al., 1990).
RESULTS AND DISCUSSION
From a remote sensing perspective, White (1983)
identifies several key factors that influence
interpretation and/or classification of the
satellite data. One factor is that relatively
clear sky conditions exist during the short dry
seasons of July and August and again in December
and January which also corresponds with the time
of probable grassland burning and leaf drop by the
semi-deciduous tree species. As a result, within
class variability in spectral response is
relatively high but between class distinctions are
relatively low. An additional factor to be
considered during interpretation is the structure
of the closed forest; emergents within the rain
forest casts shadows that lower the spectral
estimate of vegetative vigor for these areas.
Three major spectral classes are used in this
remote sensing study: CLOSED FOREST, OPER FOREST,
and NONFOREST (Table 2). Spectral classes are
widely used in land cover mapping and are related
to surface biophysical characteristics (Jensen,
1983). Visual image interpretation indicates that
there is a close correspondence between the major
vegetation types identified by White (1983) and
the spectral classes used in this research. The
spectral class CLOSED FOREST corresponds with the
closed canopy of the Guineo-Congolian rain forest
— drier types and those located on lowlands.
OPEN FOREST refers to rain forest with openings in
the canopy and/or narrow tracks of lowland forests
