Thailand/Miller, Nualchawee and Tom
The assumption that future changes in forest cover can be measured in terms of those which
occurred in the past allows a simple projection of the fate of the future forest cover within
the site. Cross-tabulation of each pair of consecutive forest cover data planes available in the
landscape model (e.g., 1968 with 1972) provides a probability transition matrix. This matrix
is applied in a Markov process to the distribution of forest cover recorded in the second or
more recent data plane to project future forest cover trends in the test site (Fig. 7.3)*. Only
a few years of validity may be assumed for these projections, as the process controlling the
observed forest cover changes during the training period (i.e., 1968 to 1972) can only be as-
sumed to persist for a few years into the future (i.e., beyond 1972). A much more complex
model has been developed which is capable of projecting the future spatial behavior of the
forest cover of the site based upon the controlling landscape features represented in the over-
laying data planes. Essentially, the spatial projection model takes all the specific changes in
forest cover as tabulated between two consecutive forest cover data planes, groups them to-
gether into like changes, and correlates their occurrence with the corresponding landscape
parameters also available in the landscape model (i.e., slope, distance to roads and trails, etc.).
Details on this approach and its results can be found in the list of references provided.
The Landsat imagery existing for the site represented five consecutive anniversary dates which
were geometrically rectified and resampled to the one hectare square cells. Difference images
were formed between like spectral bands for consecutive pairs of dates. Displays of these
easily obtained image analysis products provide a mechanism for quick reconnaissance-level
detection of areas of change between consecutive years. These areas of change represent both
shifting cultivation (Fig. 7.5) and annual shifting crop-type patterns where permanent agri-
culture is practiced (Fig. 7.6).
Landsat land cover classification in areas of significant relief is limited in accuracy by that
portion of the radiance variation created by the terrain which is independent of the surface
cover type. Overlaying slope, aspect and elevation data planes (Fig. 7.7a) in the landscape
model provided a basis for several approaches designed to minimize this impact. Overlays of
cultural features (Fig. 7.7b) as data planes were also evaluated to determine their contribu-
tion to the accuracy of the classification of the 8 cover types noted aboveT.
The value of combining these various types of ancillary landscape data planes with the Land-
sat MSS bands was tested using stepwise linear discriminant analysis as the classifier (Table
7.1). Three types of training sets were evaluated. These consisted of rectangular representa-
tions of each cover type, rectangular areas selected to represent sun-facing and opposite
aspects, and grid sampled training sets assembled for a 1/9 sample of the test site consisting of
every third row and column of one hectare image cells. Tests were also completed to assess
the value of having apriori knowledge of the composition of the area to be mapped (propor-
tional probabilities), in comparison to having no knowledge of the composition of the area
(equal probabilities).
*Land cover trend models have also been computed for the following pairs of dates: 1954 with 1972, 1975
with 1966, 1966 with 1968 and 1968 with 1972.
TTeak was omitted in the Landsat classification as its area occupied only 40 ha.
The o
direct
the 29
agains
Fig. 7
ficatic
verific
classif
propo
mapp
100%
which
airph«
woulc
able a
map 1
classi!
acy,c
and ii
Furtt
Re
| tm
*No