id. fuelwood
d (Fig. 2-d);
OWS.
that with the
'arded so that
j the classical
hood of class
ition function
aps show the
yy supervised
id non-forest,
all non-forest
P Mn
- DUM
Bl E
dm le bow
: “ » * 2 ox
da A Aha M
in
$5
ihood maps
lass.
the expansion
Ali Akbar Abkar
of agricultural areas and the process of deforestation. Two main approaches for hypothesis generation about the shape
of forest/non-forest objects are:
First, in the absence of GIS data, a method of automatic shape hypothesis generation of RS data by means of local
maximum likelihood (LML) classification is used as a shape hypothesis generator for LBSC (Abkar 1999). Fig. 4-b
shows the forest and non-forest cover map obtained from the automatic shape hypothesis generator by means of LML
classification of the 1989 TM image, which contains small-patched and sparsely distributed deforested areas.
Second, based on the availability of data and/or knowledge from the GIS the initial hypothesis with an irregular shape
was generated. Rules for the development and change of the irregularly shaped forest/non-forest objects from simple
morphological parameters are formulated to generate shape hypotheses. After which the parameters are determined
based on the RS data. Given an initial state of non-forest class (or forest cover) the "isotropic expand" operation is used
to predict the next state of the non-forest class (see Fig. 5). The expand hypothesis is parameterized by using the
number of successive "pixel" expands applied to the reference map of 1977. Isotropic expansion is selected optionally
as 4 nearest neighbor expansion.
The assumption underlying the isotropic expand operation (people cut trees at the forest boundary at the same rate), was
not reasonable, as appeared in the (residual) error maps, with error patterns indicating an incomplete geometric
hypothesis. Adding constraints such as soil to the deforestation process significantly improved the isotropic expansion
of the non-forest area (see Fig. 6). The soil map is resampled to the coordinate system of the Landsat-TM sensor (Fig.
2-d) and then the soil suitability map for agriculture is derived. Next, this constraint map, are crossed with the
deforestation hypothesis map to generate constraint hypothesis maps. Therefore, the prior knowledge in this case,
consists of an initial state represented in a reference map of 1977 and the underlying assumption for the generation
process is the process of expansion constrained by, for example, soil and land form. This is called “soil-constrained
isotropic expand" but the generated constraint hypotheses maps are non-isotropic, i.e. the generated result is non-
isotropic (e.g. Fig. 6-b).
Using global geometric hypothesis such as unconstraint and soil constraint isotropic expends have the disadvantages
that the generated hypothesis boundaries are artificial and cannot account for local changes. Consequently, the method
of isotropic and soil-constrained isotropic expands were extended to “non-isotropic expand" by using local radiometric
constraints for application of expand in the framework of hypothesis generation and parameter estimation; this is in
contrast with the pixel-based approaches. The method is capable of dealing with regular, irregular and small-patched
and sparsely distributed deforested areas (Fig. 7).
3.4 Hypothesis Evaluation & Error Analysis
In this section the model results are evaluated by the minimum cost solution of hypotheses. The graph of the cost
function and the representation of the residual error maps are used for the evaluation of the model to measure the
outcomes from the geometric hypotheses. Then, the results are evaluated by the minimum cost solution of hypotheses.
Figures 4-b, 5-b, 6-b, and 7-b show the forest and non-forest cover map obtained from the various models of hypothesis
generation using Landsat-TM images and GIS data and knowledge. The graphs of the costs are also shown in Figures 4-
a, 5-a, 6-a, and 7-a. The deforestation maps can be determined by comparing the classification results with the land
use/cover map of 1997 (see Figures 4-c, 5-c, 6-c, and 7-c).
Figure 4. Forest and non-forest cover
map obtained from the automatic shape
hypothesis generator.
(a) the graph of the cost as a function of the
number of shrink operations (Local min) for
LML classification. The graph has a min-
cost of 0.306451 at Local-min-parameter = 8.
(b) the. best result of the IML
classification
(c) deforestation map of Phrao between
1977 and 1989.
cost 0.3217
X
min( cost )
031
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 13