Ali Akbar Abkar
concentrated. Deforestation in the district is caused mainly by the expansion of agricultural activities and fuelwood
gathering of the lowland population (Ato, 1996).
The following RS and GIS data were used for the procedure of image analysis:
e The Phrao Landsat- TM obtained on February 1989 (Fig. 2-a);
e Land use/cover map of 1989 (Fig. 2-b), which serves two purposes:
— Training field selection for generation of the radiometric likelihood vectors.
— Evaluation of the classification results.
e Land use/cover map of 1977; derived from digitization on the aerial photograph of 1977 (Fig. 2-c),
e Soil map at a scale of 1:100,000 obtained from the Dept. of Land Development in Chiang Mai, Thailand (Fig. 2-d);
Bl Mixed deciduous
Dry dipterocarp
Degraded dipterocarp
Deciduous/Dipterocary
Swidden cultivation
Irrigated paddy
Mixed field crops
Crops/Paddy
Paddy/Orchards
Orchards/Crops
Crops/Orchards
Orchards
Town
Village
Village/Orchards
ES
SS
C
af
BBOB CNN ||
Figure 2. TM image & reference maps of Phrao District, Chiang Mai Province, Thailand
(a) TM image of Phrao District; natural color composite of Band 5, 4, and 3 (RGB); image size: 402 cols x 700 rows.
(b) Land use/cover map of 1989 including 15 generalized land cover classes (4 forest and 11 non-forest classes)
(c) Land use/cover map of 1977
(d) Soil map of Phrao District at a scale of 1:100,000
3.2 Radiometric Likelihood Generator
In the case of the maximum likelihood per-pixel algorithm the class to which the pixel is finally assigned is that with the
highest probability. Probabilities of class membership, on which the assignment is based, are usually disregarded so that
after classification no information on the probabilities is available. In order to overcome the limitations of the classical
techniques, such as maximum likelihood classification per-pixel, the "images" are mapped into the likelihood of class
labels per sample under the assumption of equal prior probabilities. It is assumed that the normal distribution function
will approximate the frequency distribution associated with each of the classes.
The likelihood vectors are stored in radiometric likelihood maps. The likelihood maps or evidence maps show the
probabilities that a sample belongs to each of the defined set of possible land cover classes, obtained by supervised
classification of the multi-band satellite images. Because of the interest in the boundary between forest and non-forest,
we have combined all the forest likelihood maps in one likelihood map for forest shown, in Fig. 3-a, and all non-forest
likelihood maps in one likelihood map for non-forest, which is shown in Fig. 3-b.
a wr ut.
33 Geometric Hypothesis Generator + eo
In this section, the available data and/or knowledge contained in the
GIS database about the shape of forest and non-forest objects and
processes will be used to predict the RS data about the extent of
deforestation (expansion of the non-forest area) in the study area by
shape hypotheses generation. By using a parameterized class
membership function, geometric knowledge is represented in ta
parameterized models resulting in a hypothesis map. The
hypothesis map depends on a geometrical object model with
morphological parameters.
P EE 0.0
Figure 3. Forest and non-forest likelihood maps
; : : ; a) likelihood map for forest class
Various models of hypothesis generation using Landsat-TM b) likelihood map for non-forest class.
images and GIS data and knowledge are tested. The assumption
that deforestation starts from existing agricultural lands (non-forest
classes) and spreads out towards forest areas guide the authors toward the use of morphological model for the expansion
12 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
of a
of fc
First
max
show
class
Seco
was
morj
base
to pi
num
as 4
The
not i
hypo
of th
2-d)
defoi
cons:
proce
isotr«
isotr
Usin;
that t
of isi
const
contr
and s
3.4
In th
funct
outcc
Figur
gener
a, 5-
use/c
e
cost
X
min( cc
„0.306
