2 STUDY AREA AND DATA SET
The studied area is the Krutynia river basin in the Great
Mazurian Lake district in North East Poland (fig. 2). The data
used included two Landsat TM scenes (obtained April 2nd
1990, and June 21st 1990), and manually digitised contour
lines of elevation and a digitised soil map (from map scales
1:50 000). All images were transformed to a common (local)
co-ordinate system (RMSE - 10 meters). Training data was
created (by one of us - JC) from manual interpretation of a
colour composite image made from the April data, and infra
red aerial photographs (taken October 15th 1995). Ground truth
data was collected (by both of us) during a field visit in May
1995. Classification and data sampling was done into 8 classes
(cf. table 1).
a
1 Urban area
AR Open land
S ei za Forest
t POLAND / mmm Wetland
mmm Water
& |
uy
Figure 2. Study area with Krutynia river basin.
3 METHODS
Two classification methods were tested: i) a GIS integrated
expert system, and ii) traditional maximum likelihood
classification. The expert system is developed by one of the
authors (TG), mainly as a tool for learning integrated image
classification (Gumbricht, 1996). The maximum likelihood
classification was done in IDRISI (Eastman, 1993).
184
3.1 Expert system classifier - guide
The expert system "GUIDE" is an inference tool using forward
propagation (or chaining) and declarative knowledge. It can be
used for both Boolean and fuzzy knowledge based
classification of raster images (Fig. 3). Guide is supported by
MS-DOS and is adapted to IDRISI format. Rules are either
typed into an ascii file by the user, or automatically extracted
from training data. Typically quantitative field data relations
are extracted from training data, whereas object data relations
(e.g. landuse and regolith symmetries) must be manually
inferred by domain experts. Guide can handle images of
different resolution, and segmentation according to positions
(i.e. rows and columns of the cells).
membership
1 bod Cu:
0
|
Pel 2.3.4.5 6 278 9 10° 11. siope
Figure 3. Example of fuzzy membership function to slope.
Between slopes 2 and 6 the membership is 1, whereas it
changes gradually from 0 to 1 between slope 0 and 2 and 11
and 6 respectively. In guide the above fuzzy membership
function is written “Whenimg @ 0 2 TO 6 11 Slope", the
Boolean logic is written "Whenimg (a) 2 TO 6 Slope".
Rules are compact of the form "if conditionl ... (and
condition2, and ...) then conclusion ...". Conditions are in the
form of operators (=x, <x, >X, <x<), where each operator can
be a vector of observations associated with the conclusion
(table 1). For a conclusion to become true either the same
vector observation must be true in all conditions of a single
output class, or it is enough with one observation of each
condition to be true. Categorisation of each output class is thus
possible through a single statement, still keeping a high
transparency to the user. For instance different forest classes
(e.g. deciduous, coniferous) can be merged in a single rule
without mixing of sub-category observations (illustrated in
table 1).
In Boolean mode a pixel that has been given a value will keep
this value and it can not be changed by a condition further
down in the guide-file. In the fuzzy mode each cell is given a
membership function (degree of belonging) associated with
each output category. The category with the highest
membership assigns the cell in the final output. In fuzzy mode
guide also produces an image showing the membership
function related to the assigned category for each cell
Membership function can either be cumulative (more rules give
higher mf), or averaged (maximum mf = 1). The user can
optionally choose to produce images of membership functions
for all of the output categories. Only linear membership
functions are supported (Fig. 3). If a cell does not satisfy any of
the stated conditions it will remain unclassified.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
Tabl
can |
con
"WE
