International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
ing population numbers the pressure on the forests is growing.
For the local people the forests play an important role in satisfy-
ing their daily needs (e.g. fire wood, house building material;
see Kokwaro, 1988). Other legal as well as illegal activities
since the early colonial time at the beginning of the 20th century
till today have resulted in forest degradation (Mitchell, in print).
Only small patches of intact forest are left. The heaviliy dis-
turbed Kakamega Forest is said to have been reduced to ca. 120
km? in 1980 (Kokwaro, 1988). KIFCON (1994) estimated the
offtake of fuel wood as ca. 100,000 m? per year.
3. MATERIALS AND METHODS
3.1 Data and software
"Landsat satellite imagery was ordered with the aim to reach
back as far as possible as well as to cover the time period in
regular intervals. The achieved time series encompasses 7 time
steps between 1972 and 2001 with roughly one image every
fifth year. For the time steps 1972, 1975, 1980, 1995 und 2001
two for most parts cloudless scenes from the dry and the rainy
season are available. This allows for taking seasonal variation in
vegetation patterns into account. For the remaining time steps
(1984 und 1989) only a single scene from the dry season is at
hand for classification. The scenes were detected by different
Landsat sensors, resulting in specific characteristics regarding
their spatial as well as radiometric resolutions. The data of the
first three time steps (1972, 1975 und 1980) originate of LAND-
SAT-MSS (Multispectral Scanner). The image material for
1984, 1989 und 1995 were derived by LANDSAT-TM (The-
matic Mapper), whereas the data from 2001 were collected by
LANDSAT-ETM+ (Enhanced Thematic Mapper Plus). Due to
the fact that the data was ordered in system-corrected format,
the MSS-data had been resampled to 60 x 60 m° spatial resolu-
tion, that is a multiple of the 30 x 30 m? TM/ETM- data resolu-
tion. :
For digital image processing of the remotely-sensed imagery
including a supervised multispectral classification ERDAS
Imagine 8.5 is used. This software includes the module ATCOR
3 for correcting atmospheric and terrain effects. For analysing
the classification results use is made of ArcGIS 8.3 that sup-
ports GIS overlay functionality with due accuracy.
3.2 Preprocessing
Preprocessing of the satellite imagery envolves the exact geo-
referencing of the various scenes by means of topographic map
sheets in 1:50,000 scale which are the most actual ones avail-
able but have not been updated since 1970. Here a precision of
one pixel or better (i.e. less than 30 m) is envisaged. Next, the
scenes are clipped for an area of 60 km by 65 km (UTM 36 N:
680,000 — 740,000 E by -5,000 — 60,000 N; ellipsoid Clarke
1880, datum Arc 1960 New). Atmospheric as well as terrain
shading effects are corrected via ATCOR 3 based on a digital
terrain model which was derived from the 1:50,000 topo maps
contours. For an improved visual interpretation of the scenes on
the screen a piecc-wise linear transformation (via breakpoints)
is applied for band combination 5/4/3 (TM, ETM+) and 2/4/1
(MSS), respectively. Further, in all scenes with clouds these
areas are masked-out and if available replaced by the informa-
tion of the second scene of the particular time steps, i.e. by the
pixel values of six (TM, ETM+) or four (MSS) spectral bands.
3.3 Threshold Analysis
The first step towards a classification of the landscape is the
threshold analysis aiming at the generation of binary images
distinguishing forest from ,.no forest". This distinction could
be made use of later by performing multispectral classifications
independently for these two major land cover classes (Lillesand
& Kiefer, 2000). The different spectral channels as well as
several vegetation indices (e.g. NDVI, SAVI) are evaluated
concerning their suitability. Band 2 and 1 (Green) in the case of
TM/ETM+ and MSS, respectively, turned out to be best for
separating "forest and "no forest“ (Lung, 2004). First numbers
of forested areas are derived by overlay techniques combining
the resulting threshold images and raster layers of the official
forest areas (gazetted in the 1930s), the latter derived by digitiz-
ing their boundaries from the already mentioned 1:50,000 topo
maps. However, even though a temporally changing pattern of
forest losses and replanting is observed, in total no major
change in forest-covered area is revealed. What is needed for
describing forest fragmentation and disturbances in detail is to
distinguish between more land cover classes in order to separate
near natural forest from secondary forest or even plantation
forest. Further, the results of the threshold analysis demonstrate
that a truely satisfying separation of "forest" and "no forest" is
not possible when considering just one spectral band. Therefore,
the subsequent multispectral classification is not to be per-
formed independantly for these two major classes.
3.4 Supervised Classification
The multispectral classification of the Landsat time series starts
with the most actual time step (2001) and subsequently goes
back till the earliest time step (1972). It makes use of the maxi-
mum-likelihood classificator, deriving training areas (several
per land cover class whereever possible) from a) different maps
with vegetation information, b) amateur photographs taken from
an aeroplane in 2001, and c) terrain references. While photo-
graphs and terrain references are considered as ground truth for
timestep 2001, the vegetation maps (Vegetation map 1:250,000
from 1966, Forest Department forest map 1:10,000 from 1972,
KIFCON land cover map 1:25,000 from 1991) are a valuable
source for ground truthing in the past. The development of a
methodolgy for a best-possible classification based on the satel-
lite imagery of 2001 can be subdivided in three steps: 1) Via
signature analysis of the training areas the spectral bands to be
considered are evaluated for the dry season image regarding the
spectral separation of envisaged 17 land cover classes. Because
the ETM+/TM channels 1 und 2 contribute only a very small
information amount for distinguishing the desired land cover
classes, they are disregarded in the classification process. 2) By
adding the rainy season image the classification is improved in
particular regarding the separation of grassland and agricultural
land. Thus a multiseasonal approach is to be preferred against a
monoseasonal approach. 3) A further improvement in the classi-
fication results from including the ratio 7/2 (ETM+/TM) as an
additional artificial channel. This ratio band showed to be most
suitable for differentiating between the vegetation formations
(Lung, 2004; see also Hildebrandt, 1996). Having finalized the
classification for timestep 2001, the developed methodology is
applied as exactly as possible to the data of the other time steps
in order to derive comparable results. However, minor modifi-
cations are necessary: For time steps with only one satellite
image at hand only a monoscasonal classification approach E
possible. For classifications of MSS-data other spectral band
combinations are to be used due to sensor differences. Also the
use of an artificial band did not gain improvements. To summa-
rize, for classifying ETM+/TM-data the bands 3, 4. 5, 7, and 7/2
are used, in the case of MSS-data the bands 1. 2, 3, and 4.
466
Inter
|. Ne
Seco!
2. Se
3 Bi
5. Gr
scatti
6. Gr
g Pk
Pinu.
8. Pl
Bisc}
9. Te
10. A
11. V
Inste
sat s
verifi
the y
Figur
form:
genei
been
the fc
sec ]
ondai
Psidi
“Nea
vived
forest
on ar
tions
large
the c.
Euca
the c
