ul 2004
in could
1cations
illesand
well as
valuated
' case of
best for
iumbers
mbining
official
digitiz-
00 topo
ittern of
) major
:ded for
all is to
separate
antation
onstrate
orest" is
ierefore,
be per-
es starts
tly goes
1e maxi-
(several
nt maps
en from
2 photo-
truth for
250,000
m 1972,
valuable
ent of a
he satel-
1) Via
ds to be
ding the
Because
ry small
1d cover
IS. 2 By
roved in
icultural
against a
1e classi-
V1) as an
be most
rmations
lized the
lology 1s
me steps
- modifi-
satellite
yroach 1s
ral band
Also the
summa-
Land 7/2
4.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
Table 1: Land cover classes distinguished in the supervised
Landsat imagery classification process
Class Description
Forest of lowest disturbance level, dense
canopy, older than 50 years as well as old
secondary forest (30-50 years)
1. Near natural + old
secondary forest
Mid-aged secondary forest of 20-30 years
as well as aged Maesopsis eminii (origi-
nally from Uganda) plantations mixed
with indigenous species
2. Secondary forest
Bushed areas interspersed with grasses
and herbs plus young (10-20 years) and
very young (initial state, younger than 10
years) secondary forest, also early mixed
Maesopsis eminii plantations
3. Bushland / shrubs
Colonization of guava trees (animal-
dispersed, e.g. by monkeys)
4. Secondary bushland
- Psidium guajava
5. Grassland with
scattered trees
Grassland with single bushes or trees
Grassland, partially of natural origin,
partially due to clearings, partly used as
meadows, grass used for roof hatching
6. Grassland
7. Plantation forest -
Pinus patula
Plantation of pine trees (originally from
Mexico, monocultures), maybe of cypress
Plantation of bischoffia trees (originally
from Uganda, monocultures, could be
without leaves due to pest)
8. Plantation forest -
Bischoffia javanica
9. Tea plantation Tea plantation
Cultivated land of diverse characteristics,
highly devided land with trees and bushes
along plot boundaries, mainly subsistence
agriculture, high percentage of bare
ground
10. Agricultural land
11. Water Water
Roads (tarmac or dirt track), rocks, set-
12. Others
tlements
4. RESULTS AND DISCUSSION
Instead of desired 17 classes, on the basis of the available Land-
sat satellite imagery and the reference data for ground truth
verification 12 land cover classes can be realized. A subset of
the year 2001 classification is shown for Kakamega Forest in
Figure 2. From the 12 land cover classes 6 belong to forest
formations. Thus, a differentiation of tropical rain forest in
general is possible when classifying Landsat imagery. There has
been no need in differentiating the cultivated land surrounding
the forested areas. For a more detailed description of the classes
see Table |. Classes 6 to | form successional stages, with “Sec-
ondary bushland Psidium guajava" standing out because
Psidium guajava 1s not a real forest tree species. Areas of class
“Near natural + old secondary forest” are likely to have sur-
vived over the long term or when representing old secondary
forest have regenerated to this final stage of natural succession
on areas which have been disturbed by man-kind. Forest planta-
tions can be distinguised as long as they are monocultures and
large enough in size to be reprensented by pure pixels. This is
the case for Pinus patula and Bischoffia javanica, but not for
Lucalyprus saligna and Cupressus lusitanica. A shortcome of
the classification is that Maesopsis eminii plantation cannot be
467
separated due to being planted mixed in with other indigenous
tree species. In these cases the spectral signatures are to similar
to several secondary forest stages. Forest plantation still hiding
in these classes might be later revealed at least for Kakamega
Forest by following a rule-based hybrid approach, that involves
the Forest Department forest map in 1:10,000 scale as well as
visual interpretation of the contrast-enhanced band combination
5/4/13 (ETM+/TM).
4.1 Visual evaluation of the classification
All derived classifications for the seven time steps were visually
evaluated in order to judge their accuracy. For the different
forest areas or parts of them very distinct developments can be
observed. Clear fellings of “Near natural + old secondary for-
est“ und “Secondary forest" in favour of bushland, grassland
and agricultural land are obvious all over the area, c.g. for the
western arm and the most southern parts of Kakamega Forest
(see Figure 2). Other areas, like the middle part or the most
western end of Kakamega Forest are characterized by a conti-
nous change of forest plantations and their fellings. Along the
north-eastern edge of Kakamega Forest regeneration of forest
can be noticed in younger times (1994/95 and 2001), shown by
grassland with scattered trees or even arrangements of succes-
sional stages. And, from 1994/95 onwards in the north-western
area colonization of Psidium guajava on former grassland or
agricultural land are found. Especially in the classification
results based on ETM+/TM-data with the higher resolution as
compared to MSS-data numerous scattered pixels of the classes
“Secondary Forest“ and “Bushland / shrubs“ are spread
throughout the major areas of “Near natural + old secondary
forest". This is a an indication for likely disturbance of former
prestine forest through selective logging (compare with
Mitchell, in print). For South Nandi Forest the portion of inter-
spersed "Bushland / shrubs" pixels is much higher as compared
to interspersed “Secondary Forest” pixels. Therefore, here this
process of selective logging seems to go still on to a much
higher rate as compared to Kakamega Forest, where major
disturbances by selecting and felling certain tree species seem to
have happened longer ago.
So far only a visual judgement regarding the quality of the
classifications is possible. What is still missing is an accuracy
assessment via error matrices opposing the classification results
with field reference data. Getting such reference data for all the
different time steps as covered in the time series seems to be a
big if not unsoluble effort. But at least for the most actual
timestep (2001) such an assessment should be possible and is
aimed at, as within BIOTA-East access to aerial photography of
the year 2000 is sought for. For the moment the evaluation is
based on the available ground truth reference data as well as on
interpreting the likelyhood of correct assignement to land cover
classes by putting the results of the single time steps in their
cronical order. This allows to point out typical trends but also
likely misclassifications. In general, the separation of grassland
and agricultural land is critical because their spectral distinction
depends highly on date of image acquisition, i.e. is correlated to
the development stages of the field crops. Because in this area a
large variation of crops is cultivated by the local people with up
to three harvests per year depending on the changing pattern of
rainy and dry seasons (Jitzold & Schmidt, 1982) it seems to be
almost impossible to recommend certain times of the year to be
covered by imagery. However, the classification result is im-
proved when a multiseasonal approach is followed based on at
least two scenes which represent different stages in the cultiva-
tion cycle. This seems to be not the case for the 1975 imagery
even though two scenes had been ordered with the intention to
cover dry and rainy seasons. Also difficulties arouse concerning
