Chapter 8· Analysis of Stream Macroinvertebrate Communities
133
Fig. 8.3 shows an example of grouping by the Kohonen network with a
mapping of 9 x 9 neurons (Chon et al. 1996). The convergence was mostly
reached in 500 - 1000 iterations. Communities were grouped according to
different impacts of pollution and topography of the sampie sites. The area of the
map was divided according to the main tributaries (ST, CM and SY), and
grouping was influenced by impacts of pollution.
Classification could be also conducted by the conventional clustering analysis
(Ludwig and Reynolds, 1988). The same input given to the Kohonen network
were provided to the clustering analysis utilizing the method of average linkage
between groups (Norusis 1986). The clustering results were in general similar to
those by the Kohonen network (Fig. 8.3), and confirmed overall groupings by the
Kohonen network. Benthic macroinvertebrates responded differently in groups to
anthropogenic impacts of pollution from oligo-saprobity to poly-saprobity as the
stream flowed down. Communities collected at higher saprobities showed higher
levels of similarities, suggesting higher degree of closeness obtained among the
communities collected at polluted sites.
It was generally difficult to directly compare performance of groupings by the
two methods, clustering and SOM. Since communities were grouped in an
unsupervised manner, there are no objective references of groupings to be
compared with. Based on experience with field data, however, mapping by the
Kohonen network appeared to be more realistic. The same groupings were
observed between SOM and the clustering in some case. For example, the group
of CM2SU, CM3SU, CM4SU and CM5SU (neuron (8 (x axis), 5 (y axis» and that
of SY2W, SY3W, SY5W and CM2W (neuron (0,5» in SOM in Fig. 3
correspondingly matched to the same sampie groups on clustering (Fig. 8.4). In
the other groups, however, discrepancies were observed. For example, the
sampies in the group of CMlSP, CM2A, SYlSU and SY2SU (neuron (2,8» in the
Kohonen mapping (Fig. 8.3) were all scattered in the clustering analysis (Fig. 8.4).
The communities patterned at this neuron, however, were more similar to field
data, and the grouping in the Kohonen network appeared to be more realistic.
Groupings in other cases (e.g., "SY2A and SY5A (neuron (6,2»" and "STlSU and
ST2SP (neuron (4,2»") in Fig. 8.3 also tended to reflect more field situations than
the groupings by the clustering analysis in the than groupings by the clustering
analysis.
The overall conformation of groupings was also more explainable in the map of
the Kohonen network. The neurons representing the three tributaries of the Suyong
River were clearly divided in the Kohonen network (Fig. 8.3). The Soktae Stream
(Sn - ST4) occupied mainly a large triangular area at the upper left part of the
map. The most neurons representing the Cholma Stream (CMl - CM5) were
located around the bottom right corner of the map, while those designating
Suyong Stream (SYl - SY5) generally occupied a diagonal area from the upper
right to bottom left corner of the map. The communities patterned at the upper left
corner were in general highly polluted, including some sampie sites in the Soktae
Stream. This polluted area was bordered with the area of the intermediate
pollution, diagonally starting from neuron (0, 3) to the upper fight corner of the
133
Fig. 8.3 shows an example of grouping by the Kohonen network with a
mapping of 9 x 9 neurons (Chon et al. 1996). The convergence was mostly
reached in 500 - 1000 iterations. Communities were grouped according to
different impacts of pollution and topography of the sampie sites. The area of the
map was divided according to the main tributaries (ST, CM and SY), and
grouping was influenced by impacts of pollution.
Classification could be also conducted by the conventional clustering analysis
(Ludwig and Reynolds, 1988). The same input given to the Kohonen network
were provided to the clustering analysis utilizing the method of average linkage
between groups (Norusis 1986). The clustering results were in general similar to
those by the Kohonen network (Fig. 8.3), and confirmed overall groupings by the
Kohonen network. Benthic macroinvertebrates responded differently in groups to
anthropogenic impacts of pollution from oligo-saprobity to poly-saprobity as the
stream flowed down. Communities collected at higher saprobities showed higher
levels of similarities, suggesting higher degree of closeness obtained among the
communities collected at polluted sites.
It was generally difficult to directly compare performance of groupings by the
two methods, clustering and SOM. Since communities were grouped in an
unsupervised manner, there are no objective references of groupings to be
compared with. Based on experience with field data, however, mapping by the
Kohonen network appeared to be more realistic. The same groupings were
observed between SOM and the clustering in some case. For example, the group
of CM2SU, CM3SU, CM4SU and CM5SU (neuron (8 (x axis), 5 (y axis» and that
of SY2W, SY3W, SY5W and CM2W (neuron (0,5» in SOM in Fig. 3
correspondingly matched to the same sampie groups on clustering (Fig. 8.4). In
the other groups, however, discrepancies were observed. For example, the
sampies in the group of CMlSP, CM2A, SYlSU and SY2SU (neuron (2,8» in the
Kohonen mapping (Fig. 8.3) were all scattered in the clustering analysis (Fig. 8.4).
The communities patterned at this neuron, however, were more similar to field
data, and the grouping in the Kohonen network appeared to be more realistic.
Groupings in other cases (e.g., "SY2A and SY5A (neuron (6,2»" and "STlSU and
ST2SP (neuron (4,2»") in Fig. 8.3 also tended to reflect more field situations than
the groupings by the clustering analysis in the than groupings by the clustering
analysis.
The overall conformation of groupings was also more explainable in the map of
the Kohonen network. The neurons representing the three tributaries of the Suyong
River were clearly divided in the Kohonen network (Fig. 8.3). The Soktae Stream
(Sn - ST4) occupied mainly a large triangular area at the upper left part of the
map. The most neurons representing the Cholma Stream (CMl - CM5) were
located around the bottom right corner of the map, while those designating
Suyong Stream (SYl - SY5) generally occupied a diagonal area from the upper
right to bottom left corner of the map. The communities patterned at the upper left
corner were in general highly polluted, including some sampie sites in the Soktae
Stream. This polluted area was bordered with the area of the intermediate
pollution, diagonally starting from neuron (0, 3) to the upper fight corner of the