136
T.-S. Chon . Y.S. Park· I.-S. Kwak . E.Y. Cha
an ordination tool. From the aspect of reducing dimensions, the Kohonen network
is basically similar to Principal Component Analysis (PCA): input data
dimensions are effectively contracted to a limited number of dimensions in output
(e.g., 2 or 3 dimensions). Not only for sampie communities (Q mode), different
taxa (R mode) also could be grouped on SOM. Fig. 8.6 shows mapping of the
selected taxa in benthic macroinvertebrate communities collected at Cholma,
Suyong, Heodong, and Seoktae streams in the Suyong River through training by
the Kohonen network. The species were classified according to the gradient of
pollution and abundance. The right area of the map was represented by pollution
tolerant species, while the left area was occupied by pollution intolerant species.
For example, Limnodrilus hoffmeisteri and Chironomus sp., which were collected
in streams
polluted by organic matters, were grouped in the lower right (neuron (8, 8) (No.
2)), while Viviparidae (No. 7), Ordobrevia sp. (No. 15), and Paraphaenocladius
sp. (No. 69), which were collected at relatively unpolluted streams, were located at
the lower left area of the map (neuron (0,8)). The figure also showed that the
lower area was patterned with abundant species, while the upper area was
occupied by relatively rare species. Ecological explanation on associations of
different taxa of benthic macroinvertebrates in the Suyong River will be reported
elsewhere.
Melssen et al. (1993) mentioned that the huge number of data variables may
yield a larger number of significant principal components in PCA so that it may
not retain sufficient information if only a few principal components are used for
visualizing the multi dimensional data space. Also some computational problems
might arise due to the large number of variables, such as calculating (pseudo-)
inverse of the covariance matrix. However the Kohonen network, trained in an
unsupervised fashion, could be utilized to map the multidimensional data space on
two or a few more dimensions, preserving the existing topology as much as
possible. Lohninger and Stanc (1992) compared the Kohonen mapping and the knearest neighbour clustering in classification of mass spectral data in chemical
compositions. They reported that the former was superior in all cases they tested.
The comparison between the Kohonen network and statistical clustering methods
is further discussed in Chon et al. (1996).
Although each neuron patternizes a group of similar communities and the
neurons representing communities under similar environmental conditions are
generally located in groups on the map, the distances among patternized neurons
measured on the map may not directly indicate the degree of closeness among
communities. Interpreting the distances among neurons on the trained map is a
complicate problem, considering that the original multivariate data set was
transformed into aspace of a few dimensions. Further investigations are needed
to express the degree of associations among communities in a more feasible
manner in reduced dimensions on the map.
T.-S. Chon . Y.S. Park· I.-S. Kwak . E.Y. Cha
an ordination tool. From the aspect of reducing dimensions, the Kohonen network
is basically similar to Principal Component Analysis (PCA): input data
dimensions are effectively contracted to a limited number of dimensions in output
(e.g., 2 or 3 dimensions). Not only for sampie communities (Q mode), different
taxa (R mode) also could be grouped on SOM. Fig. 8.6 shows mapping of the
selected taxa in benthic macroinvertebrate communities collected at Cholma,
Suyong, Heodong, and Seoktae streams in the Suyong River through training by
the Kohonen network. The species were classified according to the gradient of
pollution and abundance. The right area of the map was represented by pollution
tolerant species, while the left area was occupied by pollution intolerant species.
For example, Limnodrilus hoffmeisteri and Chironomus sp., which were collected
in streams
polluted by organic matters, were grouped in the lower right (neuron (8, 8) (No.
2)), while Viviparidae (No. 7), Ordobrevia sp. (No. 15), and Paraphaenocladius
sp. (No. 69), which were collected at relatively unpolluted streams, were located at
the lower left area of the map (neuron (0,8)). The figure also showed that the
lower area was patterned with abundant species, while the upper area was
occupied by relatively rare species. Ecological explanation on associations of
different taxa of benthic macroinvertebrates in the Suyong River will be reported
elsewhere.
Melssen et al. (1993) mentioned that the huge number of data variables may
yield a larger number of significant principal components in PCA so that it may
not retain sufficient information if only a few principal components are used for
visualizing the multi dimensional data space. Also some computational problems
might arise due to the large number of variables, such as calculating (pseudo-)
inverse of the covariance matrix. However the Kohonen network, trained in an
unsupervised fashion, could be utilized to map the multidimensional data space on
two or a few more dimensions, preserving the existing topology as much as
possible. Lohninger and Stanc (1992) compared the Kohonen mapping and the knearest neighbour clustering in classification of mass spectral data in chemical
compositions. They reported that the former was superior in all cases they tested.
The comparison between the Kohonen network and statistical clustering methods
is further discussed in Chon et al. (1996).
Although each neuron patternizes a group of similar communities and the
neurons representing communities under similar environmental conditions are
generally located in groups on the map, the distances among patternized neurons
measured on the map may not directly indicate the degree of closeness among
communities. Interpreting the distances among neurons on the trained map is a
complicate problem, considering that the original multivariate data set was
transformed into aspace of a few dimensions. Further investigations are needed
to express the degree of associations among communities in a more feasible
manner in reduced dimensions on the map.