Chapter 8· Analysis of Stream Macroinvertebrate Communities
139
and the output node j wh ich has minimum distance is selected as j*. If dj*(t) is
smaller than p, which is a threshold parameter in determining vigilance and
determined based on efficiency of grouping of input data, the input is assigned to
output nodej*. The weight for the nodej*, bj*i(t), then, is updated as follows:
C
1
b .'. (t + 1) = - - b .. (t)+ - - X i
(8.4)
Jl
c+l Jl
c+l
where c is the number of sampie units classified to node j*.
If dj*(t) is larger than p the input is assigned to new output. This means that
the entering input forms to a new pattern, not belonging to one of the previously
existing patterns. Then its weight briet) is newly assigned as follows:
b .'. (t + 1) = Xi .
(8.5)
J I
The benthic macroinvertebrate communities collected monthly in the Suyong
and Soktae streams in the Suyong River, Korea, from September 1993 to August
1994, were used for input data. Since the number of species (132) were too many
to train at PC (Personal Computer) level, species were grouped to 7 important taxa
such as Chironomidae, Diptera (except Chironomidae), Trichoptera,
Ephemeroptera, Miscellaneous Insecta, Oligochaeta, and Miscellaneous
Macroinvertebrate. Densities (number of individuals per e) and species richness
(number of species) in each important taxa, as well as the total density and the
total species richness, were given as inputs for training with ART. The total
number of input node for ART was sixteen.
These weights produced by ART preserve the conformational characteristic of
input data for grouping, and through them the associations among the
communities are projected into the space defined in ART (Zurada, 1992). Fig.
8.8a is an example of classification by ART (Chon et al. 2000c). The number of
trained communities was 84, while the numbers of output nodes for ART was
eleven. The threshold for vigilance, p, was set to 0.61. Similar to the case of
groupings by the Kohonen network (Fig. 8.3), communities were generally
classified according to topographical conditions and degree of pollution. Sampie
sites from the same stream had a higher tendency of grouping. Also sampie
communities collected from the polluted sites such as TCL and THP were grouped
closely, and were separated from the other less-polluted sites. The community
data collected from TKC, which were in a medium range between clean and
polluted status, appeared to be diverse, and were divided into many small groups,
separately (e.g., TKC4-2 and TKC4-3 of neuron 7 in Fig. 8.8a), or in groups with
the other similar sites (e.g., TKC3-9 and TKC4-8 of neuron 0 in Fig. 8.8a).
139
and the output node j wh ich has minimum distance is selected as j*. If dj*(t) is
smaller than p, which is a threshold parameter in determining vigilance and
determined based on efficiency of grouping of input data, the input is assigned to
output nodej*. The weight for the nodej*, bj*i(t), then, is updated as follows:
C
1
b .'. (t + 1) = - - b .. (t)+ - - X i
(8.4)
Jl
c+l Jl
c+l
where c is the number of sampie units classified to node j*.
If dj*(t) is larger than p the input is assigned to new output. This means that
the entering input forms to a new pattern, not belonging to one of the previously
existing patterns. Then its weight briet) is newly assigned as follows:
b .'. (t + 1) = Xi .
(8.5)
J I
The benthic macroinvertebrate communities collected monthly in the Suyong
and Soktae streams in the Suyong River, Korea, from September 1993 to August
1994, were used for input data. Since the number of species (132) were too many
to train at PC (Personal Computer) level, species were grouped to 7 important taxa
such as Chironomidae, Diptera (except Chironomidae), Trichoptera,
Ephemeroptera, Miscellaneous Insecta, Oligochaeta, and Miscellaneous
Macroinvertebrate. Densities (number of individuals per e) and species richness
(number of species) in each important taxa, as well as the total density and the
total species richness, were given as inputs for training with ART. The total
number of input node for ART was sixteen.
These weights produced by ART preserve the conformational characteristic of
input data for grouping, and through them the associations among the
communities are projected into the space defined in ART (Zurada, 1992). Fig.
8.8a is an example of classification by ART (Chon et al. 2000c). The number of
trained communities was 84, while the numbers of output nodes for ART was
eleven. The threshold for vigilance, p, was set to 0.61. Similar to the case of
groupings by the Kohonen network (Fig. 8.3), communities were generally
classified according to topographical conditions and degree of pollution. Sampie
sites from the same stream had a higher tendency of grouping. Also sampie
communities collected from the polluted sites such as TCL and THP were grouped
closely, and were separated from the other less-polluted sites. The community
data collected from TKC, which were in a medium range between clean and
polluted status, appeared to be diverse, and were divided into many small groups,
separately (e.g., TKC4-2 and TKC4-3 of neuron 7 in Fig. 8.8a), or in groups with
the other similar sites (e.g., TKC3-9 and TKC4-8 of neuron 0 in Fig. 8.8a).