Chapter 8· Analysis of Stream Macroinvertebrate Communities
165
preserved in the weights of the network. Y' served as trained results with the input
of X, while X' matching to the input of Y (Fig. 8.20).
Since the patterns of relation were established between X and Y in the network
through counterpropagation, the variables in the hierarchicallevels could mutually
respond to each other to new data sets. For example, if newly collected input data
set, X new ' were given to the trained network, they would produce new recognized
data set Y" corresponding to X. ew ' This could be compared with the actual field
data, X "" (Fig. 8.20).
The benthic macroinvertebrate communities in streams monthly collected in the
Suyong Stream in the Suyong River in Korea from November 1992 to December
1994 for two years were used for training by the counterpropagation network. As
previously mentioned, a wide range of organic pollution was shown in the study
area. In order to check community compositions in a limited range in
environmental impacts, we selected the sampie sites of similar saprobic status in
slightly enriched zones (ß-mesosaprobity), YIG, YCK and YSC. General
descriptions of cornmunities and ecological assessment of water quality on the
Suyong River have been reported in Kwon and Chon (1993), Kang et al. (1995),
and Yoon and Chon (1996).
For training with the network, the cornmunity data were organized according to
the taxonomic hierarchy of Genus/Species, Family, Order, and Class and
functional groups. For the convenience of handling data, as well as for alleviating
problem of the difficulty in classification, Genus and Species were pooled to be
the same hierarchical level in this study. The total number of Genus and Species
used for training was 105, and that for Family, Order, Class, and functional
feeding groups was 48, 19, 7, and 5, respectively. The data were pre-processed as
previously mentioned.
Figs. 8.21a and 8.21b show examples of the actual field data used for training,
Genus/Species (X) and Family (Y) respectively. The values of normalized density
in each hierarchical group of benthic macroinvertebrates were expressed as
different levels of contour lines. The name of each taxa is not listed since the list
is too long and the names of taxa are not the issue in this chapter. Detailed
ecological implementation of the counterpropagation will be referred to Park et al.
(2001a). Fig. 8.21c shows trained results for Genus/Species (X') matching to the
input of Family data (Y). The overall conforrnation of community dynamics
appeared to be similar between X' and X, confirrning the input data were
effectively extracted through the network. GeneraHy the "averaging effect"
appeared: dominant groups such as Chironomidae, Tipulidae (Diptera), and
Hydropsychidae (Trichoptera) occurred consistently, while groups not frequently
coHected tended to disappear in the trained results (Fig. 8.21c). The reverse
process for producing Y' (Fig. 8.20) was also possible and, in general, the trained
results matched weH with the actual data.
165
preserved in the weights of the network. Y' served as trained results with the input
of X, while X' matching to the input of Y (Fig. 8.20).
Since the patterns of relation were established between X and Y in the network
through counterpropagation, the variables in the hierarchicallevels could mutually
respond to each other to new data sets. For example, if newly collected input data
set, X new ' were given to the trained network, they would produce new recognized
data set Y" corresponding to X. ew ' This could be compared with the actual field
data, X "" (Fig. 8.20).
The benthic macroinvertebrate communities in streams monthly collected in the
Suyong Stream in the Suyong River in Korea from November 1992 to December
1994 for two years were used for training by the counterpropagation network. As
previously mentioned, a wide range of organic pollution was shown in the study
area. In order to check community compositions in a limited range in
environmental impacts, we selected the sampie sites of similar saprobic status in
slightly enriched zones (ß-mesosaprobity), YIG, YCK and YSC. General
descriptions of cornmunities and ecological assessment of water quality on the
Suyong River have been reported in Kwon and Chon (1993), Kang et al. (1995),
and Yoon and Chon (1996).
For training with the network, the cornmunity data were organized according to
the taxonomic hierarchy of Genus/Species, Family, Order, and Class and
functional groups. For the convenience of handling data, as well as for alleviating
problem of the difficulty in classification, Genus and Species were pooled to be
the same hierarchical level in this study. The total number of Genus and Species
used for training was 105, and that for Family, Order, Class, and functional
feeding groups was 48, 19, 7, and 5, respectively. The data were pre-processed as
previously mentioned.
Figs. 8.21a and 8.21b show examples of the actual field data used for training,
Genus/Species (X) and Family (Y) respectively. The values of normalized density
in each hierarchical group of benthic macroinvertebrates were expressed as
different levels of contour lines. The name of each taxa is not listed since the list
is too long and the names of taxa are not the issue in this chapter. Detailed
ecological implementation of the counterpropagation will be referred to Park et al.
(2001a). Fig. 8.21c shows trained results for Genus/Species (X') matching to the
input of Family data (Y). The overall conforrnation of community dynamics
appeared to be similar between X' and X, confirrning the input data were
effectively extracted through the network. GeneraHy the "averaging effect"
appeared: dominant groups such as Chironomidae, Tipulidae (Diptera), and
Hydropsychidae (Trichoptera) occurred consistently, while groups not frequently
coHected tended to disappear in the trained results (Fig. 8.21c). The reverse
process for producing Y' (Fig. 8.20) was also possible and, in general, the trained
results matched weH with the actual data.