Chapter 8· Analysis of Stream Macroinvertebrate Communities
143
Fig. 8.9 demonstrates the possibility of a large-seale grouping. A two
dimensional map was produeed after the training with the Kohonen network on the
sampled eommunities of benthic maeroinvertebrates eolleeted in streams of South
Korea for twelve years from 1984, which had been published in the 14 papers
from 23 tributaries in the major river system (Chon et al. 2000a). The
eommunities appeared to be grouped aeeording to the river systems (e.g., Han
River, Somjin River, ete). The Han River have been most extensively surveyed,
and the eommunities eollected from the Han River were further sub-grouped
aeeording to the degree of pollution on the map (Chon et al. 2000a). If newly
sampled eommunities are given to the network, they would be eonveniently
reeognized as mentioned previously, These proeesses of visual presentation of
large-seale data and reeognition of new data sets eould be efficiently used for
diagnosing eeologieal status of the surveyed area for a long time for sustainable
eeosystem management.
8.2.2
Grouping Community Changes
Sinee eommunity develops on time domain, either in stressful or in favorable
eonditions, groupings of eommunity "ehanges" is neeessary for the eomprehensive
perspeetive on stream eeosystems. Espeeially in aquatic eeosystems, where
eommunities are easily affected by disturbanees eaused by various natural and
anthropogenie agents (Sladeeek 1979; Hellawell 1986), it is important to pattern
eommunity ehanges in response to disturbanees. However, it is not an easy task to
classify eommunity ehanges, and fewer studies have been eondueted on this topie.
Legendre et al. (1985) and Legendre (1987) diseussed classifying eommunities in
temporal domain, utilizing ordination and segmentation techniques in multivariate
data series. Similar to the ease of static classifieation, the eonventional statistical
methods however, are limited in analyzing eomplex data for eommunity ehanges.
Artificial neural networks, however, eould be further implemented to grouping
"ehanges" in eommunity. A eombined model of artificial neural networks was
utilized in this ease. The sampled data for eommunity ehanges were trained with
the two proeesses through Adaptive Resonanee Theory (ART; Carpenter and
Grossberg 1987) and the Kohonen network (Kohonen 1989). The sehematie
diagram of the eombined network is presented in Figs. 8.7a and 8.7b. Initiallyall
the eommunity data for one-time sampling from field were trained by ART(Fig.
8.7b) as mentioned previously and subsequently the weight produeed by ART
were mapped by the Kohonen network (Fig. 8.7a) (Chon et al. 2000e).
As explained before, weights produeed by ART preserved eonformational
eharaeteristie of input data for eaeh sampling time through training (Fig. 8.8a).
Sinee one-month sampling data were aeeordingly eharaeterized by weights in the
Kohonen network (Fig. 8.8b), it was supposed that, if the weights for previous
months were appended to the target month, they would also effieiently represent
ehanges in eommunity during the specified period. The weights trained for one
month in ART were eombined sequentially for a eertain period, and were given to
the Kohonen network as inputs. If eommunity ehanges in three months are to be
143
Fig. 8.9 demonstrates the possibility of a large-seale grouping. A two
dimensional map was produeed after the training with the Kohonen network on the
sampled eommunities of benthic maeroinvertebrates eolleeted in streams of South
Korea for twelve years from 1984, which had been published in the 14 papers
from 23 tributaries in the major river system (Chon et al. 2000a). The
eommunities appeared to be grouped aeeording to the river systems (e.g., Han
River, Somjin River, ete). The Han River have been most extensively surveyed,
and the eommunities eollected from the Han River were further sub-grouped
aeeording to the degree of pollution on the map (Chon et al. 2000a). If newly
sampled eommunities are given to the network, they would be eonveniently
reeognized as mentioned previously, These proeesses of visual presentation of
large-seale data and reeognition of new data sets eould be efficiently used for
diagnosing eeologieal status of the surveyed area for a long time for sustainable
eeosystem management.
8.2.2
Grouping Community Changes
Sinee eommunity develops on time domain, either in stressful or in favorable
eonditions, groupings of eommunity "ehanges" is neeessary for the eomprehensive
perspeetive on stream eeosystems. Espeeially in aquatic eeosystems, where
eommunities are easily affected by disturbanees eaused by various natural and
anthropogenie agents (Sladeeek 1979; Hellawell 1986), it is important to pattern
eommunity ehanges in response to disturbanees. However, it is not an easy task to
classify eommunity ehanges, and fewer studies have been eondueted on this topie.
Legendre et al. (1985) and Legendre (1987) diseussed classifying eommunities in
temporal domain, utilizing ordination and segmentation techniques in multivariate
data series. Similar to the ease of static classifieation, the eonventional statistical
methods however, are limited in analyzing eomplex data for eommunity ehanges.
Artificial neural networks, however, eould be further implemented to grouping
"ehanges" in eommunity. A eombined model of artificial neural networks was
utilized in this ease. The sampled data for eommunity ehanges were trained with
the two proeesses through Adaptive Resonanee Theory (ART; Carpenter and
Grossberg 1987) and the Kohonen network (Kohonen 1989). The sehematie
diagram of the eombined network is presented in Figs. 8.7a and 8.7b. Initiallyall
the eommunity data for one-time sampling from field were trained by ART(Fig.
8.7b) as mentioned previously and subsequently the weight produeed by ART
were mapped by the Kohonen network (Fig. 8.7a) (Chon et al. 2000e).
As explained before, weights produeed by ART preserved eonformational
eharaeteristie of input data for eaeh sampling time through training (Fig. 8.8a).
Sinee one-month sampling data were aeeordingly eharaeterized by weights in the
Kohonen network (Fig. 8.8b), it was supposed that, if the weights for previous
months were appended to the target month, they would also effieiently represent
ehanges in eommunity during the specified period. The weights trained for one
month in ART were eombined sequentially for a eertain period, and were given to
the Kohonen network as inputs. If eommunity ehanges in three months are to be