Chapter 8· Analysis of Stream Macroinvertebrate Communities
173
Most examples of groupings in exergy changes by the trained Kohonen
network had the corresponding characteristics of community dynamics in relation
to environments. Examples of the first group were TSD, TYB and TYZ for the
months of February, March and April 1998 (Listed months in the figure indicate
the month at the end of the segment.), respectively representing the neurons for
(2(x axis), O(y axis», (4,0) and (6,0) (Fig. 8.25b). In these periods, the densities of
Oligochaetae and Chironomidae tended to increase in March and decrease in April
consistently. Detailed ecological descriptions could be referred to Park et al.
(2001b).
This study demonstrated that artificial neural networks could be useful for
patterning changes in exergy. Although the sampie sites were located in relatively
close locations in the same river system, the changes in exergy appeared in diverse
patterns. The Kohonen network was able to separate different patterns in the time
development of exergy and demonstrated that the trends in exergy changes could
be useful for characteristically explaining community development and
environmental impacts (park et al., 2001b).
8.5
Summary and Conclusions
Artificial neural networks were implemented to pattern and predict benthic
macroinvertbrate community in streams.
Properties of self-organization,
adaptability and flexibility made of artificial neural networks were networks
useful for extracting information out of complex community data in various ways:
grouping for classification and ordination, prediction of community dynamics,
verification of environmental impacts, and revealing organizational aspects of
community.
Based on unsupervised leaming with the Kohonen network and ART, groupings
were efficiently conduced to classify and ordinate community data. The combined
networks of ART and Kohonen were further utilized to group community changes.
Short-time predictions of community dynamics were also possible through
temporal application of artificial neural networks. The time-delayed multi-Iayer
perceptron, and the partially and fully connected recurrent networks were able to
forecast the future level of community abundance given by the previous data as
input. The recurrent networks appeared to predict the temporal development of
communities better. The fully connected recurrent network also effectively
accommodated environmental factors, and the sensitivity analyses further revealed
the impact of environmental factors on community dynamics.
The organizational informatics were also patterned by artificial neural
networks. Patterns of relationships among different hierarchical levels in benthic
macroinvertebrate communities were effectively elucidated by the
counterpropagation network. The Kohonen network and multiplayer perceptron
were further utilized to characterize exergy, an integrative parameter indicating
thermodynamic information in community. Temporal exergy changes were
173
Most examples of groupings in exergy changes by the trained Kohonen
network had the corresponding characteristics of community dynamics in relation
to environments. Examples of the first group were TSD, TYB and TYZ for the
months of February, March and April 1998 (Listed months in the figure indicate
the month at the end of the segment.), respectively representing the neurons for
(2(x axis), O(y axis», (4,0) and (6,0) (Fig. 8.25b). In these periods, the densities of
Oligochaetae and Chironomidae tended to increase in March and decrease in April
consistently. Detailed ecological descriptions could be referred to Park et al.
(2001b).
This study demonstrated that artificial neural networks could be useful for
patterning changes in exergy. Although the sampie sites were located in relatively
close locations in the same river system, the changes in exergy appeared in diverse
patterns. The Kohonen network was able to separate different patterns in the time
development of exergy and demonstrated that the trends in exergy changes could
be useful for characteristically explaining community development and
environmental impacts (park et al., 2001b).
8.5
Summary and Conclusions
Artificial neural networks were implemented to pattern and predict benthic
macroinvertbrate community in streams.
Properties of self-organization,
adaptability and flexibility made of artificial neural networks were networks
useful for extracting information out of complex community data in various ways:
grouping for classification and ordination, prediction of community dynamics,
verification of environmental impacts, and revealing organizational aspects of
community.
Based on unsupervised leaming with the Kohonen network and ART, groupings
were efficiently conduced to classify and ordinate community data. The combined
networks of ART and Kohonen were further utilized to group community changes.
Short-time predictions of community dynamics were also possible through
temporal application of artificial neural networks. The time-delayed multi-Iayer
perceptron, and the partially and fully connected recurrent networks were able to
forecast the future level of community abundance given by the previous data as
input. The recurrent networks appeared to predict the temporal development of
communities better. The fully connected recurrent network also effectively
accommodated environmental factors, and the sensitivity analyses further revealed
the impact of environmental factors on community dynamics.
The organizational informatics were also patterned by artificial neural
networks. Patterns of relationships among different hierarchical levels in benthic
macroinvertebrate communities were effectively elucidated by the
counterpropagation network. The Kohonen network and multiplayer perceptron
were further utilized to characterize exergy, an integrative parameter indicating
thermodynamic information in community. Temporal exergy changes were