Chapter 8· Analysis of Stream Macroinvertebrate Communities
147
8.3
Prediction of Community Changes
8.3.1
Multilayer Perceptron with Time Delay
In the previous section, grouping of communities was presented. Through
grouping techniques, however, actual densities of taxa in community could not be
provided. Predietion of actual values in temporal development of communities,
however, is a major concern in ecosystem management. Especially in aquatic
ecosystems, where communities are easily affected by disturbances caused by
various natural and anthropogenie agents, it is important to predict how
communities would develop in response to changes in water quality. They would
develop either progressively with further disturbances, or regressively in recovery
from pollution (Sladecek 1979; Hellawell 1986). It is essential to prediet the
future level of community abundance for monitoring as weIl as for assessing
ecological status of the target ecosystem.
As previously mentioned, data for community dynamics, however, are complex
and difficult to analyze. In temporal patterning in ecology, artificial neural
networks has been effectively implemented in estimating time development of
populations and communities such as flowering and maturity of soybean
(Elizondo et al. 1994), algal bloom (e.g., Recknagel et al. 1997; Recknagel and
Wilson 2000), dynamics of animal population (Stankovski et al. 1998), and
grassland community development (Tan and Smeins 1994). However these
models were in most cases applied in static terms; the time of input and output
were the same. Recently, attention also has been given to dynamic neural
networks (e.g., Kung 1993; Giles et al. 1994; Haykin 1996). Wray and Green
(1994) reported that artificial neural networks could be utilized for investigating
parameters in non-linear dynamics, and dynamics of ecological data were
patterned and predieted by Boudjema and Chau (1996) on sets of univariate timeseries data of tree-ring thickness
To pattern relationships between different time events of community changes,
initially a well-known multilayer perceptron was utilized as a nonlinear predictor
with the backpropagation algorithm (Wray and Green 1994; Haykin 1994) (Fig.
8.12a). The architecture is multiplayer perception, however input and output data
were provided with time delay. The input vector is defined in terms of the past
sampies, X(t-l), X(t-2), ... , X(t-q), where q, prediction order, is the number of the
total delays. The current data, X(t), was given as matching output.
For field data, sampies were collected in a relatively short distance within 200
meters in the Yangjae Stream, a tributary of the Han River. The Yangjae Stream is
located in the Seoul metropolitan area, on the middle part of the Korean Peninsula,
and is highly polluted with poly-saprobity. In selecting data, attention was given
to taxa more frequently and abundantly collected while the data for rare species
147
8.3
Prediction of Community Changes
8.3.1
Multilayer Perceptron with Time Delay
In the previous section, grouping of communities was presented. Through
grouping techniques, however, actual densities of taxa in community could not be
provided. Predietion of actual values in temporal development of communities,
however, is a major concern in ecosystem management. Especially in aquatic
ecosystems, where communities are easily affected by disturbances caused by
various natural and anthropogenie agents, it is important to predict how
communities would develop in response to changes in water quality. They would
develop either progressively with further disturbances, or regressively in recovery
from pollution (Sladecek 1979; Hellawell 1986). It is essential to prediet the
future level of community abundance for monitoring as weIl as for assessing
ecological status of the target ecosystem.
As previously mentioned, data for community dynamics, however, are complex
and difficult to analyze. In temporal patterning in ecology, artificial neural
networks has been effectively implemented in estimating time development of
populations and communities such as flowering and maturity of soybean
(Elizondo et al. 1994), algal bloom (e.g., Recknagel et al. 1997; Recknagel and
Wilson 2000), dynamics of animal population (Stankovski et al. 1998), and
grassland community development (Tan and Smeins 1994). However these
models were in most cases applied in static terms; the time of input and output
were the same. Recently, attention also has been given to dynamic neural
networks (e.g., Kung 1993; Giles et al. 1994; Haykin 1996). Wray and Green
(1994) reported that artificial neural networks could be utilized for investigating
parameters in non-linear dynamics, and dynamics of ecological data were
patterned and predieted by Boudjema and Chau (1996) on sets of univariate timeseries data of tree-ring thickness
To pattern relationships between different time events of community changes,
initially a well-known multilayer perceptron was utilized as a nonlinear predictor
with the backpropagation algorithm (Wray and Green 1994; Haykin 1994) (Fig.
8.12a). The architecture is multiplayer perception, however input and output data
were provided with time delay. The input vector is defined in terms of the past
sampies, X(t-l), X(t-2), ... , X(t-q), where q, prediction order, is the number of the
total delays. The current data, X(t), was given as matching output.
For field data, sampies were collected in a relatively short distance within 200
meters in the Yangjae Stream, a tributary of the Han River. The Yangjae Stream is
located in the Seoul metropolitan area, on the middle part of the Korean Peninsula,
and is highly polluted with poly-saprobity. In selecting data, attention was given
to taxa more frequently and abundantly collected while the data for rare species