130
T.-S. Chon . Y.S. Park' I.-S. Kwak . E.Y. Cha
In this chapter, based on our experiences on filed data, we try to demonstrate
how artificial neural networks could be utilized as a general tool for analyzing and
predicting macroinvertebrate communities in streams. We try to elucidate
feasibility of artificial neural networks in grouping of communities as
cIassification and ordination, predicting multivariate community dynamics,
verifying environmental impacts, and revealing organizational aspects of
community.
8.1.1
Grouping Through Self-Organization
8.1.1.1
Static Grouping
Kohonen Network
Community data are complex and difficult to analyze as mentioned previously.
After collection of sampIes, the first step required is to have a comprehensive view
on the overall pattern of the collected community sampIes. This could be
generally conducted by cIassification or ordination through conventional statistical
methods. Through cIustering the communities were grouped in a hierarchical
manner dependent upon the degree of similarity among the sampled communities
(Ludwig and Reynolds 1988). Based on eigen analysis approach, associations
among sampIe communities (e.g., Q mode) or variables such as taxa and
environmental factors (e.g., R mode) could be revealed on principal factors
through ordination (Legendre and Legendre 1987; Ludwig and Reynolds 1988).
As mentioned previously, however, the conventional methods are generally
limited to linear data. Artificial neural networks is an alternative tool for
community cIassification, and the self-organizing mapping (SOM) is useful for
grouping non-linear data. The Kohonen network (Kohonen 1989) is one of the
most frequently used models for self-organizing, and the network has been
successfully implemented to patterning community data (e.g., Chon et al. 1996;
Foody 1999; Giraudel et al. 2000). The Kohonen network extracts information out
of multi-dimension data and maps onto the space of the reduced dimension (e.g., 2
or 3). In the Kohonen network, in this study, a linear array of M 2 artificial neurons
(i.e., computation nodes), with each neuron being represented as j (Fig. 8.2) is
arranged in two dimensions for the convenience of visual understanding (Chon et
al. 1996). Suppose a community data containing N species (Le., N dimensions),
and the density of species, i, is expressed as a vector xi' The vector xi is
considered to be an input layer to the Kohonen network. In the network each
neuron, j, is supposed to be connected to each node, i, of the input layer. The
connectivities are represented as weights, wij(t), adaptively changing at each
iteration of calculations, t. Initially the weights are randomly assigned in sm all
values. When the input vector is sent through the network, each neuron of the
network computes the summed distance between the weight and input as shown
below:
T.-S. Chon . Y.S. Park' I.-S. Kwak . E.Y. Cha
In this chapter, based on our experiences on filed data, we try to demonstrate
how artificial neural networks could be utilized as a general tool for analyzing and
predicting macroinvertebrate communities in streams. We try to elucidate
feasibility of artificial neural networks in grouping of communities as
cIassification and ordination, predicting multivariate community dynamics,
verifying environmental impacts, and revealing organizational aspects of
community.
8.1.1
Grouping Through Self-Organization
8.1.1.1
Static Grouping
Kohonen Network
Community data are complex and difficult to analyze as mentioned previously.
After collection of sampIes, the first step required is to have a comprehensive view
on the overall pattern of the collected community sampIes. This could be
generally conducted by cIassification or ordination through conventional statistical
methods. Through cIustering the communities were grouped in a hierarchical
manner dependent upon the degree of similarity among the sampled communities
(Ludwig and Reynolds 1988). Based on eigen analysis approach, associations
among sampIe communities (e.g., Q mode) or variables such as taxa and
environmental factors (e.g., R mode) could be revealed on principal factors
through ordination (Legendre and Legendre 1987; Ludwig and Reynolds 1988).
As mentioned previously, however, the conventional methods are generally
limited to linear data. Artificial neural networks is an alternative tool for
community cIassification, and the self-organizing mapping (SOM) is useful for
grouping non-linear data. The Kohonen network (Kohonen 1989) is one of the
most frequently used models for self-organizing, and the network has been
successfully implemented to patterning community data (e.g., Chon et al. 1996;
Foody 1999; Giraudel et al. 2000). The Kohonen network extracts information out
of multi-dimension data and maps onto the space of the reduced dimension (e.g., 2
or 3). In the Kohonen network, in this study, a linear array of M 2 artificial neurons
(i.e., computation nodes), with each neuron being represented as j (Fig. 8.2) is
arranged in two dimensions for the convenience of visual understanding (Chon et
al. 1996). Suppose a community data containing N species (Le., N dimensions),
and the density of species, i, is expressed as a vector xi' The vector xi is
considered to be an input layer to the Kohonen network. In the network each
neuron, j, is supposed to be connected to each node, i, of the input layer. The
connectivities are represented as weights, wij(t), adaptively changing at each
iteration of calculations, t. Initially the weights are randomly assigned in sm all
values. When the input vector is sent through the network, each neuron of the
network computes the summed distance between the weight and input as shown
below: