Chapter 8· Analysis of Stream Macroinvertebrate Communities
159
c10ser to field data in July 1997 than the predictions with the community data
alone as previously mentioned (Fig. 8.16a). In July 1997, the correlation
coefficient between the predicted and the matching field data was 0.94 when the
network was trained with the input of the community plus environmental data. In
contrast, the correlation coefficient was 0.55 when only the community data were
given to the network. High densities of chironomids were present in the Yangjae
Stream during the flooding season of the training period. This relationship
between precipitation and the occurrence of chironomids was reflected in the
network trained with the community and the environmental data. As shown in Fig.
16a, the predicted densities of Chironomidae and Chironomus, trained with the
community and the environmental data, for example, weIl matched the field data,
while the network trained only with the community data consistently
underestimated the densities of Chironomidae and Chironomus (Chon et al.
2000b).
However, predictions based on environmental data were not always superior to
predictions based only with the community data. In seasons without strong
environmental effects, the predictability seems to be similar between these two
types of training. For example, in the prediction of the community data of
November 1997 in winter (Fig. 8.16b), the respective Pearson's correlation
coefficients were 0.60 and 0.59 for the training based on the community plus
environmental data, and that based only on the community data. During the
training period there were no serious environmental disturbances occurred in this
case. When the effects of relationships on training factors were complex
environmental, influences may be negatively related. The detailed explanation
could be referred to Chon et al. (2001).
It is also useful to investigate causality relationships that how environmental
factors influence community dynamics through the sensitivity analysis. We
conducted sensitivity analyses on the recurrent network so that the variational
impact of environmental factors could be revealed on community changes.
Variation around the mean value (ranging +50 % and -50 %) was provided to
each input value of the environmental variables. For the simplicity of the
sensitivity analysis of this recurrent neural network, variation term was given only
to the input of the last month.
In terms of different training periods and selected taxa, the sensitivity tests
effectively showed important environmental variables in determining community
changes. For the data of July 1997, when the flooding occurred during this period
used for training, all four environmental variables of organic matter, depth,
velocity and substrates (smaller than 5 mm), caused a large variation of
communities in a wide range (Fig. 8.17a). For the data ofNovember 1998, wh ich
had no strong environmental effects, in contrast, all environmental variables did
not produce variations in community dynamics (Fig. 8.17b).
159
c10ser to field data in July 1997 than the predictions with the community data
alone as previously mentioned (Fig. 8.16a). In July 1997, the correlation
coefficient between the predicted and the matching field data was 0.94 when the
network was trained with the input of the community plus environmental data. In
contrast, the correlation coefficient was 0.55 when only the community data were
given to the network. High densities of chironomids were present in the Yangjae
Stream during the flooding season of the training period. This relationship
between precipitation and the occurrence of chironomids was reflected in the
network trained with the community and the environmental data. As shown in Fig.
16a, the predicted densities of Chironomidae and Chironomus, trained with the
community and the environmental data, for example, weIl matched the field data,
while the network trained only with the community data consistently
underestimated the densities of Chironomidae and Chironomus (Chon et al.
2000b).
However, predictions based on environmental data were not always superior to
predictions based only with the community data. In seasons without strong
environmental effects, the predictability seems to be similar between these two
types of training. For example, in the prediction of the community data of
November 1997 in winter (Fig. 8.16b), the respective Pearson's correlation
coefficients were 0.60 and 0.59 for the training based on the community plus
environmental data, and that based only on the community data. During the
training period there were no serious environmental disturbances occurred in this
case. When the effects of relationships on training factors were complex
environmental, influences may be negatively related. The detailed explanation
could be referred to Chon et al. (2001).
It is also useful to investigate causality relationships that how environmental
factors influence community dynamics through the sensitivity analysis. We
conducted sensitivity analyses on the recurrent network so that the variational
impact of environmental factors could be revealed on community changes.
Variation around the mean value (ranging +50 % and -50 %) was provided to
each input value of the environmental variables. For the simplicity of the
sensitivity analysis of this recurrent neural network, variation term was given only
to the input of the last month.
In terms of different training periods and selected taxa, the sensitivity tests
effectively showed important environmental variables in determining community
changes. For the data of July 1997, when the flooding occurred during this period
used for training, all four environmental variables of organic matter, depth,
velocity and substrates (smaller than 5 mm), caused a large variation of
communities in a wide range (Fig. 8.17a). For the data ofNovember 1998, wh ich
had no strong environmental effects, in contrast, all environmental variables did
not produce variations in community dynamics (Fig. 8.17b).