Chapter 7 . Stream Ecosystem Analysis
117
7.4.1
Correlations and Predictions of Environmental Variables
Data of physical and chemical water variables suffer from missing or wrong
values due to instrument failures. It is possible to model costly or difficult to
measure variables from cheaper and more precisely measured factors.
Applications are error detection and correction of data sets as weil as to fill gaps in
data bases with more reliable values.
LNNs and a relatively small MLP network with only five neurons in one
hidden layer were applied to model individual or groups of variables (e.g.
conductivity, pH-value, 02, NH 4 -N, and N0 3 -N). The resulting accuracy was high.
Networks with only one output neuron had improved generalisation performance
only for NH 4 -N. This agrees with the narrow correlation of the variables
conductivity, oxygen and pH among themselves and with the low correlation of
these variables with the nitrogen variables.
All water quality variables regarded here were predicted with good accuracy
(B>0.7) by a reduced number of network input variables (Borchardt et al 1997a).
For dimension reduction the two methods 'regression' and 'sensitivity analysis'
proved to be most suited (Dapper 1998).
7.4.2
Dependencies of Colonisation Patterns of Macro-Invertebrates on
Water Quality and Habitat Characteristics
The search for interdependencies, coherence or even causality of environmental
variables and a community is a fundamental challenge of ecology. The abundance
of species and the community assemblage depend on the environment. An
important aim in ecologically based research is the description of a species'
environmental requirements facilitating the prediction of a community under
given abiotic conditions.
7.4.2.1
Aquatic Insects in a Natural Stream, the Breitenbach
One goal was to predict the monthly species abundance of aquatic insects using a
'sliding time-window' on the data, first, of all environmental variables (discharge,
precipitation, and water temperature) of the actual and the 12 preceding months,
respectively and second, the species abundance of the 12 preceding months
(Dapper 1998). This period comprises the one-year life-cycle of these insects.
Second, the five most relevant predictors for ANNs were detected by a 'stepwise
linear regression' (SPSS; Brosius 1995), and further, predictors were identified by
neural sensitivity analysis (Dapper 1998). The resulting five-dimensional input
vectors were computed with 20 MLPs. Finally, generalization ability of the best
ANNs was visualized and compared to nets with all 51 input variables. The
resulting correlation coefficients of the models of 17 EPT species from the
Breitenbach are compared in Table 7.1.
117
7.4.1
Correlations and Predictions of Environmental Variables
Data of physical and chemical water variables suffer from missing or wrong
values due to instrument failures. It is possible to model costly or difficult to
measure variables from cheaper and more precisely measured factors.
Applications are error detection and correction of data sets as weil as to fill gaps in
data bases with more reliable values.
LNNs and a relatively small MLP network with only five neurons in one
hidden layer were applied to model individual or groups of variables (e.g.
conductivity, pH-value, 02, NH 4 -N, and N0 3 -N). The resulting accuracy was high.
Networks with only one output neuron had improved generalisation performance
only for NH 4 -N. This agrees with the narrow correlation of the variables
conductivity, oxygen and pH among themselves and with the low correlation of
these variables with the nitrogen variables.
All water quality variables regarded here were predicted with good accuracy
(B>0.7) by a reduced number of network input variables (Borchardt et al 1997a).
For dimension reduction the two methods 'regression' and 'sensitivity analysis'
proved to be most suited (Dapper 1998).
7.4.2
Dependencies of Colonisation Patterns of Macro-Invertebrates on
Water Quality and Habitat Characteristics
The search for interdependencies, coherence or even causality of environmental
variables and a community is a fundamental challenge of ecology. The abundance
of species and the community assemblage depend on the environment. An
important aim in ecologically based research is the description of a species'
environmental requirements facilitating the prediction of a community under
given abiotic conditions.
7.4.2.1
Aquatic Insects in a Natural Stream, the Breitenbach
One goal was to predict the monthly species abundance of aquatic insects using a
'sliding time-window' on the data, first, of all environmental variables (discharge,
precipitation, and water temperature) of the actual and the 12 preceding months,
respectively and second, the species abundance of the 12 preceding months
(Dapper 1998). This period comprises the one-year life-cycle of these insects.
Second, the five most relevant predictors for ANNs were detected by a 'stepwise
linear regression' (SPSS; Brosius 1995), and further, predictors were identified by
neural sensitivity analysis (Dapper 1998). The resulting five-dimensional input
vectors were computed with 20 MLPs. Finally, generalization ability of the best
ANNs was visualized and compared to nets with all 51 input variables. The
resulting correlation coefficients of the models of 17 EPT species from the
Breitenbach are compared in Table 7.1.