Chapter 7 . Stream Ecosystem Analysis
111
Analysis (PCA) leads to reduced computational efforts and easier to manage
models.
7.2.3
Artificial Neural Network Types
The software for all applied ANN simulations was developed by the Research
Group Neural Networks at the University Kassel (Germany). The suitability of the
following ANN types was tested: Kohonens self organising map as an example for
unsupervised learning networks and feed-forward networks such as conventional
and modified Multi-Layer-Perceptrons (MLP, Rumelhart et al. 1986; Senso
Neural Networks, Dapper 1998), General Regression Neural Networks (GRNN,
Specht 1991), Motoric Feature Maps (MFM, Ritter et al. 1994), Linear Neural
Networks (LNN) and also Radial Basis Function Neural Networks (RBF, Bishop
1995).
7.2.4
Dimension Reduction
Ecological variables are often correlated and thus data contain a certain amount of
redundant information. For example, many variables in aquatic environments
depend direcdy or indirecdy on the amount of oxygen available. Dimension
reduction finally resulted in:
1. an improvement of the proportion of the number of input dimensions and the
amount of available data with an improved generalization quality of ANNs
2. reduced computation effort, particularly during ANNs training
3. recognition ofrelevant predictors (to be compared also with expert knowledge)
4. simplified and easy to analyse models based on a manageable number of
variables.
Data compression by linear PCA factor analysis and botde-neck nets (Dapper
1998; Bishop 1995) provided independent variables. The computed factor
"pollution with organic matter" contains compressed information on P tot ' 02'
BOD" COD, and is interpreted as a major stressor in many anthropogenically
altered aquatic ecosystems. In contrast, selection of relevant predictors by
sensitivity analysis (Dapper 1998), stepwise methods, genetic algorithms
(Goldberg 1989) simply reduces the number of relevant variables to neglect
irrelevant and redundant information.
7.2.5
Quality Measures
Model accuracy is usually measured as the differences between predicted and
observed data. Usual error measures are the sum of squared errors (SQE), the
111
Analysis (PCA) leads to reduced computational efforts and easier to manage
models.
7.2.3
Artificial Neural Network Types
The software for all applied ANN simulations was developed by the Research
Group Neural Networks at the University Kassel (Germany). The suitability of the
following ANN types was tested: Kohonens self organising map as an example for
unsupervised learning networks and feed-forward networks such as conventional
and modified Multi-Layer-Perceptrons (MLP, Rumelhart et al. 1986; Senso
Neural Networks, Dapper 1998), General Regression Neural Networks (GRNN,
Specht 1991), Motoric Feature Maps (MFM, Ritter et al. 1994), Linear Neural
Networks (LNN) and also Radial Basis Function Neural Networks (RBF, Bishop
1995).
7.2.4
Dimension Reduction
Ecological variables are often correlated and thus data contain a certain amount of
redundant information. For example, many variables in aquatic environments
depend direcdy or indirecdy on the amount of oxygen available. Dimension
reduction finally resulted in:
1. an improvement of the proportion of the number of input dimensions and the
amount of available data with an improved generalization quality of ANNs
2. reduced computation effort, particularly during ANNs training
3. recognition ofrelevant predictors (to be compared also with expert knowledge)
4. simplified and easy to analyse models based on a manageable number of
variables.
Data compression by linear PCA factor analysis and botde-neck nets (Dapper
1998; Bishop 1995) provided independent variables. The computed factor
"pollution with organic matter" contains compressed information on P tot ' 02'
BOD" COD, and is interpreted as a major stressor in many anthropogenically
altered aquatic ecosystems. In contrast, selection of relevant predictors by
sensitivity analysis (Dapper 1998), stepwise methods, genetic algorithms
(Goldberg 1989) simply reduces the number of relevant variables to neglect
irrelevant and redundant information.
7.2.5
Quality Measures
Model accuracy is usually measured as the differences between predicted and
observed data. Usual error measures are the sum of squared errors (SQE), the