122
I.M. Schleiter . M. Obach· R. Wagner· H. Werner· H.-H. Schmidt
. D. Borchardt
N and N tol (RMSE=14.5% of range). Data sets of a reduced number of species
improved model quality for e.g. N0 3 -N, pH, 02' BOD s . These species can be
successfully used for bioindication.
For the majority of variables, presence-absence data provided better models
compared with abundance dasses. Predietability of hydro-morphological variables
was sufficient only for the discharge regime and substrate diversity.
Morphological quality dass (Hessisches Ministerium für Umwelt, Landwirtschaft
und Forsten 2000) was not predictable by macro-zoobenthos with the required
accuracy. Saprobic index was estimated weil based on only ten out of 60 indieator
species (RescaIRMSE=O.I).
Prediction of chemieal water quality dass (LA WA 1998) was easy and
efficient. A linear regression model with presence-absence data of three predictors
was sufficient. However, this predietion was not accurate on all test sites, because
there are many reasons for a species' absence beyond the water quality. Only
when other variables were verified by expert knowledge (e.g. intact stream
morphology, community with several species), was the model reliable (Schleiter et
al. 2001).
Depending on the network type and the selection method, different species
groups were chosen for most, even for correlated output variables, and verified the
results of Schleiter et al. (1999). Some species were useful for the prediction of
several targets (e.g. Gammarus pulex: conductivity, saprobie index, oxygen, P lOl '
N0 3 -N; Chironomus thummi: conductivity, chemical water quality dass, BOD s '
N0 3 -N), whereas others appeared in only one model. Models with species groups
selected here may not be generalized due to restrictions in the basic data set
(narrow geographieal region and limited abiotie gradients). Probably, other
species not occuring in the DIN table (Friedrieh 1990) may be suitable
bioindicators for the saprobity in small streams. The results obtained need further
validation based on additional data and expert knowledge.
The selection of relevant variables and the use of presence-absence data
provided less complex, easier to understand and handle models and a drastic
reduction of computational effort. This allowed an increased number of repetitions
to provide more relevant results for generalisation. However, one problem is small
training and test data sets. A low generalisation error may be accidental.
7.5
Assessment of Model Quality and Visualisation
Possibilities: Hybrid Networks
A disadvantage of most ANNs are the complicated, difficult to comprehend
internal network processes so that in many cases the neural networks are
considered as black boxes.
Usually, the quality of neural network outputs is measured based on the
difference between observed and predicted values. For quality assessments of the
network outputs, the error value alone is not sufficient because it does not provide
any information concerning the reason of the errOf. Large differences can have
I.M. Schleiter . M. Obach· R. Wagner· H. Werner· H.-H. Schmidt
. D. Borchardt
N and N tol (RMSE=14.5% of range). Data sets of a reduced number of species
improved model quality for e.g. N0 3 -N, pH, 02' BOD s . These species can be
successfully used for bioindication.
For the majority of variables, presence-absence data provided better models
compared with abundance dasses. Predietability of hydro-morphological variables
was sufficient only for the discharge regime and substrate diversity.
Morphological quality dass (Hessisches Ministerium für Umwelt, Landwirtschaft
und Forsten 2000) was not predictable by macro-zoobenthos with the required
accuracy. Saprobic index was estimated weil based on only ten out of 60 indieator
species (RescaIRMSE=O.I).
Prediction of chemieal water quality dass (LA WA 1998) was easy and
efficient. A linear regression model with presence-absence data of three predictors
was sufficient. However, this predietion was not accurate on all test sites, because
there are many reasons for a species' absence beyond the water quality. Only
when other variables were verified by expert knowledge (e.g. intact stream
morphology, community with several species), was the model reliable (Schleiter et
al. 2001).
Depending on the network type and the selection method, different species
groups were chosen for most, even for correlated output variables, and verified the
results of Schleiter et al. (1999). Some species were useful for the prediction of
several targets (e.g. Gammarus pulex: conductivity, saprobie index, oxygen, P lOl '
N0 3 -N; Chironomus thummi: conductivity, chemical water quality dass, BOD s '
N0 3 -N), whereas others appeared in only one model. Models with species groups
selected here may not be generalized due to restrictions in the basic data set
(narrow geographieal region and limited abiotie gradients). Probably, other
species not occuring in the DIN table (Friedrieh 1990) may be suitable
bioindicators for the saprobity in small streams. The results obtained need further
validation based on additional data and expert knowledge.
The selection of relevant variables and the use of presence-absence data
provided less complex, easier to understand and handle models and a drastic
reduction of computational effort. This allowed an increased number of repetitions
to provide more relevant results for generalisation. However, one problem is small
training and test data sets. A low generalisation error may be accidental.
7.5
Assessment of Model Quality and Visualisation
Possibilities: Hybrid Networks
A disadvantage of most ANNs are the complicated, difficult to comprehend
internal network processes so that in many cases the neural networks are
considered as black boxes.
Usually, the quality of neural network outputs is measured based on the
difference between observed and predicted values. For quality assessments of the
network outputs, the error value alone is not sufficient because it does not provide
any information concerning the reason of the errOf. Large differences can have