Chapter 7 . Stream Ecosystem Analysis
121
The inappropriate ratio of cases and variables, the incompletely known
ecological demands of species, and difficult analysis and training of networks with
multiple output neurons led to models with single target variables (Borchardt et al.
1997a, Schleiter et al. 1999; 2001).
One of the most widespread organisms even in anthropogenically altered
streams is Gammarus pu/ex. Abundance classes of this species were predicted
with GRNN (i.e. kerne! regression estimators) based on 14 stream features.
Genetic algorithms selected the 7 most relevant variables: minimum O 2 , the
maximum of water temperature, COD, and NH 4 -N, as weIl as channel
characteristics like stream bed and bank structure, and longitudinal development
of channel morphology. Importance (fitness) of predictors was calculated with a
leave-one-out cross validation, suitable for small data sets.
Ecological model reliability is high because the selected variables are in good
concordance with empirical knowledge of the species' distribution and related
water quality measures (i.e. high abundance at intermediate water temperatures,
high oxygen availability, and low pollution). Each figure (Fig. 7.5.) provides
information on the variability of two predictors, the remaining were kept constant
at their mean values (Obach 1998).
7.4.3
Bioindication
In the section above, we predicted community assemblage from a set of
environmental variables. Here, we model the magnitude of abiotic characteristics,
based on presence-absence and, in addition, abundance class data of indicator
organisms. This is a basic task of bioindication. Because the variables may change
in short time scales, biological indicators are adequate long-term probes for
environmental quality. These organisms require and thus represent a defined
environmental quality.
Initial experiments with species selected by MLPs from the anthropogenically
altered stream data modeIled, e.g. conductivity with a high precision based on
abundance classes of only 40, 20, 10 or even 5 benthic macro-invertebrates
(Schleiter et al. 1999; 2001). The reduction from initially 248 to only 5 taxa
decreased the amount of input data by about 98% (Borchardt et al. 1997b).
Furthermore, various methods for predictor detection (e.g. sensitivity analysis
on MLPs, linear regression analysis, genetic algorithms) were tested and the
generalisation performance of different network types (MLPs with an additional
special input layer, MFM, GRNN, LNN) was evaluated.
Eleven physical and chemical water quality measures, seven hydromorphological habitat characteristics, and three combined quality indices
(chemical and morphological water quality class, saprobic index) were modelled
using presence-absence and abundance data of 127 (out of 248) species present on
at least 10% of the 46 sites.
Conductivity was excellently modelled with both, 127 most frequent and
presence of 10 species (RMSE=2.8 and 4.5% of range respectively). Other
chemical variables were predicted with high accuracy, with the exception of NH 4 -
121
The inappropriate ratio of cases and variables, the incompletely known
ecological demands of species, and difficult analysis and training of networks with
multiple output neurons led to models with single target variables (Borchardt et al.
1997a, Schleiter et al. 1999; 2001).
One of the most widespread organisms even in anthropogenically altered
streams is Gammarus pu/ex. Abundance classes of this species were predicted
with GRNN (i.e. kerne! regression estimators) based on 14 stream features.
Genetic algorithms selected the 7 most relevant variables: minimum O 2 , the
maximum of water temperature, COD, and NH 4 -N, as weIl as channel
characteristics like stream bed and bank structure, and longitudinal development
of channel morphology. Importance (fitness) of predictors was calculated with a
leave-one-out cross validation, suitable for small data sets.
Ecological model reliability is high because the selected variables are in good
concordance with empirical knowledge of the species' distribution and related
water quality measures (i.e. high abundance at intermediate water temperatures,
high oxygen availability, and low pollution). Each figure (Fig. 7.5.) provides
information on the variability of two predictors, the remaining were kept constant
at their mean values (Obach 1998).
7.4.3
Bioindication
In the section above, we predicted community assemblage from a set of
environmental variables. Here, we model the magnitude of abiotic characteristics,
based on presence-absence and, in addition, abundance class data of indicator
organisms. This is a basic task of bioindication. Because the variables may change
in short time scales, biological indicators are adequate long-term probes for
environmental quality. These organisms require and thus represent a defined
environmental quality.
Initial experiments with species selected by MLPs from the anthropogenically
altered stream data modeIled, e.g. conductivity with a high precision based on
abundance classes of only 40, 20, 10 or even 5 benthic macro-invertebrates
(Schleiter et al. 1999; 2001). The reduction from initially 248 to only 5 taxa
decreased the amount of input data by about 98% (Borchardt et al. 1997b).
Furthermore, various methods for predictor detection (e.g. sensitivity analysis
on MLPs, linear regression analysis, genetic algorithms) were tested and the
generalisation performance of different network types (MLPs with an additional
special input layer, MFM, GRNN, LNN) was evaluated.
Eleven physical and chemical water quality measures, seven hydromorphological habitat characteristics, and three combined quality indices
(chemical and morphological water quality class, saprobic index) were modelled
using presence-absence and abundance data of 127 (out of 248) species present on
at least 10% of the 46 sites.
Conductivity was excellently modelled with both, 127 most frequent and
presence of 10 species (RMSE=2.8 and 4.5% of range respectively). Other
chemical variables were predicted with high accuracy, with the exception of NH 4 -