124
I.M. Schleiter . M. Obach . R. Wagner' H. Werner' H.-H. Schmidt
. D. Borchardt
We focussed on single network outputs because models with multiple outputs
were more difficult to train and interpret. The quality of the models was described
by tbe comparison of an ANN model error with trivial (persistence, naIve
prediction) or easy to calculate (linear model, long-term mean) prediction errors.
Our preferred error measure was the RMSE of [O,I]-scaled data, the ratio of the
expected mean error from the range of the output variable's values. However, the
RMSE in original units also provides valuable information to ecologists. These
global errors do not necessarily estimate local reliability, wbich depends on e.g.
the local variability of the output variable and data density. Tbe combination of
SOM and RBF networks (RBFSOM) combines good prediction properties on weIl
supported input data with a warning function, if the particular input is not
supported by training data, and hence the output information may not be valid. VMatrix, Sammon map, visualization of input vector component planes and the
display of neuron activities on the SOM codebook vectors are some graphical
representation possibilities of the unsupervised trained SOMs. Feedforward
network outputs can be displayed as 3D surfaces. The example of RBFSOM
shows that it is profitable to connect different ANNs to a hybrid network in order
to combine their special capabilities. Combined with GRNN for input relevance
detection RBFSOMs become capable, efficient and transparent prediction tools.
We applied ANNs on data sets with environmental variables and communities
to model interrelations in pristine and anthropogenically altered streams. The
results confirmed interrelations between colonisation patterns of benthic macroinvertebrates, chemical and hydro-morphological habitat characteristics in lotic
ecosystems. In a pristine stream discharge predominantly determined species
abundances and community assemblage. Water temperature and other variables
had smaller effects. High determination coefficients on test data were surprising,
because of large proportions of trivial zero predictions. This inspired the
application of more adequate tools and error measures.
On anthopogenic altered streams similarities of sampling sites as weIl as of
individual variables were visualized with SOMs. Extraordinary sites were detected
with Sammon maps and V-Matrix displays. Component planes were useful to
analyze the responsible factors and to designate correlations among variables.
Furthermore, leaps in the trajectories of individual streams on a map indicated
abrupt changes of water quality.
Dependencies of species on their environment were modelled with ANNs
circumscribing ecological niches. The effects of two important abiotic factors on
the abundance classes of Gammarus pulex were displayed in 3D surface figures.
A reversed task is the prediction of environmental features from communities.
Even information on the presence instead of the abundance of selected sub sets of
macro-invertebrates was adequate to perform bioindication of e.g. conductivity,
oxygen and water quality classes with sufficient accuracy.
However, the last step in data analysis is the interpretation and the check of
plausibility and the interpretation based on expert knowledge. ANNs have been
used for more than a decade in ecology, but there are still many research fields
left. Further generalisation of the models beyond the area of Bunter Sandstone in
Central Germany is the main future task beside the quantitative and qualitative
extension of the data base.
I.M. Schleiter . M. Obach . R. Wagner' H. Werner' H.-H. Schmidt
. D. Borchardt
We focussed on single network outputs because models with multiple outputs
were more difficult to train and interpret. The quality of the models was described
by tbe comparison of an ANN model error with trivial (persistence, naIve
prediction) or easy to calculate (linear model, long-term mean) prediction errors.
Our preferred error measure was the RMSE of [O,I]-scaled data, the ratio of the
expected mean error from the range of the output variable's values. However, the
RMSE in original units also provides valuable information to ecologists. These
global errors do not necessarily estimate local reliability, wbich depends on e.g.
the local variability of the output variable and data density. Tbe combination of
SOM and RBF networks (RBFSOM) combines good prediction properties on weIl
supported input data with a warning function, if the particular input is not
supported by training data, and hence the output information may not be valid. VMatrix, Sammon map, visualization of input vector component planes and the
display of neuron activities on the SOM codebook vectors are some graphical
representation possibilities of the unsupervised trained SOMs. Feedforward
network outputs can be displayed as 3D surfaces. The example of RBFSOM
shows that it is profitable to connect different ANNs to a hybrid network in order
to combine their special capabilities. Combined with GRNN for input relevance
detection RBFSOMs become capable, efficient and transparent prediction tools.
We applied ANNs on data sets with environmental variables and communities
to model interrelations in pristine and anthropogenically altered streams. The
results confirmed interrelations between colonisation patterns of benthic macroinvertebrates, chemical and hydro-morphological habitat characteristics in lotic
ecosystems. In a pristine stream discharge predominantly determined species
abundances and community assemblage. Water temperature and other variables
had smaller effects. High determination coefficients on test data were surprising,
because of large proportions of trivial zero predictions. This inspired the
application of more adequate tools and error measures.
On anthopogenic altered streams similarities of sampling sites as weIl as of
individual variables were visualized with SOMs. Extraordinary sites were detected
with Sammon maps and V-Matrix displays. Component planes were useful to
analyze the responsible factors and to designate correlations among variables.
Furthermore, leaps in the trajectories of individual streams on a map indicated
abrupt changes of water quality.
Dependencies of species on their environment were modelled with ANNs
circumscribing ecological niches. The effects of two important abiotic factors on
the abundance classes of Gammarus pulex were displayed in 3D surface figures.
A reversed task is the prediction of environmental features from communities.
Even information on the presence instead of the abundance of selected sub sets of
macro-invertebrates was adequate to perform bioindication of e.g. conductivity,
oxygen and water quality classes with sufficient accuracy.
However, the last step in data analysis is the interpretation and the check of
plausibility and the interpretation based on expert knowledge. ANNs have been
used for more than a decade in ecology, but there are still many research fields
left. Further generalisation of the models beyond the area of Bunter Sandstone in
Central Germany is the main future task beside the quantitative and qualitative
extension of the data base.