266
H. Wilson . F. Recknagel
models have the potential to have lower prediction bias than existing inductive
approaches.
French and Recknagel (1994) were among the first applications to demonstrate
the potential of ANNs for modelling phytoplankton growth dynamics. These
authors demonstrated that feedforward ANNs, where variables describing water
quality were used as inputs and cell counts of the phytoplankton species of interest
formed the outputs, were able to predict algal bloom events when trained with a
time series of observations from a German lake.
Since then, applications have been demonstrated for the Murray River
(Australia) (Maier et al. 1998), Lake Kasumigaura (Japan), Lake Biwa (Japan),
Lake Tuusulanjarvi (Finland), and the Darling River (Australia) (Recknagel et al.
1997), and two Turkish reservoirs (Karul et al 2000). Scardi (1996) developed an
ANN application for estimating phytoplankton production in a estuarine
environment (Chesapeake Bay and Delaware Bay on the US east coast).
Recknagel and Wilson (2000) claim that, in general, ANN approaches to algal
bloom prediction have produced models with greater predictive accuracy than has
been achieved with conventional deductive and inductive approaches.
14.2
Issues to be Addressed by ANN Aigal Bloom Models
14.2.1
Input Layer Design
In general, input layer designs for preceding ANN models have considered the
following factors:
1. Theories regarding the causes of phytoplankton growth.
2. A vailability of variables in historical water quality monitoring databases.
3. Serial correlation of time series data.
4. The results of data analysis such as a visual or statistical analysis or a-priori
ANN studies.
In practice, input layer designs have been implemented in an ad-hoc nature
depending on the specific characteristics of the case study. In particular, the
second point outlined above (i.e. data availability) is crucial in determining the
scope of potential inputs. In this study, it is hypothesised that it is possible to
achieve a predictive ANN model with a more generic input layer design that is
compatible with a wide range of databases. Such a generic design will have the
following benefits:
• Comparisons between models developed for different lakes using established
elucidation techniques for ANNs such as sensitivity analyses will be more
meaningful.
• It provides scope for increasingly general ANN eutrophication models trained
with data aggregated from many lakes or classes of lakes.
H. Wilson . F. Recknagel
models have the potential to have lower prediction bias than existing inductive
approaches.
French and Recknagel (1994) were among the first applications to demonstrate
the potential of ANNs for modelling phytoplankton growth dynamics. These
authors demonstrated that feedforward ANNs, where variables describing water
quality were used as inputs and cell counts of the phytoplankton species of interest
formed the outputs, were able to predict algal bloom events when trained with a
time series of observations from a German lake.
Since then, applications have been demonstrated for the Murray River
(Australia) (Maier et al. 1998), Lake Kasumigaura (Japan), Lake Biwa (Japan),
Lake Tuusulanjarvi (Finland), and the Darling River (Australia) (Recknagel et al.
1997), and two Turkish reservoirs (Karul et al 2000). Scardi (1996) developed an
ANN application for estimating phytoplankton production in a estuarine
environment (Chesapeake Bay and Delaware Bay on the US east coast).
Recknagel and Wilson (2000) claim that, in general, ANN approaches to algal
bloom prediction have produced models with greater predictive accuracy than has
been achieved with conventional deductive and inductive approaches.
14.2
Issues to be Addressed by ANN Aigal Bloom Models
14.2.1
Input Layer Design
In general, input layer designs for preceding ANN models have considered the
following factors:
1. Theories regarding the causes of phytoplankton growth.
2. A vailability of variables in historical water quality monitoring databases.
3. Serial correlation of time series data.
4. The results of data analysis such as a visual or statistical analysis or a-priori
ANN studies.
In practice, input layer designs have been implemented in an ad-hoc nature
depending on the specific characteristics of the case study. In particular, the
second point outlined above (i.e. data availability) is crucial in determining the
scope of potential inputs. In this study, it is hypothesised that it is possible to
achieve a predictive ANN model with a more generic input layer design that is
compatible with a wide range of databases. Such a generic design will have the
following benefits:
• Comparisons between models developed for different lakes using established
elucidation techniques for ANNs such as sensitivity analyses will be more
meaningful.
• It provides scope for increasingly general ANN eutrophication models trained
with data aggregated from many lakes or classes of lakes.