Friedrich Recknagel

Artificial neural network approach for modelling and prediction of algal blooms

Abstract Following a comparison of current alternative approaches for modelling and prediction of algal blooms, artificial neural networks are introduced and applied as a new, promising model type. The neural network applications were developed and validated by limnological time-series from four different freshwater systems. The water-specific time-series comprised cell numbers or biomass of the ten dominating algae species as observed over up to twelve years and the measured environmental driving variables. The resulting predictions on succession, timing and magnitudes of algal species indicate that artificial neural networks can fit the complexity and nonlinearity of ecological phenomena apparently to a high degree.

Applications of machine learning to ecological modelling

The paper provides a summary of paper presentations at the 2nd International Conference on Applications of Machine Learning to Ecological Modelling and a preview of forthcoming developments in this area. Artificial neural networks were demonstrated to be very useful for nonlinear ordination and visualization of ecological data by Kohonen networks, and ecological time-series modelling by recurrent networks. Genetic algorithms proved to be very innovative for hybridizing deductive models, and evolving predictive rules, process equations and parameters. Newly emerging adaptive agents provide a novel framework for the discovery and forecasting of emergent ecosystem structures and behaviours in response to environmental changes.

ANNA - Artificial neural network model for predicting species abundance and succession of blue-green algae

Predictive potential of deductive and inductivephytoplankton models are compared regarding theirusefulness for forecasting and control of harmfulalgal blooms. While applications of deductive modelsstill seem to be restricted by lack of knowledge, ad hocinductive models sometimes prove to bestraightforward and useful. The inductive neuralnetwork model ANNA is documented by means of anapplication to Lake Kasumigaura, Japan. ANNA wasvalidated for five blue-green algae species wherepredictive accuracy has improved with increased eventand time resolution of training data. A scenarioanalysis on species succession has demonstrated thepotential of ANNA for hypothesis testing. Finally,implications for use of ANNA for operational algalbloom control are discussed.

Prediction and elucidation of phytoplankton dynamics in the Nakdong River (Korea) by means of a recurrent artificial neural network

A recurrent artificial neural network was used for time series modelling of phytoplankton dynamics in the hypertrophic Nakdong River system. The model considered meteorological, hydrological and limnological parameters as input variables and chl. a concentration as output variable. It was trained and validated by means of a complex database measured from 1994 to 1998 at a study site 27 km upstream of the river mouth. The validation results for 1994 indicated that the recurrent training algorithm and a 3 days time lag of input data predict reasonably accurate the timing and magnitudes of chl. a. A comprehensive sensitivity analysis of the model revealed relationships between seasons, specific input variables and chl. a that correspond well with theoretical assumptions and literature findings.

Predicting chlorophyll-a in freshwater lakes by hybridising process-based models and genetic algorithms

This paper describes the application of several machine learning techniques to modify a process-based difference equation. The original process equation was developed to model phytoplankton abundance based on measured limnological and climate variables. A genetic algorithm is shown to be capable of calibrating the constants of the process model, based on the data describing a lake environment. The resulting process model has a significantly improved performance based on unseen test data. A symbolic genetic algorithm is then applied to the process model to evolve new expressions for the grazing term of the equation. The results indicate that this approach can be used to explore new process formulations and to improve the generalisation and predictive response of process models.

Towards a generic artificial neural network model for dynamic predictions of algal abundance in freshwater lakes

A generalised architecture of a feedforward ANN for the prediction of algal abundance is suggested. It simplifies practical model applications, rationalises data collection and preprocessing, improves model validity, and enables meaningful comparison of ANN models between lakes. The generic ANN model considers the key driving variables of algal growth such as phosphorous, nitrogen, underwater light and water temperature as input nodes and predicts algal species abundance or biomass as output. Two model structures were used; one for same-day and one for 30-days ahead predictions of algal abundance. ANN models with and without hidden layers were compared to determine the impact of the addition of non-linear processing capabilities on model performance. A bootstrap aggregation method was found to reduce test set prediction error and to mitigate the effects of overfitting. The model was validated by means of time-series data from six different freshwater lakes.

Response of stream macroinvertebrates to changes in salinity and the development of a salinity index

Many streams and wetlands have been affected by increasing salinity, leading to significant changes in flora and fauna. The study investigates relationships between macroinvertebrate taxa and conductivity levels (µ Sc m −1 ) in Queensland stream systems. The analysed dataset contained occurrence patterns of frequently found macroinvertebrate taxa from edge (2580 samples) and riffle (1367 samples) habitats collected in spring and autumn over 8 years. Sensitivity analysis with predictive artificial neural network models and the taxon-specific mean conductivity values were used to assign a salinity sensitivity score (SSS) to each taxon (1—very tolerant, 5— tolerant, 10—sensitive). Salinity index (SI) based on the cumulative SSS was proposed as a measurement of change in macroinvertebrate communities caused by salinity increase. Changes in macroinvertebrate communities were observed at relatively low salinities, with SI rapidly decreasing to ∼800-1000 µ Sc m −1 and decreasing further at a slower rate. Natural variability and water quality factors were ruled out as potential primary causes of the observed changes by using partial canonical correspondence analysis and subsets of the data with only good water quality.

Predicting eutrophication effects in the Burrinjuck Reservoir (Australia) by means of the deterministic model SALMO and the recurrent neural network model ANNA

Abstract Two modelling paradigms were applied to the prediction of phytoplankton abundance in the Burrinjuck Reservoir: the deductive model SALMO and the inductive model ANNA. While SALMO is driven by process-based differential equations, the model ANNA is designed as recurrent feedforward neural network trained by time series data. Predictions of chlorophyll-a for the years 1979–1982 by both models were validated by means of measured data. Results showed that SALMO is able to predict annual average trends not only of chlorophyll-a but other chemical and biological state variables as well. It supports decision making by evaluating alternative scenarios for strategic eutrophication control. The model ANNA achieved reasonable accuracy in predicting timing and magnitudes of algal biomass up to 7 days ahead. The recurrent feedforward architecture of ANNA proved to be most efficient in order to model and predict seasonal dynamics of chlorophyll-a and its forecasting results can be utilized for early warning and tactical control of algal blooms in freshwater lakes. A sensitivity analysis conducted by ANNA revealed that algal abundance in Burrinjuck Reservoir is not only driven by physical and chemical characteristics of the water body but to a large extend by hydrological characteristics such as water depth as well.

Knowledge discovery for prediction and explanation of blue-green algal dynamics in lakes by evolutionary algorithms

This paper applies an evolutionary algorithm to the problem of knowledge discovery on blue-green algae dynamics in a hypertrophic lake. Patterns in chemical and physical parameters of the lake and corresponding presence or absence of highly abundant blue-green algae species such as Microcystis spp, Oscillatoria spp and Phormidium spp are discovered by the machine learning algorithm. Learnt patterns are represented explicitly as classification rules, which allow their underlying hypothesis to be examined. Models are developed for the filamentous blue-green algae Oscillatoria spp and Phormidium spp, and the colonial blue-green algae Microcystis spp. Hypothesized environmental conditions which favour blooms of the three species are contrasted and examined. The models are evaluated on independent test data to demonstrate that models can be evolved which differentiate algae species on the basis of the environmental attributes provided.

An inductive approach to ecological time series modelling by evolutionary computation

Building time series models for ecological systems that can be physically interpreted is important both for understanding the dynamics of these natural systems and the development of decision support systems. This work describes the application of an evolutionary computation framework for the discovery of predictive equations and rules for phytoplankton abundance in freshwater lakes from time series data. The suggested framework evolves several different equations and rules, based on limnological and climate variables. The results demonstrate that non-linear processes in natural systems may be successfully modelled through the use of evolutionary computation techniques. Further, it shows that a grammar based genetic programming system may be used as a tool for exploring the driving processes underlying freshwater system dynamics.