Information discovery in ecological systems by artificial neural networks: Algal blooms at Gippsland Lakes

This paper aims to discuss two aspects of working with large ecological data sets; analysis and modelling of ecological data sets, and subdivision of data into smaller subsets for the purpose of analysis and modelling. Different approaches to the information discovery in ecological systems based on Artificial Neural Networks (ANNs) are considered ANNs are powerful modelling tools. Their strength is derived from their ability to model complex, non-linear relationships. However, a drawback of ANNs is that they cannot distinguish between noise and actual data in a system. Ecological systems are prone to greater noise than many other systems. The solution therefore lies in applying ANNs to ecological problems more creatively. In algal blooms and similar ecological problems the use of ANNs has been primarily limited to the predictive modelling and sensitivity analysis. This paper proposes a multi-stage analysis comprising of predictive function modelling, sensitivity analysis, principal component analysis (PCA) and non-linear principal component analysis (NLPCA). The most common method of data subdivision for training, validating and testing is a method of random or redundant subsets. This method of data subdivision is not always appropriate because ecological systems represent open sets with complex relationships. Ecological data are often incomplete and contaminated by noise. This paper proposes a systematic approach to subdivision of data into training, validating and testing datasets.