William J. Walley

Neural network predictors of average score per taxon and number of families at unpolluted river sites in Great Britain

Abstract Biological monitoring of river water quality in the United Kingdom and several other European and Commonwealth countries is based on the Biological Monitoring Working Party (BMWP) system. Central to the present day application of this system is the prediction of “unpolluted” average score per taxon (ASPT) and number of families present (NFAM). The paper outlines the need for such predictions and proceeds to develop predictors of ASPT and NFAM using neural networks. The basic principles of neural networks are outlined and a brief introduction to their structure and function is given via a typical example. Important preliminary considerations are fully discussed, such as model selection, training and testing procedures and the selection of relevant input variables. The results of impact analyses, designed to optimise the structures of the networks, are reported and discussed. In-depth analyses of the performance of the networks on independent test data and also relative to the industrys current model, RIVPACS III, are presented. The results of investigations into bias and error in the predicted values of ASPT and NFAM are discussed and related to some possible inadequacies in the database. It is concluded that: predictions of ASPT are significantly more reliable than those of NFAM; the neural networks performed marginally better than RIVPACS III; ASPT and NFAM can be predicted directly, without reference to site type or biological community, from a few key environmental variables; and there is scope for improved predictions if additional relevant environmental data are collected.

Using machine learning techniques in the construction of models. II. Data analysis with rule induction

Abstract Artificial intelligence and machine learning methods can be used to automate the acquisition of ecological knowledge, i.e., automate the construction of ecological models. This paper describes a particular methodology of machine learning, called rule induction, and its application to data analysis in several ecological domains. These include the biological classification of British rivers based on bioindicator data, the analysis of the influence of physical and chemical parameters on selected bioindicator organisms in Slovenian rivers and the biological classification of Slovenian rivers based on physical and chemical parameters as well as bioindicator data. In all three cases, valuable models (knowledge) in the form of rules were extracted from data acquired through environmental monitoring and/or expert interpretation of the acquired samples. This provides positive evidence for the utility of machine learning in ecological modelling.

A computer-based development of the Biological Monitoring Working Party score system incorporating abundance rating, site type and indicator value

Abstract A method of deriving BMWP (Biological Monitoring Working Party) scores and indicator values that incorporates the effects of abundance rating and site type is presented. This is an extension of earlier work by the authors, which reappraised overal BMWP family scores using biological data from the 1990 River Quality Survey of England and Wales. The method is only briefly explained but its formulation in mathematical terms is fully documented. Full details are given of the overall, site-related and site-abundance-related derived scores and indicator values of 34 selected families. New definitions of average score per taxon (ASPT) based on the derived scores and indicator values are presented, and the potential impact of these on existing BMWP site scores and ASPTs is briefly examined via two examples. Some interesting variations in the derived scores with respect to site type and abundance rating are noted, and explanations of these are given from an ecological point of view. It is concluded that the method described offers a means of significantly improving the reliability and hence utility of the BMWP score system.

Unsupervised pattern recognition for the interpretation of ecological data

The paper describes a novel pattern recognition system (MIR-max) that was developed to facilitate the construction of a river pollution diagnostic system for the British Environment Agency. MIR-max is a non-neural self-organising map based on information theory, which, unlike Kohonens Self-organised map (SOM), separates the processes of clustering and ordering. It first clusters the input samples into a pre-defined number of classes by maximising the mutual information between the samples and the classes. The classes are then ordered in a two-dimensional output space by maximising the correlation coefficient (r) between the Euclidean distances separating the classes in data space and their corresponding distances in output space. This produces a map of the classes which when labelled can be used for the classification/diagnosis of new samples. A novel feature of MIR-max is that it permits the disaggregation of the classes in the output map, thus permitting exceptional classes to separate from their neighbours. MIR-max is designed specifically for use with ordinal data, but can also be used for interval-valued data. Its application in the ecological field is demonstrated via two examples based on data from the 1995 river quality survey of England and Wales. In the first example, MIR-max is used to classify biological samples into 100 river quality classes for each of five site types. These classifiers are then tested against two corresponding neural network classifiers, and are shown to provide better performance. In the second example, MIR-max is used to classify combined biological and environmental (i.e. physical characteristics of the site) data directly into 500 quality classes. The way in which this pattern classifier has been used to produce a river pollution diagnostic system is then explained. The advantages of the system over traditional river quality assessment systems, like RIVPACS, are outlined. It is concluded that MIR-max has considerable potential for use in the visualisation and interpretation of multivariate ecological data.

Biological Monitoring: a Comparison between Bayesian, Neural and Machine Learning Methods of Water Quality Classification.

Biological methods of monitoring river water quality have enormous potential but this is not presently being realised owing to inadequacies in methods of data interpretation and classification. This paper describes the development and testing of several classification models based on Bayesian, neural and machine learning techniques, and compares their performance with two traditional models. It is demonstrated, using an expertly classified test data set, that ‘naive’ Bayesian models and multi-layered perceptrons can significantly out-perform the traditional methods. It is concluded that these two techniques presently provide the most promising means of realising the full potential of bio-monitoring, either acting separately or jointly as complementary ’experts’.

Rule-based versus probabilistic approaches to the diagnosis of faults in wastewater treatment processes

The need for computer-based diagnostic tools in wastewater management is outlined. Rule-based and probabilistic approaches to the development of diagnostic expert systems are critically reviewed, and it is demonstrated that the rule-based approach has serious limitations which make it unsuitable for diagnostic tasks under conditions of uncertainty. It is shown that Bayesian belief networks (BBNs), a probabilistic approach, has none of these limitations and is well-suited to diagnosis under uncertainty. The theory and application of BBNs are outlined and illustrated by a simple example based on a wastewater treatment plant. A brief case study is presented of the development of a full-scale BBN for the diagnosis of faults in a wastewater treatment plant. It is concluded that BBNs are far superior to rule-based systems in their ability to diagnose faults in complex systems like wastewater treatment processes, whose behaviour is inherently uncertain.

Self-Organising Maps for the Classification and Diagnosis of River Quality from Biological and Environmental Data

The paper addresses the problem of how to classify and diagnose the state of health of a river from the composition of its biological community. It is claimed that experts use two complex mental processes when interpreting such data, knowledge-based reasoning and pattern recognition. It is argued that existing classification methods are inadequate and that the application of advanced computer-based techniques is vital to the realisation of the full potential of biological monitoring. The paper then concentrates on a pattern recognition approach and demonstrates how Self Organising Maps (SOM), a type of unsupervised-learning neural network, can be used to classify and diagnose river quality. A brief introduction is given to the theory of SOMs and the interpretation of their output, as expressed in feature maps and class templates. SOMs are developed using two different methods of accounting for the confounding effects of environmental factors, and their relative performances are compared. Some improvements to the SOM architecture and functionality that are currently being implemented are briefly described, together with plans to use information theory for the assessment of performance. Finally, it is concluded that the methods of classification/diagnosis described in the paper have considerable potential not only in river quality monitoring, but also in other environmental fields.

Identification of macro-invertebrate taxa as indicators of nutrient enrichment in rivers

Abstract Eutrophication of fresh waters, especially from diffuse sources, is often a priority environmental issue for industrialised countries. Understanding the relationships between nutrient pressures and their impacts on ecology is essential for predicting the likely benefits of a programme of remedial measures to return nutrient concentrations to former levels. The aim of this study was to use mutual information to analyse the strength of association between macroinvertebrate families and nutrient levels (Total Oxidised Nitrogen and Total Reactive Phosphorus) in data covering rivers in England and Wales. Prior to the analysis the dataset was screened to minimise the confounding effects of organic pollution and split according to site type and season. Significance thresholds for the values of mutual information were calculated and the most significant indicator taxa were identified for each site type, season and nutrient pressure. It was found that in upland rivers the most significant indicator taxa were generally positive indicators, that is, their presence is indicative of high levels of at least one nutrient. In addition the number of significant indicators was greatest in upland rivers and least in lowland rivers.

An Information Theoretic Self-Organising Map with Disaggregation of output Classes

The paper describes an unsupervised pattern recognition system with feature maps, in which the tasks of clustering samples into classes and of ordering classes into two-dimensional maps are treated as separate processes. The clustering process is based on the maximisation of mutual information between the classes and the attributes of the data. The ordering process is based on the maximisation of the correlation coefficient between corresponding distances in data space and output space. The main strengths of the system are shown to be its abilities to: a) allow the disaggregation of the output classes into ‘natural’ clusters; b) produce good global ordering of the output map; c) optimise ordering globally or locally; and d) cluster ordinal and nominal data. An example application, based on river quality data, is given in which the performance of MIR-max is compared with that of the SOM and GTM unsupervised neural networks. It is concluded that MIR-max: a) outperformed SOM and GTM with respect to clustering and global ordering; b) offers some powerful new features; and c) is especially suited for use on ordinal data.