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 industry's 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.