An expert system to estimate the pesticide contamination of small streams using benthic macroinvertebrates as bioindicators II. The knowledge base of LIMPACT

The development and the evaluation of a biological indicator system for pesticide pollution in streams are presented. For small headwater streams with an agricultural catchment area, the expert system Limpact estimates the pesticide contamination according to the four classes: Not Detected (ND); Low (L); Moderate (M) and High (H) contamination without any specification of the chemical agents. The input parameters are the abundance data of benthic macroinvertebrate taxa within four time frames in a year (March/April; May/June; July/August; September/October) and nine basic water-quality and morphological parameters. The heuristic knowledge base was developed with the shell-kit D3 and contains 921 diagnostic rules with scores either to establish or to de-establish a diagnosis. The 418 rules had less than 3 symptoms, and only 47 rules had more than 4 symptoms in their rule condition. We differentiate between positive indicator (PI) taxa, which indicate contamination by high abundance values and positive abundance dynamics, and negative indicator (NI) taxa, a high abundance of which rules out contamination and indicates an uncontaminated site. We analysed 39 taxa and found 13 positive and 24 negative indicators. The database comprises 157 investigations per stream and year with rainfall event-controlled pesticide sampling and repeated benthic sampling as described in Part 1 (Neumann et al., this issue). For the evaluation of Limpact, we used the same cases. The correct class for the 157 investigations per stream and year is established by Limpact in 66.7–85.5% of the cases, with better results for uncontaminated sites. The overall alpha error probability (false positive) is 9.6% while the beta error probability (false negative) varied between 0 and 8% depending on the contamination class. If each stream is considered only once in the system (n = 104), the correct diagnosis is established by Limpact in 51.9–88.6% of the cases. In most of the remaining cases no diagnosis is established instead of a wrong one. © 2002 Elsevier Science Ltd. All rights reserved.