A new approach to application of pattern recognition methods in analyticalchemistry - II. Prediction of missing values in water pollution grid usingmodified KNN-method

This paper (the second in the series devoted to new applications of pattern recognition methods in analytical chemistry[1]) deals with prediction of p ollutants concentrations in apparently stagnant water Determination of wate r quality indexes in lakes or ponds play a crux role in environmental prote ction, mainly for two reasons: in classification of a given lake as a bathi ng resort, and classification of the lake itself as a source of drinking wa ter supply. Quantitative estimation of chemical and bacteriological contami nation is performed according to standardised procedures, gathering water s amples in a specified collection sites, mainly in the flume, or sometimes i n the vicinity of the lake shore. Generally, it might be required to determ ine a few dozen of various chemical and bacteriological species for a given sample collection point, and depending on the size of the lake-the number of measurement points may be enormously large. In consequence, it leads to development of the so called water pollution grid. In our research, perform ed in carefully controlled conditions, the chemical and bacteriological con tamination of water in one of barrier lakes in South East Poland have been measured. The ultimate goal of research was testing the modified KNN-method (named "look-ahead-and-back-KNN" [2,3]) in forecasting concentrations of s elected pollutants in a chosen location, basing on the known concentrations of these pollutants in neighbour nodes of the grid. Taking into account th e known concentrations of pollutants in particular nodes of the grid (made intentionally sparse), multi-category predictions were carried out, and res ults obtained were compared with respective real values. It was found, that concentrations of some pollutants can be predicted with satisfactory preci sion, what in turn allows to create a dense pollution grid, consisting of m easured and - to some extent - predicted data points. It may be assumed tha t methodology developed in our research may be securely applied in estimati on of pollutants concentrations in various parts of a lake, thus providing a way to decrease the number of measuring nodes of the grid, and to cut dow n cost and time of its rendering.