Cluster analysis and chemometric classification of sud-Paris waters

Multivariate classification models for the analytical control of a high number of water sources localized in the south-est region of Paris is proposed. These models are able to perform a global quality evaluation of the water by the simultaneous use of a very high number of analytic parameters, representing a valid device for a continuous water monitoring. For building the models, two advanced chemometric methods have been adopted: Cluster Analysis and Classification, applied by means of dedicate software. The objects set was composed of 69 drawing points, each described by 9 selected analytic parameters, carried out during the year 2004. In a first step, the examined sources have been distributed in clusters by applying an agglomerative hierarchical algorithm to the data matrix, using the Euclidean Distance as discriminative parameter. 4 clusters have so been defined, each composed by a changing number of sources. On the data matrix having clusters as dependent variable, a PLS1 regression algorithm (Partial Least Squares) was used for the multivariate model building. The obtained model has been optimized and validated. The final models have been applied to a new matrix of data, concerning the analyses done in 2005 on analogous samples, in such a way to appreciate eventual variations in the Classification of the sources. A lot of water drawing points was demonstrated to maintain the same class of the prior year, with the exception of 8 sources classified in different clusters, due to statistical significative change of some characteristic analytic parameters.Le but de ce travail consiste ? proposer des mod?les de classification multivari?s pour le contr?le analytique d?un nombre ?lev? de sources d?eau situ? au sud-est de Paris. Ces mod?les sont susceptibles d?effectuer une ?valuation globale des ?chantillons par compulsion simultan?e d?un grand nombre de param?tres analytiques et ils repr?sentent ensuite un outil valide pour une t?l?surveillance continue des eaux. Pour la construction des mod?les, deux m?thodes chimiom?triques tr?s avanc?es ont ?t? utilis?es : Analyse de Clusters et Classification, par recours ? des logiciels de derni?re g?n?ration. L?ensemble des objets est compos? de 69 points de pr?l?vement, chacun d?crit par 9 param?tres analytiques s?lectionn?s, concernant des analyses effectu?es durant l?ann?e 2004. Dans une premi?re phase on a effectu? la distribution des sources en cluster en appliquant ? la matrice de donn?es, l?algorithme de type agglom?rative hi?rarchique Average Linkage, en utilisant comme discriminant le calcul de la Distance Euclidienne entre les objets. Pour l?examen de la similarit?, 4 clusters ont ?t? d?finis, chacun compos? par un nombre variable de sources. En ce qui concerne la matrice de donn?es dans laquelle les clusters repr?sentent la variable d?pendante, on a appliqu? l?algorithme de r?gression PLS1 (Partial Least Squares) pour la construction d?un mod?le multivari?. Le mod?le obtenu a donc ?t? soumis aux ? proc?s ? d?optimisation et de validation. Le mod?le final a ?t? ainsi appliqu? ? une nouvelle matrice de donn?es, concernant les analyses effectu?es en 2005 sur des ?chantillons analogues, pour appr?cier les variations ?ventuelles dans la Classification des sources consid?r?es. Il a ?t? possible de souligner que la plus grande partie des points de pr?l?vement, dans le laps de temps d?une ann?e, s?est plac?e ? l?int?rieur des clusters d?origine, ? l?exception de 8 sources qui ont ?t? class?es dans un autre Cluster ? la suite de variations statistiquement significatives de quelques param?tres analytiques.