performance of classification methods to evaluate groundwater (case study: shoosh aquifer)

the objective of this study was to classify the shoosh aquifer to several zones with different water quality in khuzestan province, iran. in this regard, the performance of classification methods (discriminant function and cluster analysis) for the classification of groundwater based on the level of pollution with an emphasis on the problem of over-fitting in training data were considered. an over-fitted model will generally have poor predictiveperformance, as it can exaggerate minor fluctuations in the data. cluster analysis(ca) was adopted to spatially explain the similarity of sampling stations with respect to measured parameters. three methods for variable selection were used including regularized discriminant analysis, principal component analysis and wilks's lambda method. the best algorithm for variable selection was wilks'lambda which resulted in reducing the generalization error of the test sample to 0.1 for leave-one-out and 4-fold cross-validation. the second best performed algorithm was regularized discriminant function with 0.167 and 0.133 misclassification error for the two above-mentioned methods, respectively. principal component analysis did not proved to be a promising algorithm for variable selection in the classification methods.