Use of multivariate analysis in radioecological and environmental radoactivity studies - advantages and limitations

Abstract

Some of the most commonly occuring problems in radioecological and environmental radioactivitiy studies when applying traditional statistical models are multivariate and multiscale structures of data. Spatial data analysis of radioactively contaminated areas are particularly complex for many reasons: uncertainty of the source term, high spatial and temporal variability of pollution patterns, spatial and temporal nonstationarity and multivariate nature of the phenomenon with linearly and nonlinearly correlated variables. There are only few studies on employing the multivariate approach to describe the correlation between locations and radioactive contamination (Kanevski, 1996; Kanevski, 1997). In this work the feasibility of using multivariate analysis techniques, principal component analysis (PCA), linear discriminant analysis (LDA), k-nearest neighbours (kNN), soft independent modelling of class analogy (SIMCA) and artificial neural networks (ANN), to predict soils and bioindicators origin based on their radionuclide content was examined.