Comparative study of two stochastic models using the physicochemical characteristics of river sediment to predict the concentration of toxic metals

Environmental models,whether deterministic or stochastic,are based upon the relationships between environmental variables and the characteristics of the dependent variables. several models have been constructed from linear correlation or multiple linear regression (mlr). this procedure assumes a linear relationship between the variables or their functions,which is rather rare in the case of environmental data. on the other hand,artificial neural network (ann) systems have recently demonstrated their capacity to process non-linear relationships. their performance has already been proven in various areas of science,even for modelling environmental processes. in this work,we have studied the predictive capacity of mlr with that of ann for the estimation of four heavy metals (cd,cr,cu and pb) concentration from eight physicochemical variables (organic matter,water content,fine fraction,ph,caco3,carbon and phosphorus in the sediment,and the suspended matter in the water column) of the beht river basin in morocco. performed with the mlr method,the determination coefficients ranged from r2=0.26 for cd to r2=0.83 for cr,with intermediate values for cu (r2=0.55) and pb (r2=0.67). trained with the back-propagation algorithm of ann,we obtained the determination coefficient at r2=0.88 for cd,r2=0.93 for cr,r2=0.96 for cu and r2=0.80 for pb. these results were much better than those obtained by the mlr method,and they demonstrate the higher predictive performance of ann. the ann can thus be a powerful alternative in comparison with the traditional techniques of modelling. its ability to give good prediction with non-linearly related variables,as is commonly encountered in environmental sciences,opens promising horizons in this research field.