Unbiased sensitivity analysis and pruning techniques in neural networks for surface ozone modelling

This paper presents the use of artificial neural networks (ANNs) for surface ozone modelling. Due to the usual non-linear nature of problems in ecology, the use of ANNs has proven to be a common practice in this field. Nevertheless, few efforts have been made to acquire knowledge about the problems by analysing the useful, but often complex, input-output mapping performed by these models. In fact, researchers are not only interested in accurate methods but also in understandable models. In the present paper, we propose a methodology to extract the governing rules of trained ANN which, in turn, yields simplified models by using unbiased sensitivity and pruning techniques. Our proposal has been evaluated in thousands of trained ANNs under different conditions to establish a relationship between present contaminants (or several atmospheric variables) and surface ozone concentrations. The technique presented has demonstrated to be unbiased and stable with regard to the interpretability of the models and the good results obtained.