Multivariate influence through neural networks ensemble: Study of Saharan dust intrusion in the Canary Islands

Abstract

Analyzing and predicting the concentration of airborne dust is vital to the economic activity and to the health of the population. In this study, we use a set of artificial neural networks that we structure through ensemble learning to yield a complex variable, such as the concentration of dust, based on actual data such as air temperature, relative humidity, atmospheric pressure and wind speed. The statistical performance indices obtained, show the effectiveness of the proposed approach through the application of a cross-validation committee. It is thus vital to have a reliable calculation method for determining relative importances that can be applied to this type of ensemble architecture by way of artificial neural networks. Unlike other relative importance methods, where calculations are done based directly on the weights in the artificial neural network and whose results in ensemble sets exhibit high dispersion, we propose our own procedure, which selectively chooses the variation in the inputs to readjust the architecture of the neural network. This allows us to measure those variables with the greatest effect on the target variable, thus obtaining the multivariate influence on the surface dust concentration through a computational model. This method thus provides a real alternative for calculating and estimating relative importance that can be generalized to any type of problem for multivariate systems modeled using artificial neural networks for both, a simple configuration, and an ensemble architecture