Geography-based Neural Networks for the Simulation of Air Pollution

Abstract

Most environmental variables, including air pollution, are characterized by high variability in time and space. The classical approach to modelling these variables is developing physically based models with high data and computational requirements or their surrogate version, often implemented through a neural structure. We suggest that adding the geographical coordinates to the other standard inputs of the neural network architectures provides more accurate results without additional measurement costs and with minor increments of computer time. We demonstrate the advantages of the proposed approach in an application to Northern Italy.