Does the Zero Carry Essential Information for Artificial Neural Network learning to simulate the contaminant transport in Urban Areas?

Abstract

The release of hazardous materials in urbanized areas is a considerable threat to human health and the environment. Therefore, it is vital to detect the contamination source quickly to limit the damage. In systems localizing the contamination source based on the measured concentrations, the dispersion models are used to compare the simulated and registered point concentrations. These models are run tens of thousands of times to find their parameters, giving the model output’s best fit to the registration. Artificial Neural Networks (ANN) can replace in localization systems the dispersion models, but first, they need to be trained on a large, diverse set of data. However, providing an ANN with a fully informative training data set leads to some computational challenges. For example, a single simulation of airborne toxin dispersion in an urban area might contain over 90% of zero concentration in the positions of the sensors. This leads to the situation when the ANN target includes a few percent positive values and many zeros. As a result, the neural network focuses on the more significant part of the set - zeros, leading to the non-adaptation of the neural network to the studied problem. Furthermore, considering the zero value of concentration in the training data set, we have to face many questions: how to include zero, scale a given interval to hide the zero in the set, and include zero values at all; or limit their number? This paper will try to answer the above questions and investigate to what extend zero carries essential information for the ANN in the contamination dispersion simulation in urban areas. For this purpose, as a testing domain, the center of London is used as in the DAPPLE experiment. Training data is generated by the Quick Urban & Industrial Complex (QUIC) Dispersion Modeling System.