Air pollution is an environmental health problem which affects several countries. This is the case of Latin America where particulate air pollution is of higher concern and a rapidly growing problem. Latin America is the most urbanized region of the developing world with 75% of its population living in cities. Buenos Aires which is one of the thirty most populated cities of the world and the third in Latin America has only two air pollution monitoring stations from which only one measures particulate matter. In comparison with other cities, the coastal location and terrain flatness of Buenos Aires is considered to facilitate the dilution of air pollutant concentrations, but it has never been quantified due to the lack of motoring networks. Therefore, the objective of this thesis was to provide a cost-effective methodology for the estimation or prediction of particulate matter (PM) concentrations in the urban air of Buenos Aires city and integrate the non-linear relationship of the factors affecting PM concentrations. Firstly, the proposed methodology consisted of an analysis of the results of the measurements performed during one year in Buenos Aires within the framework of the international project BARUCA (Buenos Aires Research on Urban Climate and Air Pollution). With the evaluation of the measurement data a thoroughly understanding of the main factors affecting the behavior and levels of PM concentrations was gained. As a result, neural network models were proposed which contribute to determine in a reliable way the spatial and temporal distribution of PM10 and PM2.5 in Buenos Aires city. Neural network models in the field of air quality are mainly used because of their ability to model and learn any complex relationship between different variables. Given the importance of the selection of the input data for the training and learning of the neural networks, an analysis of the influence of the input parameters on the targets (PM concentrations) was also done. A model for the determination of the PM spatial distribution was developed which was trained with the measured PM concentrations at 67 sites and variables such as traffic categorization, land use (residential use, commercial use and green areas), population density and UTM-coordinates. Furthermore, two additional models were proposed to estimate the temporal variation of PM at two monitoring stations for the periods when no measurement data were available. The models were trained with measured PM data from the permanent monitoring site and meteorological parameters (wind speed, wind direction, temperature and humidity).
Based on the correlation coefficients, it was found that each neural network learns the relationship between the input data and the measured concentrations and the results demonstrate a comprehensible interpretation of the influence of the input data to the networks’ outputs. With the proposed models a better and more information about the PM pollution in Buenos Aires city is obtained. In this way, appropriate actions can be taken in order to reduce PM levels and consequently protect the population’s health and environment. The proposed networks provide an effective tool, based on land use maps and temporal measurement campaigns, which can be used to assess PM pollution for urban areas like in the case of Buenos Aires city where air quality monitoring networks are limited.