Forecasting air pollutants using classification models: a case study in the Bay of Algeciras (Spain)

• Autor(a) principal: Rodríguez-García, M. I.
• Data de Publicação: 2023
• Outros Autores: Ribeiro, Conceição, González-Enrique, J., Ruiz-Aguilar, J. J., Turias, I. J.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.1/20095
• topic: Air pollution forecasting; Classification models; Minimum redundancy maximun relevance; Maritime traffic; Artificial neural networks

The main goal of this work is to obtain reliable predictions of pollutant concentrations related to maritime traffic (SO2, PM10, NO2, NOX, and NO) in the Bay of Algeciras, located in Andalusia, the south of Spain. Furthermore, the objective is to predict future air quality levels of the principal maritime traffic-related pollutants in the Bay of Algeciras as a function of the rest of the pollutants, the meteorological variables, and vessel data. In this sense, three scenarios were analysed for comparison, namely Alcornocales Park and the cities of La Linea and Algeciras. A database of hourly records of air pollution immissions, meteorological measurements in the Bay of Algeciras region and a database of maritime traffic in the port of Algeciras during the years 2017 to 2019 were used. A resampling procedure using a five-fold cross-validation procedure to assure the generalisation capabilities of the tested models was designed to compute the pollutant predictions with different classification models and also with artificial neural networks using different numbers of hidden layers and units. This procedure enabled appropriate and reliable multiple comparisons among the tested models and facilitated the selection of a set of top-performing prediction models. The models have been compared using several quality classification indexes such as sensitivity, specificity, accuracy, and precision. The distance (d(1)) to the perfect classifier (1, 1, 1, 1) was also used as a discriminant feature, which allowed for the selection of the best models. Concerning the number of variables, an analysis was conducted to identify the most relevant ones for each pollutant. This approach aimed to obtain models with fewer inputs, facilitating the design of an optimised monitoring network. These more compact models have proven to be the optimal choice in many cases. The obtained sensitivities in the best models were 0.98 for SO2, 0.97 for PM10, 0.82 for NO2 and NOX, and 0.83 for NO. These results demonstrate the potential of the models to forecast air pollution in a port city or a complex scenario and to be used by citizens and authorities to prevent exposure to pollutants and to make decisions concerning air quality.