Abstract
Air quality assessment regarding metals and metalloids using experimental measurements is expensive and time consuming due to the cost and time required for the analytical determination of the levels of these pollutants. According to the European Union (EU) Air Quality Framework Directive (Directive 2008/50/EC), other alternatives, such as objective estimation techniques, can be considered for ambient air quality assessment in zones and agglomerations where the level of pollutants is below a certain concentration value known as the lower assessment threshold. These conditions occur in urban areas in Cantabria (northern Spain). This work aims to estimate the levels of As, Cd, Ni and Pb in airborne PM10 at two urban sites in the Cantabria region (Castro Urdiales and Reinosa) using statistical models as objective estimation techniques. These models were developed based on three different approaches: partial least squares regression (PLSR), artificial neural networks (ANNs) and an alternative approach consisting of principal component analysis (PCA) coupled with ANNs (PCA-ANN). Additionally, these models were externally validated using previously unseen data. The results show that the models developed in this work based on PLSR and ANNs fulfil the EU uncertainty requirements for objective estimation techniques and provide an acceptable estimation of the mean values. As a consequence, they could be considered as an alternative to experimental measurements for air quality assessment regarding the aforementioned pollutants in the study areas while saving time and resources.
Highlights
Mathematical modelling for air quality assessment purposes has become increasingly important in recent years
For pollutants like the ones under study in this work, the performance of complex models is often equal to that of simpler methodologies. This fact highlights the interest of statistical models to estimate the ambient air concentration of atmospheric pollutants even though a wide range of deterministic models, as reviewed by El-Harbawi (2013), have been already developed and studied in the literature. Techniques such as partial least squares regression (PLSR), which presents advantages over other statistical linear regression techniques because it combines features from factor analysis statistical methods, such as principal component analysis (PCA) and linear regression techniques, as multiple linear regression (MLR), may potentially lead to more accurate estimations than those provided by MLR or principal component regression (PCR)
PLSR provides a fractional bias (FB) index lower than those obtained for artificial neural networks (ANNs) and PCA-ANNs because the mean metal concentration estimated by the PLSR models are equal —up to two significant figures— to the corresponding observed values and that, according to the corresponding equation (Table 3), yields lower FB index values
Summary
Mathematical modelling for air quality assessment purposes has become increasingly important in recent years. Quality Framework Directive establishes that in all zones and agglomerations where the level of pollutants is below the lower assessment threshold (LAT), which is expressed as a percentage of the corresponding target/limit value, modelling techniques or objective estimation techniques (or both) shall be sufficient for the assessment of the ambient air quality (European Council Directive 2008/50/EC) Both statistical and deterministic methods are currently used in regulatory air pollution forecasting by environmental authorities. For pollutants like the ones under study in this work, the performance of complex models is often equal to that of simpler methodologies This fact highlights the interest of statistical models (e.g., linear regression techniques and non-linear modelling techniques) to estimate the ambient air concentration of atmospheric pollutants even though a wide range of deterministic models, as reviewed by El-Harbawi (2013), have been already developed and studied in the literature. Non-linear models performed relatively better than the linear PLSR models
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
