Science of the Total Environment 488–489 (2014) 197–207

Contents lists available at ScienceDirect

Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv

Can artificial neural networks be used to predict the origin of ozone episodes? T. Fontes a,b,⁎, L.M. Silva c,d, M.P. Silva a, N. Barros a, A.C. Carvalho e a

University Fernando Pessoa, Global Change, Energy, Environment and Bioengineering Center (CIAGEB), Praça 9 de Abril, 349, 4249-004 Porto, Portugal University of Aveiro, Department of Mechanical Engineering/Centre for Mechanical Technology and Automation, Campus Universitário de Santiago, 3810-193 Aveiro, Portugal University of Aveiro, Department of Mathematics, Campus Universitário de Santiago, 3810-193 Aveiro, Portugal d INEB — Instituto de Engenharia Biomédica, Rua do Campo Alegre, 823, 4150-180 Porto, Portugal e New University of Lisbon, Faculty of Sciences and Technology/Center for Environmental and Sustainability Research (CENSE), Quinta da Torre, 2829-516 Caparica, Portugal b c

H I G H L I G H T S • ANN can classify the origin of an O3 episode with a mean error around 2-7%. • The best classification is obtained when a simpler input combination is used. • ANN can help authorities to foster O3 action plans to control exceedances.

a r t i c l e

i n f o

Article history: Received 5 February 2014 Received in revised form 7 April 2014 Accepted 20 April 2014 Available online xxxx Editor: P. Kassomenos Keywords: Human health Ozone Stratosphere Troposphere Classification Artificial neural network

a b s t r a c t Tropospheric ozone is a secondary pollutant having a negative impact on health and environment. To control and minimize such impact the European Community established regulations to promote a clean air all over Europe. However, when an episode is related with natural mechanisms as Stratosphere–Troposphere Exchanges (STE), the benefits of an action plan to minimize precursor emissions are inefficient. Therefore, this work aims to develop a tool to identify the sources of ozone episodes in order to minimize misclassification and thus avoid the implementation of inappropriate air quality plans. For this purpose, an artificial neural network model – the Multilayer Perceptron – is used as a binary classifier of the source of an ozone episode. Long data series, between 2001 and 2010, considering the ozone precursors, 7Be activity and meteorological conditions were used. With this model, 2–7% of a mean error was achieved, which is considered as a good generalization. Accuracy measures for imbalanced data are also discussed. The MCC values show a good performance of the model (0.65–0.92). Precision and F1-measure indicate that the model specifies a little better the rare class. Thus, the results demonstrate that such a tool can be used to help authorities in the management of ozone, namely when its thresholds are exceeded due natural causes, as the above mentioned STE. Therefore, the resources used to implement an action plan to minimize ozone precursors could be better managed avoiding the implementation of inappropriate measures. © 2014 Elsevier B.V. All rights reserved.

1. Introduction Due to its oxidative characteristics, the adverse effects of tropospheric ozone (O3) on human health and on the environment are fully recognized (Agrawal et al., 2003; Dimitriou et al., 2011; Heal et al., 2013). The major health issues found to be related with high ozone concentrations are associated with the respiratory system, provoking or increasing lung irritations and asthma (Halonen et al., 2010; Hamade et al., 2008; ⁎ Corresponding author at: University of Aveiro, Department of Mechanical Engineering/ Centre for Mechanical Technology and Automation, Campus Universitário de Santiago, 3810-193 Aveiro, Portugal. E-mail address: [email protected] (T. Fontes).

http://dx.doi.org/10.1016/j.scitotenv.2014.04.077 0048-9697/© 2014 Elsevier B.V. All rights reserved.

Ebi and McGregor, 2008; WHO, 2006; Srebot et al., 2009). This pollutant is classified as a greenhouse gas and is responsible for inducing a reduction of the photosynthetic process thus affecting vegetation growth and reproduction, and hence crop productivity (EEA, 2011). Thereby, the prediction of ozone concentrations and, as possible the identification of its sources are fundamental to improve the effectiveness of public awareness and policies for human health protection, as well as vegetation, and increase the knowledge on the interactions between the ozone concentrations, weather and climate. To promote a cleaner air in Europe, guidelines, programmes and standards from the World Health Organization (WHO) were included in the Directive 2008/50/EC, of 21 May 2008. This Directive defines the main rules concerning the ambient air quality as well as strategies

198

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

to reduce, minimize and inform citizens about the adverse effects of air pollution. As a result, information thresholds (180 μg·m−3) and alert thresholds (240 μg·m−3) were defined for ozone. Moreover, for places where one of these limits is exceeded, Member States shall ensure the implementation of a programme, and/or an action plan, to reduce the ozone concentration and minimize its effects. Nevertheless, the recommended measures may give results when anthropogenic sources are the main ozone precursors and can be useless if the origin of the episode is natural. The main origin of tropospheric ozone is a series of complex photochemical reactions regulated by natural and anthropogenic precursor emissions, such as volatile organic compounds (VOCs) and nitrogen oxides (NOx), in the presence of solar radiation (wavelength b 424 nm) (Crutzen et al., 1999; Fishman and Crutzen, 1978). However, the ozone concentrations can also increase due to Stratosphere– Troposphere Exchanges (STE). Fishman and Crutzen (1978) reasoned that the enhancement of tropospheric ozone concentration over the Northern Hemisphere is based on photochemical reactions but its flux is not sufficient to explain the measured concentrations. Stevenson et al. (2006), and Hess and Zbinden (2013) show the importance of the STE in the tropospheric ozone burden. Studies conducted by Hess and Lamarque (2007), and Solomon et al. (2007), conclude that STE are responsible for nearly 20% of the global tropospheric ozone burden and Hegglin and Shepherd (2009) suggest that the stratospheric flux of ozone has been increasing at a nearly constant rate in the North Hemisphere of approximately 2% per decade since 1970. Therefore, this last process cannot be neglected as a dynamical mechanism that also contributes to the tropospheric ozone budget. The exchange of air masses between the stratosphere and the troposphere in the extratropics results from irreversible diabatic erosion of cut-off lows and diffusive mixing connected with tropopause folds (Elbern et al., 1997), associated with upper level frontogenesis and rapid surface cyclogenesis (Davies and Schuepbach, 1994). Other mechanisms include mesoscale convective systems, thunderstorms and breaking gravity waves (Stohl et al., 2000). The shape of the STE in the extratropics is highly variable in space and time, which induces a high variability in the surface measured ozone concentrations influenced by these dynamical processes. Stratospheric intrusions are responsible for high or very high levels of ozone in the troposphere with an unknown impact at a regional to local scale (Arsićet et al., 2011; Barros et al., 2004; Carvalho et al., 2005; Fontes et al., 2013; San José et al., 2005). Generally, the signature of this phenomenon follows fine three-dimensional atmospheric movements that only by chance are measured. Very few studies report these high ozone levels. In Madrid (Spain) between 2 h00 and 6 h00 in the 29th of April, 2000 several monitoring stations reported ozone concentration levels up to 1190 μg·m−3 (San José et al., 2005). In Bor (Servia) ozone concentration reached values of 3000 μg·m−3 in two episodes, which lasted 3–5 days in November 2010 (Arsić et al., 2011). Other studies report similar results (e.g. Fadnavis et al., 2010; Ganguly, 2012). In spite of scarcity of extremely high surface ozone measurements, the actual contribution to ozone episodes near the legal values presented above is up to now, difficult to justify and forecast. Unknowing the real reason for an ozone episode may be costly when action plans are enforced by law (EEA, 1999). Thus, to improve the effectiveness of these plans more research on this field is needed in order to identify the origin of these episodes. In the last decades, the research has been focused in quantifying the ozone levels from different sources and understanding the main processes related with their formation. In fact, few studies have been developed in order to present tools which can help authorities to optimize the costs of an action plan and thus minimize the negative effects of ozone. Some exceptions are the studies presented by Grewe (2006) and Emmons et al. (2012). Grewe (2006) applied a chemistry–climate model E39/C to determine the origin of ozone (using a classification model), and Emmons et al. (2012) present a technique of tagging

ozone from various source regions by using artificial tracers of NO and its oxidation products. These are interesting approaches to identify the origin of ozone, but complex and time consuming to implement on the air quality management networks and services. Instead of using deterministic models, as the ones used in the abovementioned works, other groups of researchers have been focused on the prediction of ozone concentrations using statistical models. These linear and non-linear methods can be used for constructing non-deterministic models that can be used not only to forecast the ozone concentration but also to predict the concentrations of other pollutants as PM10 and NOx (using regression model). Although linear models are acceptable, non-linear models capture the non-linearity of ozone response and thus have been one of the preferred statistical tools for ozone prediction in the last years (Chattopadhyay and Bandyopadhyay, 2007; Hájek and Olej, 2012; Sellitto et al., 2011; Taormina et al., 2011; Tsai et al., 2009; Wang et al., 2003). In these studies, chemical and/or meteorological input variables were used. Most of the studies use chemical variables related with ozone and its precursors' concentrations (NO, NO2) (e.g. Chattopadhyay and Chattopadhyay, 2012; Moustris et al., 2012). In addition, in some of the studies other pollutants, such as carbon oxide — CO, particles — PM (PM10 and/or PM2.5), and sulfur dioxide — SO2, are also included (e.g. Hájek and Olej, 2012; Özbay et al. 2011; Zhang et al., 2012). Concerning meteorological variables, the most commonly used variables are temperature, wind speed, wind direction, solar radiation and relative humidity. Additionally, some authors include other meteorological variables such as the Solar Zenith Angle (SZA), the reflectance data and the Total Ozone Column (TOC) (Chattopadhyay et al., 2012; Sellitto et al., 2011), rain (Chattopadhyay and Chattopadhyay, 2012; Özbay et al., 2011) and the cloud cover (Chattopadhyay and Chattopadhyay, 2012; Tsai et al., 2009). In general, the literature review shows a lack of knowledge about the prediction of the origin of ozone episodes, with the underlying assumption that the near surface ozone concentrations are mainly resulting from photochemistry reaction products. In addition, the literature review on ozone regression models shows that the large majority of studies were developed to predict the surface ozone concentrations using chemical and meteorological data. Although these regression models (e.g. Hájek and Olej, 2012; Sellitto et al., 2011) are simpler than those classification models (e.g. Emmons et al., 2012; Grewe, 2006) they give information only on the ozone concentration level but their origin is not addressed. Therefore, the main purpose of the work presented here is to fill this gap and include the ozone origin into an artificial neural network model for ozone forecast purpose. To achieve this goal the MLP architecture was used to build a binary classifier of the origin of ozone episodes, anthropogenic vs natural. The objective of the development here presented is to understand if an intelligent computational tool can be used to help authorities to improve the efficiency of ozone concentration management mainly when ozone concentrations exceed threshold values due to natural causes, like STE. Although STE are classified as rare events, in this case the implementation of control programmes is clearly ineffective. For this reason, knowing the origin of ozone may be of paramount importance in order to improve the actuation of the air quality management services improving the use of the available resources. 2. Materials and methods To predict the origin of ozone episodes between anthropogenic (e.g. photochemical due to the presence of anthropogenic precursors) and natural (in the present study mainly due to STE), four main steps were defined: (i) firstly, air quality and meteorological data were collected from different sources (see Subsection 2.1); (ii) secondly, the classification of ozone episodes was then performed by experts in the field and also based on knowledge available in the literature (see Subsection 2.2); (iii) in the third step, different input scenarios were evaluated, in

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

199

2013). Because the 7Be isotope is a chemical tracer of atmospheric air masses of high troposphere and low stratosphere, it is the one of the parameters to consider in the study, as also indicated by the WMO (2001). Therefore, the 7Be measurements collected in the region of Lisbon and Tagus Valley were included in the database. This location is the only area over Portugal's mainland where 7Be activity (mBq·m−3) measurements are available on a regular basis since 2001. The 7Be activity was measured in aerosol particles collected from the Portuguese Nuclear Technological Institute (ITN) weekly between 2001 and 2010 (as described in Carvalho et al., 2013). Additionally, air quality pollutant concentrations and meteorological data were acquired, inside the study domain, by the air quality and meteorological national network stations, respectively. Concentrations

order to identify which variables are more important for this discrimination problem as well as to foresee locals where a small number of variables are available (see Subsection 2.3); and (iv) finally, the whole modelling process is described, including a discussion on alternative accuracy measures for imbalanced data (see Subsection 2.4). Fig. 1 presents an overview of the overall methodology. 2.1. Data collection High beryllium activities, in its isotope 7 (7Be), may be an indicative for stratospheric intrusions. Several studies consider 7Be as a parameter to identify stratospheric air masses (eventually with high ozone concentrations) (Allen et al., 2003; Cristofanelli et al., 2006; Carvalho et al.,

Data collection: •7Be activity •Air pollutants concentrations •Meteorological data

Classification of episodes by experts

Statistical analysis

Definition of scenarios

For each scenario: Dataset randomization

5-fold

Cross-validation

Training set (

)

Test set (

)

Model search

No 30 repetitions? Yes Model evaluation

Fig. 1. Methodology overview for the ANN model construction.

200

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

(μg·m−3) of O3, NO2 and NO, recorded in 25 air quality stations located at the Lisbon and Tagus Valley agglomeration were used. In this analysis the type of environment of each air quality station was analysed (1: urban; 2: suburban; and 3: rural). Pressure (hPa), precipitation (mm·h−1), temperature (°C), relative humidity (%) and specific humidity (g·kg−1) were measured at the Gago Coutinho meteorological station (serving the Lisbon International Airport). Fig. 2 shows the location of the different air quality and meteorological stations used. Table 1 presents the various variables considered in this study. Data from the air quality stations were validated according to the requirements of Directive 2008/50/EC. To ensure accuracy of the measurements and compliance with the data quality objectives, this Directive requires an integrated system of quality control and maintenance of the measurement equipment fulfilling of ISO/IEC 17025:2005 norm (IPQ, 2005).

2.2. Classification of episodes From all the available data only cases with hourly concentrations of ozone above 180 μg·m−3 were kept. This value is the threshold over which a brief exposure is a risk to human health, particularly for the sensitive sections of the population for which immediate and appropriate information is necessary (Directive 2008/50/EC). In these situations authorities need to implement an action plan in order to control emission precursors and minimize the effects of ozone concentrations. At this point, the data consists of a matrix with 153 rows (number of cases) and 22 columns (number of variables measured). The values were

summarized on a weekly mean basis, in order to be in accordance with the 7Be sample frequency. Two origins for ozone episodes are defined: (i) only local/mesoscale photochemical origin (class 1); or (ii) other origins, stratospheric (STE) and/or advection (class 2). The episode labelling was then performed by two of the authors of the paper, experts on this research area. This work was specifically developed for this paper. Previous knowledge based on a literature review was also considered. The analysis of the literature shows important details that can influence the ozone concentrations, nevertheless, the definition of specific thresholds to classify the origin of these concentrations is difficult to define. Meteorological variables are important elements that can directly influence ozone production, accumulation and depletion (Zhang et al., 2012). The study performed by June and Hui-Jun (2010) suggests that regional climate (e.g. temperature and wind), in a linear way, is responsible for 80% of the variance of total ozone. Some studies proved a high correlation between air temperature and ozone concentration, resulting in an increase of surface ozone concentrations with the rise of temperature during photochemical ozone episodes (Dawson et al., 2007; Sousa et al., 2007; David and Nair, 2011; Fernández-Fernández et al., 2011). Moustris et al. (2012) explain that during an advection ozone episode, wind and temperature influence surface ozone concentrations due to its effect in the distribution of air pollutants by dispersion and horizontal transport. Moreover, some studies show a high correlation between ozone and relative humidity which correspond to a variable with direct influence on the ozone processes (Mintz et al., 2005; Sousa et al., 2007). Camalier et al. (2007) confirm that ozone generally increases with increasing temperature and decreases with increasing relative humidity.

Fig. 2. Spatial distribution of the 7Be, the meteorological and air quality stations in Lisbon and Tagus Valley used in the construction of the ANN models.

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

201

Table 1 Measured variables used in the non-linear model construction. Variable General variables Beryllium activity concentration Air quality variables

Meteorological variables

Month of occurrence Type of environment Ozone concentration Nitrogen monoxide concentration Nitrogen dioxide concentration Number of hours of ozone episode Temperature Pressure Partial pressure Relative humidity Specific humidity

This last relation could be linked to the existence of greater cloud cover and atmospheric instability leading to reduction of photochemical process. Specific humidity and potential vorticity are two of the tracers well known to determinate the presence of ozone in the upper troposphere (Felker et al., 2011). When photochemical processes are more intense, high correlations are found between ozone concentrations and several chemical parameters as NO, NO2 and NOx (Im et al., 2013, Wang et al., 2003). Another important factor that may explain surface ozone concentration is the7Be isotope activity measured at surface. This isotope is a chemical tracer of atmospheric air masses with origin in the high troposphere–low stratosphere (WMO, 2001). Reiter et al. (1983) explain that the presence of high 7Be activity is related with stratospheric intrusions. Several recent studies consider this isotope a parameter to identify stratospheric air masses (Allen et al., 2003; Cristofanelli et al., 2006; Carvalho et al., 2013). In the Zugspitze station (2962 m a.s.l.) Slàdkovic and Munzert (1990) and Stohl et al. (2000) identified the value of 8 mBq·m− 3 as the threshold value to characterize stratospheric intrusions. This value was also proposed by Reiter et al. (1983) as a threshold above which an air mass probably has stratospheric characteristics. Although this value cannot be applied to other study domains Cristofanelli et al. (2006) used it to selected stratospheric intrusion episodes in the Cimone mountain (2165 m a.s.l.).

Acronym

Description/Unit

Time period

M 7 Be TE O3 NO NO2 NH T P PP RH SH

1: Jan; 2: Feb; 3: Mar, … mBq·m−3 1: urban; 2: suburban; 3: rural μg·m−3 μg·m−3 μg·m−3 – °C hPa mm·h−1 % g·kg−1

– Weekly Hourly Hourly Hourly Hourly Hourly Hourly Hourly Hourly Hourly Hourly

meteorological group of variables moderate correlations are found with high statistical significance (p b 0.001). Table 3 presents the significance value for the t-test and the Mann–Whitney test for each variable. Significant values (p b 0.05 or p b 0.001) mean that for one of the classes the values of the specific variable are significantly higher (regardless of the specific variable). This hints some discriminative power of that variable. On the other hand, for variables with p ≥ 0.05, significant differences between the classes are not expected. Note, however, that despite the fact that the variable shows lack of discriminative power per se, it might give a good contribution for the discriminative power of a set of variables. It may be observed from Table 2 that some variables such as the type of environment, the number of hours with concentrations higher than 100 μg·m− 3, the NO and pressure have a potential lack of discriminative power and are candidates to be discarded in some scenario. All the other variables have, in some way, significant differences between the classes and can be considered as potentially useful for discrimination. Based on the results provided in Tables 2 and 3, nine scenarios were defined (S1–S9) based on different combinations of the input variables. S1 uses all the available variables. The remaining combinations were created by analysing the statistical significance values of the Spearman

2.3. Input scenarios It is usually difficult in real world situations to collect a great diversity of variables, such as the case of this study. Therefore, it is important to understand the impact on the model's accuracy when different/reduced subsets of variables are used. Thereby, to select the main variables with predictive capacity of ozone origin, nine scenarios were defined considering different sets of inputs. For that purpose, a statistical evaluation of the input variables was performed in two steps. Firstly, the correlation between ozone, its precursors and some meteorological parameters (NO, NO2, T, P, PP, RH and SH) was analysed. Precursors and meteorological variables without significant correlation (p N 0.05) with ozone are candidates to be discarded. Spearman's correlation is used as the Kolmogorov–Smirnov test shows no evidence that the variables follow a normal distribution. Secondly, the same data are used to analyse the discriminative power of each variable by applying a t-test or a Mann–Whitney test (groups are defined by the classes). The Mann–Whitney test is preferred whenever the Kolmogorov–Smirnov test shows no evidence that the groups follow a normal distribution. This analysis was conducted using SPSS Version 21 (IBM, 2012). Table 2 presents the Spearman correlation coefficient (rs) between ozone concentrations and its precursors as well as with meteorological parameters. Statistically significant (p b 0.05 or p b 0.001) values of correlation are appropriately marked (* 5% level; ** 1% level). With some few exceptions low correlations are found. In the air quality and in the

Table 2 Spearman's correlation coefficient (rs) between different O3 measures (minimum, average and maximum) and several air quality and meteorological measures. O3

Variable General variables

Minimum

Month of occurrence Type of environment No. of hours with O3 ≥ 100 μg·m-3 No. of hours with O3 ≥ 180 μg·m-3

Air quality measurements

NO2

NO

Pressure

Temperature Meteorological measurements

Relative humidity Specific humidity

Minimum Average Maximum Minimum Average Maximum Minimum Average Maximum Minimum Average Maximum Minimum Average Maximum Minimum Average Maximum

Average

Maximum

-0.013

0.184*

0.128

0.199* 0.317** -0.014

0.112 0.624** 0.117

-0.077 0.147 0.664**

-0.020 -0.415** -0.500** 0.198* -0.545** -0.648** -0.033 -0.261** -0.212** 0.056 -0.001 -0.221** 0.113 -0.163* -0.286** -0.050 -0.109 -0.328**

0.243** 0.019 -0.294** 0.105 -0.313** -0.438** -0.122 -0.545** -0.612** 0.324** 0.383** -0.427** 0.384** -0.548** -0.433** -0.109 -0.081 -0.204*

-0.045 0.108 0.228** -0.082 0.091 0.109 -0.260** -0.122 -0.103 -0.089 -0.008 0.185* 0.101 0.128 0.106 0.192* 0.154 0.137

Notes: Bold values point out significant values of correlation at: *5% level; and **1% level. Strong (|rs| ≥ 75%), Moderate (75% N |rs| ≥ 50%), Low (|rs| b 50%). Correlation:

202

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

Table 3 Significance value (p-value) for the t-test or Mann–Whitney test for the groups defined by the two classes of ozone episodes. Variable General variables Air quality measurements

Meteorological measurements

Spearman correlation and S9 includes all variables with significant values of the t-test or Mann–Whitney test. Table 4 shows a summary of the variables considered by these scenarios.

p-Value Month of occurrence Type of environment No. of hours with O3 ≥ 100 μg·m−3 No. of hours with O3 ≥ 180 μg·m−3 Minimum O3 Average Maximum Minimum NO2 Average Maximum NO Minimum Average Maximum Pressure Minimum Average Maximum Temperature Minimum Average Maximum Relative humidity Minimum Average Maximum Specific humidity Minimum Average Maximum

2.4. Modelling

0.001 0.670 0.472 0.000 0.057 0.042 0.680 0.000 0.000 0.004 0.021 0.096 0.376 0.182 0.349 0.100 0.022 0.090 0.041 0.715 0.001 0.565 0.000 0.000 0.000

ANNs are a broad set of statistical models developed with an inspiration from the biological neural networks of the brain. They have been used with success in several areas of knowledge, from engineering applications to social sciences. In the present section we describe the specific model used, how it is optimized and how it is applied to obtain binary predictions for the origin of ozone episodes. For a comprehensive approach on ANNs please refer to Bishop (1995) or Haykin (2009). All the implementations and computations were performed using MATLAB (MathWorks, 2012). 2.4.1. The model We consider the Multilayer Perceptron (MLP) a feedforward artificial neural network model that maps a set of inputs to a set of outputs using nonlinear compositions of functions. An MLP can be seen as a directed graph with a stacked arrangement of layers composed of processing units (neurons). Layers between the inputs and the outputs are known as hidden layers. We restrict here to the case of a single hidden layer as this type of MLP is known to be a universal approximator provided that the number of neurons in that layer (hidden neurons) is sufficient (Cybenko, 1989). Also, one output neuron is sufficient as we are performing binary classification. In a formal way, for d inputs and nhid hidden neurons, the model can be expressed as:

Bold values: p-value b 0.05.

correlation (S2–S5), and the statistical significance values of the t-test or Mann–Whitney test (S6–S9). For each of these groups, the procedure follows with the analysis of the significance values of the minimum (S2 and S6), average (S3 and S7), and maximum values (S4 and S8) of each variable. In these groups the month of occurrence, the type of environment and/or the number of hours of ozone concentration higher than 100 or 180 μg·m−3 were also included , whenever statistical significance is achieved. S5 includes all variables with significant values of

0 y ¼ φ@

1 nhid X ð2Þ ð2Þ A wj hj þ b

0 j¼1 1 ! nhid d X X ð1Þ ð1Þ ð2Þ A @ wð2Þ φ þb wkj xk þ b j ; j ¼φ

j¼1

ð1Þ

k¼1

hj

Table 4 Set of input variables for each scenario. Variables

Scenario S1

General variables Air quality measurements

Meteorological measurements

Total

Month of occurrence Type of environment No. of hours with O3 ≥ 100 μg·m−3 No. of hours with O3 ≥ 180 μg·m−3 Minimum O3 Average Maximum Minimum NO2 Average Maximum NO Minimum Average Maximum Pressure Minimum Average Maximum Temperature Minimum Average Maximum Relative humidity Minimum Average Maximum Specific humidity Minimum Average Maximum

x x x x x x x x x x x x x x x x x x x x x x x x x 25

S2

S3

S4

S5

S6

S7

S8

S9

x

x

x

x

x x x x

x

x

x

x x x x

x x x

x

x x x x x x x

x

x

x x x

x x

x x x

x x x x x x x x

x

x

x x x x x x x x x x x x x x x x

x

x

x

x x x x

x

x

x x

x x

x

x x

x 13

x 15

5

x 21

6

6

x

x 5

x x x 15

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

where: k-th input variable; Parameter/weight connecting input k to hidden neuron j; Bias term connected to hidden neuron j; Output of hidden neuron j; Parameter/weight connecting hidden neuron j to the output neuron; Bias term connected to the output neuron; Output/response of the MLP. Activation function: φðaÞ ¼ 1þe1 −a :

xk w(1) kj b(1) j hj w(2) j b(2) y φ

The MLP architecture can thus be represented in the form d : nhid : 1, which means it has d inputs (depending on the complexity of each scenario used, see Subsection 2.3), nhid hidden neurons and 1 output neuron. 2.4.2. Model search To search for the optimal set of weights (also known as training phase), the batch backpropagation iterative algorithm was used along with the gradient descent optimization technique. Two different objective (cost) functions were used. Let n be the number of instances, ti ∈ {0, 1} is the target value (class code) for instance i and yi ∈ [0, 1] is the MLP output for instance i. The Cross– Entropy (CE) cost function (Bishop, 1995) is expressed as: RCE ¼ −

n X

yi logðyi Þ þ ð1−yi Þ logð1−yi Þ;

ð2Þ

i¼1

while the Exponential (EXP) cost function is given by REXP ¼

n X i¼1

τ exp

! ðt i −yi Þ2 : τ

ð3Þ

The latter is a parameterized family of cost functions with the ability to emulate the behaviour of other cost functions (Silva et al., 2008). EXP has the advantage of adapting the cost function to the particular problem at hand by controlling τ. The learning rate η controls the rate of weight updating and is an essential parameter in the training phase. We used an adaptive learning rate as described in Marques de Sá et al. (2012). To set the number of training epochs (iterations of the optimization algorithm), the number of hidden neurons nhid and the τ parameter of EXP a series of preliminary experiments were carried out as now described. For each value of nhid in the case of CE and each pair (nhid, τ) in the case of EXP, 10 repetitions of a stratified 5-fold cross-validation were performed. For each repetition, the whole dataset was randomized and each training fold was normalized, to have inputs with zero mean and unit standard deviation, and the corresponding test fold was normalized using the parameters of the training fold; the MLP was then initialized with small random weights and trained during 3000 epochs. Along this process, we kept track of the test error (misclassifications) which was then used to choose an appropriate number of epochs, nhid and τ. For this purpose, nhid was varied from 2 to 20 and τ was varied in the interval [−3, 3]. The parameters were then chosen using a combination of the following three rules: lowest mean test error and standard deviation; lowest complexity (lowest number of hidden neurons); and homogeneity of the model across the scenarios (i.e. as much as possible the same architecture across scenarios). This strategy prevents picking overfitted models which happen when a too complex model fits the training data and not the distribution of the data (despite the training error decreases, the test error increases). The complete set of parameters for each pair scenario/cost function is shown in Table 6. As you can see, depending on the inputs and the

203

cost function of the model, different optimal configurations were recorded. However, no correlation was found between the architecture and the complexity of the model (number of inputs and hidden neurons), which show the non-linearity of the problem. Moreover, for each scenario, no significant differences were found between the mean errors for the different architectures (~error b 5%), which suggest that simple models can provide satisfactory results. 3. Results and discussion Although we want a simple model with the lowest mean error, accuracy measures must be used to assess the degree of closeness of the model predictions with the “real” values. This evaluation can be very important, when we have imbalanced data, as in this case. Thus, in order to analyse these questions, Subsection 3.1 presents the obtained results and discusses the imbalanced problem and accuracy estimation. In addition, policy implications related with the application of the model are discussed in Subsection 3.2. 3.1. The imbalance problem and accuracy estimation It is known that imbalanced data can compromise learning in two ways. First, by compromising the performance of learning algorithms that are not usually prepared for such imbalanced class distributions. Second, traditional accuracy measures are distribution dependent and do not provide a clear picture of the classifier's functionality. Several strategies have been suggested, and tested, to solve (or at least, alleviate) the effects of class imbalance, either by adapting, in some sense, the traditional classifiers (usually by using sampling strategies) or by considering different measures of performance. We followed the latter approach. For that purpose we considered a set of seven performance measures based on the confusion matrix of the test set: Accuracy, Error rate, Precision, Recall, F1-measure, balanced error rate (BER) and Matthews Correlation Coefficient (MCC) (see Appendix A for details). The estimation of such quantities is carried out with a similar experimental procedure described before (Subsection 2.4.2) with the exception that now 30 repetitions of a 5-fold cross-validation are performed (the data is randomly shuffled in each repetition). Moreover, instead of the single mean test error, we now keep track of all the confusion matrices generated. Figs. 3 and 4 present the results of the above accuracy measures using the CE and EXP cost functions respectively, obtained for each of the nine scenarios previously defined. It can be observed that the results are, in general, promising with expected differences across scenarios. A mean error between 2% and 7% was observed. This means that if we have 100 ozone episodes with the application of this model at least 93 can be classified correctly. However, to understand if our rare class (and our main goal) is correctly classified, an analysis of the other accuracy measures must be done. A ranking obtained with MCC reveals both for CE and EXP that the best scenarios are from S1 to S5, pointing S3 and S4 as the best ones. For CE, S3 is the best scenario as the Precision and F1-measure values are slightly higher, indicating that the model specifies a little better in the rare class (natural origin of ozone), while EXP elects S4 as the best one. The latter has the advantage of being a simpler model (requiring less inputs). Hence, depending on the available information (variables collected) at the specific site one may choose to use S3 or S4. Counter intuitively, we note that S1 is not the best scenario, indicating, eventually, that some of the variables are uninformative or redundant (but all of them bring some noise) and can be discarded from the model. This fact supports the importance of defining and studying different (input) scenarios. It is also interesting to observe the impact of the class imbalance on the results, namely when comparing Error with BER. The latter essentially doubles the value of the former and, therefore, the accuracy in the most prevalent class is masking the behaviour of the model in the other class. This is more evident for scenarios S6 through S9 where

204

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

Fig. 3. Classification results using the CE cost function: a) Precision, Recall and F1-measures; and b) Error, BER and MCC. Except MCC, all the measures are presented in %.

BER is above 10%. This can be related with the type of variables used to perform the prediction. In fact, these scenarios (S6–S9) did not include any variable related with pressure. The overall error should not, in this case, be the preferred measure to analyse. Note also that the standard deviations are high in some cases. This is expected as the dataset is small; a single misclassified instance has a great impact on the estimates. It is not clear whether one of the cost functions performs better than the other. Despite the trial to choose τ as similar as possible across scenarios, it was found that EXP was quite sensitive to the choice of this parameter essentially due to the small size of the dataset. Nevertheless, EXP consistently prefers (except for S9) lower complexity models. It was also verified which variables are most important for this discrimination problem. For example, when the number of hours with ozone concentrations higher than 180 μg·m−3, the minimum pressure and the minimum specific humidity variables are introduced on the S3 scenario, corresponding to S5, a lower performance is obtained. The same happens between S4 and S5. Moreover, the scenarios which include the maximum NO2 and maximum temperature can produce better results.

3.2. Policy implications In this work, several scenarios were analysed in order to understand which variables allow a better prediction (binary classification) of the origin of ozone episodes. The results across scenarios have shown to be very similar. This suppleness is very important because it demonstrates that different combinations of inputs provide similar results. In addition, the accuracy of the model gives very promising results. The rare events, as the ones introduced into the surface through stratospheric–tropospheric exchanges (STE), are very well classified showing mean errors of 2% to 7%. Accuracy measures confirm the good performance of the model showing MCC values ranging between 0.65 and 0.92. Precision and F1-measure indicate that the model specifies a little better the rare class. Hence, although the model was developed to the Lisbon area, the results obtained suggest that for other areas with limited information it can be possible to adapt the inputs, maintaining the same error levels. This capacity of adaptation of the model is very important since high costs of ozone pollution have been recorded in the last decade. EU estimated, for the year 2000, that ozone pollution had an impact on health costs and crop losses of about 11.9 and 2.8 billion euros, respectively

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

205

Fig. 4. Classification results using the EXP cost function: a) Precision, Recall and F-measures; and b) Error, BER and MCC. Except MCC, all the measures are presented in %.

(EEA, 2008). Although the proposed tool did not cut the problem, some of those features demonstrate that such a tool has a high potential in helping authorities to develop ozone action plans with more accuracy. Moreover, coupling this classification model with meteorological and air quality prediction models, is possible to predict the origin of ozone exceedances and thus improve the use of the resources available to minimize the impacts on environment and human health. This is specially important during the occurrence of STE where the control of ozone emission precursors as VOC and NOx has a reduced impact in the tropospheric ozone levels. Thus instead of using such resources to control

ozone emission precursors, those resources can be used to reinforce the plans of human health protection. 4. Conclusions A multilayer perceptron with one hidden layer was applied to automate the classification of the origin of ozone episodes given several air quality and meteorological variables. It was found that with a small complexity model (with 4 to 10 hidden neurons) a mean error around 2% to 7% may be achieved (depending on the scenario at hand) which

Table 5 Number of epochs, hidden neurons and τ selected for each of the nine scenarios. Cost function

CE EXP

Parameters

Epochs nhid Epochs nhid τ

Scenarios S1

S2

S3

S4

S5

S6

S7

S8

S9

1000 10 1500 8 3

1000 10 2000 6 1

1000 10 2500 6 1

2000 10 2000 4 2

1000 10 2500 6 2

2000 10 3000 6 2

2000 8 100 6 2

2000 8 100 6 1

2000 4 100 6 1

206

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207

Accuracy or error rate can be readily computed from the confusion matrix by

Table 6 Confusion matrix for a binary classification problem. Predicted values

Actual values

Positive (P) Negative (N)

Positive (P)

Negative (N)

True positives (TP) False positives (FP)

False negatives (FN) True negatives (TN)

can be considered as a good generalization. The analysis of the accuracy measures shows that the learning process is affected by the class imbalance present in the data. Thus, measures such as the mean error (or accuracy) are not the most appropriate. However, the MCC values show a good performance of the model (0.65–0.92). In addition, Precision and F1-measure indicate that the model specifies a little better the rare class. Although, the best results are obtained when a simpler input combination is included, as is the case of scenarios S3 and S4 (please see details in Table 5) no correlation was found between the model complexity and accuracy measures. Due to the small dataset size available for this study, results should benefit from the collection of more data or the application of sampling strategies as suggested by some authors. The present work demonstrates that such a tool has a potential application value on helping authorities to foster ozone action plans, in particular to control ozone exceedances due to natural causes (as the ones introduced into the surface through stratospheric–tropospheric exchanges — STE). This allows governments to better justify to the European Commission and also better assess the potential benefits of the introduction of an action plan to minimize emissions of ozone precursors, therefore minimizing the implementation of inappropriate air quality plans. Conflict of interest There are no known conflicts of interest associated with this publication and there has been no significant financial support for this work that could have influenced its outcome. Acknowledgements This work was partially funded by FEDER Funds through the Operational Programme “Factores de Competitividade — COMPETE” and by National Funds through FCT — Portuguese Science and Technology Foundation within the projects PTDC/CTE — ATM/105507/2008, PTDC/ EIA — EIA/119004/2010 and the Strategic Project PEst-C/EME/UI0481/ 2014. The authors also acknowledges to the Campus Tecnológico e Nuclear, of the Instituto Superior Técnico-Universidade de Lisboa and to the Portuguese Environmental Agency (APA) that made the beryllium activity and the air quality available, respectively. Appendix A We discuss accuracy measures for the case of a two-class imbalanced problem, that is, where one of the classes has fewer cases (say, less than 10%) than the other one. The basis for this approach is the confusion matrix, a two-way table that summarizes the performance of a classifier in each class. Considering one of the classes as the positive (P) class (usually the rare one) and the other as the negative (N) class, four quantities may be defined: the true positives (TP), the true negatives (TN), the false positives (FP) and the false negatives (FN). For example, TP are the number of instances of the positive class that were correctly classified. The confusion matrix is represented in Table 6.

Accuracy ¼

TP þ TN ; TP þ TN þ FP þ FN

Error rate ¼ 1−Accuracy:

ð4Þ

ð5Þ

However, these measures are not the most appropriate to analyse in the presence of imbalanced data and so, several measures have been suggested for this purpose (a good overview can be found in He and Garcia (2009)). In this work we have considered: Precision ¼

Recall ¼

TP ; TP þ FP

TP ; TP þ FN

F 1 −measure ¼

BER ¼

2  Precision  Recall ; Precision þ Recall

  1 FP FN þ ; 2 TN þ FP TP þ FN

TP  TN−FP  FN MCC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : ðTP þ FPÞðTP þ FNÞðTN þ FP ÞðTN þ FNÞ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

ð10Þ

Precision gives an insight on how the classifier specifies on the positive class, giving a measure of how many positive predicted instances are in fact positive. This does not mean however that all the positive instances have been correctly classified. On the other hand, Recall gives an insight of the classifier's performance on the positive class, measuring how well the whole positive class is recognized. In general, Precision and Recall share an inverse relationship, that is, increasing Precision implies a reduction in Recall (but not necessarily, as in a perfect classification scenario, with no errors, one would have Precision = Recall = 1). It is obvious that these measures should not be analysed separately and the F1-measure is an attempt to combine both (it is the harmonic mean of Precision and Recall). It varies between 0 and 1 and higher values imply higher (trade-off) values for Precision and Recall. BER stands for Balanced Error Rate and essentially computes an equally weighted average of the errors in each class (in contrast to Accuracy which weights class errors by their proportions in the data). This gives a fairer estimate of the performance and functionality of the classifier. Finally, MCC stands for Matthews Correlation Coefficient and is considered as one of the best measures to summarize a confusion matrix (in a single value) (He and Garcia (2009)). Note that this measure uses all the TP, TN, FP and FN values. The classifier is better as MCC tends to one (its maximum value) and no better than a toss of a coin for MCC = 0 (MCC also takes value in [− 1,0] but in this case one would just change the class labels to get MCC back to [0,1]). Except for MCC, all the measures can be multiplied by 100 to get the results in %. The joint analysis of this pack of measures can give a more clear view of the classifier's performance and value. References Agrawal M, Singh B, Rajput M, Marshall F, Bell J. Effect of air pollution on periurban agriculture: a case study. Envron Pollut 2003;126(3):323–39. Allen DJ, Dibb JE, Ridley B, Pickering KE, Talbot RW. An estimate of the stratospheric contribution to springtime tropospheric ozone maxima using TOPSE measurements and beryllium-7 simulations. J Geophys Res 2003;108(D4):8355.

T. Fontes et al. / Science of the Total Environment 488–489 (2014) 197–207 Arsićet M, Nikolić D, Djordjević P, Mihajlović I, Živković Z. Episodes of extremely high concentrations of tropospheric ozone in the urban environment in Bor — Serbia. Atmos Environ 2011;45(32):5716–24. Barros N, Borrego C, Fontes T, Carvalho AC, Moreira N, Leitão P, et al. Stratospheric ozone into the troposphere over Portugal In Air Pollution 2004, Rhodes, Greece, 30/06 a 02/ 07; 2004. p. 315–25. Bishop CM. Neural networks for pattern recognition. Oxford, U.K.: Oxford University Press; 1995 Camalier L, Cox W, Dolwick P. The effects of meteorology on ozone in urban areas and their use in assessing ozone trends. Atmos Environ 2007;41(33):7127–37. Carvalho A, Moreira N, Leitão P, Fontes T, Barros N, Borrego C. A winter ozone episode over Portugal. Geophys Res. Abs. European Geosciences Union 24–29/04, Vol. 7, (SRef-ID: 1607-7962/gra//EGU05-A-06926). Vienna, Austria; 2005. Carvalho AC, Reis M, Silva L, Madruga MJ. A decade of 7Be and 210Pb activity in surface aerosols measured over the Western Iberian Peninsula. Atmos Environ 2013;67: 193–202. Chattopadhyay S, Bandyopadhyay G. Artificial neural network with backpropagation learning to predict mean monthly total ozone in Arosa, Switzerland. Int J Remote Sens 2007;28(20):4471–82. Chattopadhyay G, Chattopadhyay S. Modeling and prediction of monthly total ozone concentrations by use of an artificial neural network based on principal component analysis. Pure Appl Geophys 2012;169:1891–908. Chattopadhyay G, Chattopadhyay S, Chakraborthy P. Principal component analysis and neurocomputing-based models for total ozone concentration over different urban regions of India. Theor Appl Climatol 2012;109:221–31. Cristofanelli P, Bonasoni P, Tositti L, Bonafe U, Calzolari F, Evangelisti F, et al. A 6-year analysis of stratospheric intrusions and their influence on ozone at Mt. Cimone (2165 m above sea level). J Geophys Res 2006;111:D03306. Crutzen P, Lawrence M, Poschl U. On the background photochemistry of tropospheric ozone. Tellus 1999;A 51:123–46. Cybenko G. Approximation by superpositions of a sigmoidal function. Math Control Signals Syst 1989;2:303–14. David LM, Nair PR. Diurnal and seasonal variability of surface ozone and NOx at a tropical coastal site: association with mesoscale and synoptic meteorological conditions. J Geophys Res 2011;116:D10303. Davies TD, Schuepbach E. Episodes of high ozone concentrations at the earth's surface resulting from transport down from the upper troposphere/lower stratosphere: a review and case studies. Atmos Environ 1994;28(1):53–68. Dawson JP, Adam PJ, Pandis SN. Sensitivity of ozone to summertime climate in the eastern USA: a modeling case study. Atmos Environ 2007;41:1494–511. Dimitriou K, Kassomenos PA, Paschalidou AK. Assessing the relative risk of daily mortality associated with short term exposure to air pollutants in the European Union through air quality indices. Ecol Indic 2011;27:108–15. Directive 2008/50/EC of the European Parliament and of the Council of 21 May 2008 on ambient air quality and cleaner air for Europe. Off J Eur Union, L 152/1–44. Ebi K, McGregor G. Climate change, tropospheric ozone and particulate matter, and health impacts. Environ Health Perspect 2008;116(11):1449–55. EEA. Ozone position paper (final version). Prepared by the Ad-Hoc Working Group on Ozone Directive and Reduction Strategy Development. Luxembourg: Office for Official Publications of the European Communities; 1999. EEA. Costs of ozone pollution. [Available on:] http://www.eea.europa.eu/themes/ air/air-quality/more-about-air-pollutants/ozone-1/impacts/costs-of-ozonepollution, 2008. EEA. Air quality in Europe — 2011 report. Copenhagen: Publications Office of the European Union; 2011. Elbern H, Kowol J, Sladkovic R, Ebel A. Deep stratospheric intrusions. A statistical assessment with model guided analyses. Atmos Environ 1997;31(19):3207–26. Emmons LK, Hess PG, Lamarque J-F, Pfister GG. Tagged ozone mechanism for MOZART-4, CAM-chem and other chemical transport models. Geosci Model Dev 2012;5: 1531–42. Fadnavis S, Chakraborty T, Beig G. Seasonal stratospheric intrusion of ozone in the upper troposphere over India. Ann Geophys 2010;28:2149–59. Felker SR, Moody JL, Wimmers AJ, Osterman G, Bowman K. A multi-sensor upper tropospheric ozone product (MUTOP) based on TES Ozone and GOES water vapour: derivation. Atmos Chem Phys 2011;11:6515–27. Fernández-Fernández MI, Gallego MC, García JA, Acero FJ. A study of surface ozone variability over the Iberian Peninsula during the last fifty years. Atmos Environ 2011; 45:1946–59. Fishman J, Crutzen P. The origin of ozone in troposphere. Nature 1978;274:855. Fontes T, Silva LM, Silva MP, Barros N, Carvalho AC. Predicting the origin of ozone episodes using artificial neural networks. 16th International Union of Air Pollution Prevention and Environmental Protection Associations (IUAPPA) Congress, Cape Town, South Africa; 2013. p. 94. Ganguly N. Influence of stratospheric intrusion on the surface ozone levels in India. ISRN Meteorology; 2012 [Article ID625318]. Grewe V. The origin of ozone. Atmos Chem Phys 2006;6:1495–511. Hájek P, Olej V. Ozone prediction on the basis of neural networks, support vector regression and methods with uncertainty. Ecol Informat 2012;12:31–42. Halonen J, Lanki T, Tiittanen P, Niemi J, Loh M, Pekkanen J. Ozone and cause-specific cardiorespiratory morbidity and mortality. J Epidemiol Community Health 2010;64: 814–20.

207

Hamade A, Rabold R, Tankersley C. Adverse cardiovascular effects with acute particulate matter and ozone exposures: interstrain variation in mice. Environ Health Perspect 2008;116(8):1033–9. Haykin S. Neural networks and learning machines. 3rd Edition. Prentice Hall; 2009. He H, Garcia E. Learning form imbalanced data. IEEE Trans Knowl Data Eng 2009;21(9): 1263–84. Heal MR, Heaviside C, Doherty RM, Vieno M, Stevenson DS, Vardoulakis S. Health burdens of surface ozone in the UK for a range of future scenarios. Environ Int 2013;61:36–44. Hegglin MI, Shepherd TG. Large climate induced changes in ultraviolet index and stratosphere-to-troposphere ozone flux. Nat Geosci 2009;2:687–91. Hess PG, Lamarque J. Ozone source attribution and its modulation by the arctic oscillation during the spring months. J Geophys Res 2007;112:D11303. Hess PG, Zbinden R. Stratospheric impact on tropospheric ozone variability and trends: 1990–2009. Atmos Chem Phys 2013;13:649–74. IBM. IBM SPSS Statistics, Version 21.0, 2012, Chicago, IL, USA; 2012. Im U, Incecik S, Guler M, Tek A, Topcu A, Unal YS, et al. Analysis of surface ozone and nitrogen oxides at urban, semi-rural and rural sites in Istanbul, Turkey. Sci Total Environ 2013;443:920–31. IPQ. Norma Portuguesa - Requisitos gerais de competência para laboratórios de ensaio e calibração (ISO/IEC 17025:2005). Instituto Português da Qualidade; 2005. June W, Hui-Jun W. The relationship between total ozone and local climate at Kunming using Dobson and TOMS data. Atmos Ocean Sci Lett 2010;3:207–12. Marques de Sá J, Silva LM, Santos JM, Alexandre LA. Minimum error entropy classification. Studies in computational intelligence. Springer; 2012. p. 420. MathWorks MATLAB and Statistics Toolbox Release 2012. Natick, Massachusetts, United States: The MathWorks, Inc.; 2012. Mintz R, Young BR, Svrcek WY. Fuzzy logic modeling of surface ozone concentrations. Comput Chem Eng 2005;29:2049–59. Moustris KP, Nastos PT, Larissi IK, Paliatsos AG. Application of Multiple Linear Regression Models and Artificial Neural Networks on the surface ozone forecast in the Greater Athens Area, Greece. Adv Meteorol 2012;2012. Article ID 894714, 8. Özbay B, Keskin GA, Doğruparmak ŞÇ, Ayberk S. Predicting tropospheric ozone concentrations in different temporal scales by using multilayer perceptron models. Ecol Informat 2011;6(3–4):242–7. Reiter R, Munzert K, Kanter HJ, Poetzl K. Cosmogenic radionuclides and ozone at a mountain station at 3.0 km a.s.l. Arch Meterol Geophys Bioklimatol 1983:131–60. San José R, Stohl A, Karatzas K, Bohler T, James P, Pérez JL. A modelling study of an extraordinary night time ozone episode over Madrid domain. Environ Model Software 2005;20(5):587–93. Sellitto P, Bojkov BR, Liu X, Chance K, Del Frate F. Tropospheric ozone column retrieval at northern mid-latitudes from the Ozone Monitoring Instrument by means of a neural network algorithm. Atmos Meas Tech 2011;4:2375–88. Silva L, Marques de Sá J, Alexandre LA. Data classification with multilayer perceptrons using a generalized error function. Neural Netw 2008;21(9):1302–10. Slàdkovic R, Munzert K. Lufthygienisc-klimatologische Uberwachung im bayrischen Alpenarum: Ozonspitzen auf der Zugspitze durch Zustrom aus der Stratosphäre. Final report 908080 — VI.4. Fraunhofer-Institut für Atmosphärische Umweltforschung; 1990. Solomon S, Qin D, Manning M, Chen Z, Marquis M, Averyt KB, et al. Contribution of Working Group I to the fourth assessment report of the Intergovernmental Panel on Climate Change. Cambridge, United Kingdom and New York, USA: Cambridge University Press; 2007 [http://www.ipcc.ch/publications_and_data/ar4/wg1/en/ch7s74-4.html#table-7-9]. Sousa SIV, Martins FG, Alvim-Ferraz MCM, Pereira MC. Multiple linear regression and artificial neural networks based on principal components to predict ozone concentrations. Environ Model Software 2007;22(1):97–103. Srebot V, Gianicolo EAL, Rainaldi G Trivella M, Sicari R. Ozone and cardiovascular injury. Cardiovasc Ultrasound 2009;7:30. Stevenson DS, Dentener FJ, Schultz MG, van Noije TPC Ellingsen K, Wild O, Zeng G, et al. Multimodel ensemble simulations of present-day and near-future tropospheric ozone. J Geophys Res 2006;111(D8):27. Stohl A, Spichtinger-Rakowsky N, Bonasoni P, Feldmann H, Memmesheimer M, Scheel HE, et al. The influence of stratospheric intrusions on alpine ozone concentrations. Atmos Environ 2000;34(9):1323–54. Taormina R, Mesin L, Orione F, Pasero E. Forecasting tropospheric ozone concentrations with adaptive neural networks. Proceedings of International Joint Conference on Neural Networks, San Jose, California, USA, July 31–August 5, 2011; 2011. Tsai CH, Chang LC, Chiang HC. Forecasting of ozone episode days by cost-sensitive neural network methods. Sci Total Environ 2009;407(6):2124–35. Wang W, Lu W, Wang X, Leung A. Prediction of maximum daily ozone level using combined neural network and statistical characteristics. Environ Int 2003;29:555–62. WHO. Air quality guidelines — global update 2005. Copenhagen, Denmark: World Health Organization Regional Office for Europe; 2006. WMO. WMO TD No. 1073 — global atmosphere watch measurements guide. [Last access on 30 November of 2012] http://www.wmo.int/, 2001. Zhang H, Zhang W, Palazoglu A, Sun W. Prediction of ozone levels using a Hidden Markov Model (HMM) with Gamma distribution. Atmos Environ 2012;62:64–73.

Can artificial neural networks be used to predict the origin of ozone episodes?

Tropospheric ozone is a secondary pollutant having a negative impact on health and environment. To control and minimize such impact the European Commu...
1MB Sizes 2 Downloads 3 Views