People with asthma, children, and older adults are the groups most at risk
2.3.2 Context Aware Outdoor Air Quality Monitoring and Prediction
As presented in the previous section, many algorithms for the forecasting of air pollutants lev-els have been developed. Some existing systems have applied these algorithms and have been designed considering context awareness. In this section of the document, those sys-tems, that were developed for the urban AQ monitoring and prediction using context aware-ness, are presented and compared to each other.
In (Dutta et al., 2009) the authors describe a personal and portable AQ monitoring system, that tells the users the air conditions around them in real time using the same AQI (US-EPA) presented in previous sections of this document. A test with 16 participants was undertaken
and each person carried a sensor that measured CO, O3 and NO2 levels, and other mete-orological factors, such as temperature, atmospheric pressure and relative humidity, during their daily commutes. They concluded that the visualization of the data and the information presented was helpful for the users. Nevertheless, this system can be further improved by adding prediction algorithms, depending on where a user will commute next, besides letting the person add their health conditions, to make the AQI output even more accurate.
In (Kurt and Oktay, 2010), the authors propose an air pollution forecasting solution that in-cludes an ANN and geographic relationship models for the sensor stations. They state that only using the time series data for each individual station’s pollutant prediction is not enough and that considering the geographical distribution of the sites can increase the accuracy of the predictions. Given that there are many stations in a certain area, they select the most in-fluential ones towards the predicting station, by using their statistical influence on each other’s data and by assigning weights to the relation according to the distance between each of them.
They conclude that their results improve the prediction over other simpler models that do not take geographical elements, which count as an extended context, into account.
The authors in (Catalano and Galatioto, 2017) state that pollution cases should be treated as individual cases that depend on the location and context they are studied in. Their idea was to develop a self-managing model framework, or meta-model, able to select, for different emission and dispersion factors of airborne pollutants, the passing prediction model from a group of alternative AQ models, which are set to consider a wide spectrum of contexts and situations. This means that if a model from a city A suits the one for a city B under certain circumstances, like rare high peaks of NO2, then it should be used when needed. They applied their model on two sites in the United Kingdom (UK) and compared it with the meta-model that decided from each of the two meta-models when needed. The meta-meta-model takes the decisions based on a simple Euclidean distance between the attribute points in the attribute space. They also compared it to a cross-site model, where they trained a MLP with data from both sites. In the first case, the meta-model showed improvement as high as 11% when predicting low frequency pollution exceeding peaks. In the next case, it showed a maximum gain of up to an impressive 113% in the site with very high NO2 hourly concentrations. They consider the context of each site for the forecasting algorithm to be applied, which is a good approach to get more accurate outputs. This approach can be even further improved by using other prediction techniques presented in the previous section of this document and by adding the user’s context as well.
Finally, in (Chen et al., 2016), a prediction approach is presented, which expands the usual context used in all the other algorithms (mainly pollutants and meteorological variables). First,
they divide the region under monitoring in a grid of equally sized squares, and consider a set of different attributes in each one of them. The characteristics considered are: i) traffic related features: a) vehicle speed and b) variance of vehicle speed over the lastn hours, ii) road-network related features: a) road density and b) road types, iii) Point of Interest (POI) related features: a) location types, i.e. stadiums, parks, factories, etc., iv) check-in features: a) human mobility in the region, v) nearby monitoring-related features: a) the AQ from neighbouring grid areas is considered as it influences the current region directly. On the data collected for these variables a semi-supervised ensemble pruning technique is applied. This method, in contrast to all the previous uses only external context to predict AQ. The results are encouraging for context driven approaches, even though they are not nearly as good as the ones obtained with DNNs, they can be used to improve the data driven techniques that have problems rec-ognizing sudden changes in a pollutant’s level. It also allows to give users a more thorough explanation of what the sources of the a given pollutant’s level are, expanding the contextual understanding.
With this final thought in mind we conclude the background section and start the introduction to our proposed system. In it we have combined the strengths of the aforementioned techniques to overcome the known issues with AQ prediction.