Content uploaded by Ari Karppinen
Author content
All content in this area was uploaded by Ari Karppinen on Nov 18, 2014
Content may be subject to copyright.
30
The Extraction and Fusion of Meteorological and Air
Quality Information for Orchestrated Services
Lasse Johansson
The Finnish Meteorological Institute,
Dept. of Atmospheric composition
Erik Palmenin aukio 1
00101, Helsinki, Finland
lasse.johansson@fmi.fi
Victor Epitropou
and Kostas Karatzas
Aristotle University of Thessaloniki,
Dept. of Mechanical Engineering,
54124 Thessaloniki, Greece
Leo Wanner
Catalan Institute for Research and
Advanced Studies,
Dept. of Information and
Communication Technologies,
Pompeu Fabra University, Barcelona,
Spain
Ari Karppinen
and Jaakko Kukkonen
The Finnish Meteorological Institute,
Dept. of Atmospheric composition
Stefanos Vrochidis
and Ioannis Kompatsiaris
Information Technologies Institute, Centre for Research
and Technology Hellas, Thessaloniki, Greece
ABSTRACT
The PESCaDO system (Personal Environmental Service
Configuration and Delivery Orchestration) aims at providing
accurate and timely information about local air quality and
weather conditions in Europe. The system receives environment
related queries from end users, discovers reliable environmental
multimedia data in the web from different providers and
processes these data in order to convert them into information and
knowledge. Finally, the system uses the produced information to
provide the end user a personalized response. In this paper, we
present the general architecture of the above mentioned system,
focusing on the extraction and fusion of multimedia
environmental data. The main research contribution of the
proposed system is a novel information fusion method based on
statistical regression modelling that uses as input data land use
and population density masks, historic track-record of data
providers as well as an array of atmospheric measurements at
various locations. An implementation of this fusion model has
been successfully tested against two selected datasets on air
pollutant concentrations and ambient air temperatures.
1. INTRODUCTION
Recently, the emergence of social media, personalized web
services and the increased public awareness of environmental
conditions that impact the quality of life have resulted in the
demand for easier access to environmental information tailored to
personal requirements. In particular, in case of the atmospheric
environment, there is a need for an integrated assessment of the
impact of air pollution, allergens and extreme meteorological
conditions on public health [9], [8]. In addition, this information
has to be disseminated to citizens in an easily accessible form [7].
Getting a direct answer to a seemingly simple question such as
“How will the air quality be tomorrow in Glasgow?” involves
extensive manual search and expert interpretation of the often
contradictory and heterogeneous information found on various
web sites. Furthermore, a significant portion of air quality and
meteorological information is published on the Internet only in
the form of colour-mapped, geo-referenced images [1]. Also the
quality of information might vary significantly in reliability and
relevance with respect to the queried location and time. On the
other hand, even biased and inaccurate information about air
quality could be utilized effectively by data fusion methods in
order to provide reliable information. The success of fusing
multiple model results is evident in the case of models with no
major deviation of forecasting performance, and has been
demonstrated in many related studies [22].
In this context, in [5] it has been presented an approach to
provide air quality information for any location within a large
geographical domain, by fusing air quality data from multiple
sources, by using a statistical air pollution model (RIO). In a
review of land use regression (LUR) models it has been stated
that LUR-models have been very successful in predicting annual
mean concentrations of NO2 and PM2.5 in urban environments [4].
However, these state-of-the-art LUR models are difficult to utilize
for the accurate prediction of hourly concentration of air
pollutants – a more dynamic approach is needed. Another
complication is the extremely heterogeneous nature of input data
which may contain model forecasts and observations, both with
varying reliability, time of validity and location. Spatial and
temporal gaps are also a matter of concern; there are only a finite
number of measurement stations, and forecasting models also
have a finite spatial and temporal resolution. These considerations
lead to the need to use some form of data interpolation either in
space or time, or both.
In this paper, we aim to describe the general architecture of the
PESCaDO system, focusing especially on the fusion of extracted
information [20], [21]. First, we discover environmental nodes
(i.e. web resources that include environmental measurements),
which are relevant to the area of interest. Then, a specific service
called AirMerge is presented, which is capable of performing
extraction and fusion of information from a wide range of online
Chemical Weather (CW) forecasting systems. The online fusion
service is then presented; this is a general method for the fusion of
Copyright © by the paper’s authors. Copying permitted only for private
and academic purposes.
In: S. Vrochidis, K. Karatzas, A. Karpinnen, A. Joly (eds.): Proceedings of
the International Workshop on Environmental Multimedia Retrieval
(EMR2014), Glasgow, UK, April 1, 2014, published at http://ceur-ws.org
31
processed meteorological and air quality data, and is also the
main topic of this paper. There are many definitions of data
fusion, as it is a method that is applied to various scientific
domains, such as remote sensing, meteorological forecasting,
sensor networks, etc. [19]. We use the term “fusion” to describe
the process of integration of multiple data and knowledge into a
consistent, accurate, and useful representation. An evaluation of
the performance of this fusion system is presented for two
selected cases: i) the fusion of atmospheric temperature forecasts
and (ii) the fusion of measured NO2 concentrations.
2. FRAMEWORK
We present here an overview of the general architecture of the
PESCaDO system. For a more detailed description, the reader is
referred to [20], [21].
2.1 An overview of the PESCaDO system
The purpose of the PESCaDO system is to address the need for
timely personalized environmental information (see
www.pescado-project.eu for more information). It first processes
user queries, based on the personal information on the user,
formulated in terms of a user profile. For instance, health
conditions such as asthma may affect the displayed warnings and
recommendations while the user group (e.g. citizen or
administrative expert) affects the level of detail and technicality
of the response.
Figure 1a-b: A simplified schematic diagram of the PESCaDO
system, starting from the user defined query and ending at
the delivery of response (a). An example response for the user
is presented in figure (b).
The queries are formulated in terms of PESCaDO’s Problem
Description Language via an interactive web interface. First, the
system discovers environmental nodes that contain measurements
for the areas of interest. Then, for each query, (i) relevant
environmental data sources are orchestrated, (ii) data from textual
and image formats in the sources are identified, extracted, fused
and reasoned over to assess the relevance of the data for the user,
and (iii) his query and the outcome are presented in terms of a
bulletin in the language of the preference of the user.
Figure 1a illustrates the information flow of PESCaDO from
the viewpoint of the Fusion Service, which is the backbone of the
system. The system includes two uncoupled process chains, called
here as pipelines, that operate in offline and online modes. In the
offline pipeline, environmental websites that cover the region
targeted by the user are searched for in the web and data are
extracted from the identified sites and fed into the database of the
system. We use the term ‘offline’ here since at the time of user
query the data used by the pipeline has already been retrieved,
processed and stored into a local database. In the online pipeline
user queries are processed and answered. The online pipeline
starts from the specification of personal information and query by
the user. With this information, the system first determines which
aspects of environmental and contextual knowledge (e.g.
temperature, CO2 concentration, etc.) are relevant to the user and
his query (cf. Fig, 1, Relevant aspects determination). Next, the
Fusion Service (FS) is given a request to produce fused
information about the identified relevant aspects. At this stage,
the system retrieves information from the database and starts to
process it. The ‘relevant aspects’ could be, for instance, “NO2
concentration and ambient air temperature, tomorrow between
12:00 and 18:00 in a specified region in Helsinki, given the
reported traffic density”. Furthermore, the user profile
(administration personnel vs. citizen; healthy individual vs.
allergic, etc.) affects the way the response is ultimately presented
to the user (relevant aspects determination).
The Data Retrieval Service (DRS) serves as an interface,
through which other PESCADO services can retrieve information
(i.e. environmental measurements) from the database. The Fusion
Service queries the DRS to receive environmental data available
for the requested geographic areas and time periods for all related
environmental aspects. After the FS fuses the data retrieved from
the DRS these are inserted to the PESCaDO Knowledge Base
(KB).
The PESCaDO’s KB contains, manages and provides
information represented with the PESCaDO ontology to other
services [13]. This KB also provides the Fusion Service with
supporting information needed in the fusion process. This
includes source identification and fixed coordinates if available,
and source reliability. Furthermore, the PESCaDO ontology helps
to translate verbal ratings into numeric form if needed. For
instance, the expression “heavy rain” can be converted into mm/h
numeric value with the help of the concept definitions in the
ontology. More specifically, the KB is queried about the upper
and lower limit for “heavy rain” in the specified region and then
the average value of the returned limits can be taken to represent
the input in numeric format - an approach related to the use of
fuzzy logic methods in air quality problems [6].
Once all input data are in numeric form, the FS fuses the data
by one variable (e.g. temperature, wind speed, NO2 or O3) at a
time, utilizing available uncertainty metrics for each information
source given by the Uncertainty Metrics tool (UMT). Fused data
are stored in the KB and then the tasks, including the selection,
structuring and presentation of the information resulting from the
fusion to the user can be carried on. In parallel, the retrieved
32
information, which can be used for performance evaluation later
on, is passed to UMT and stored. Using this stored information,
UMT evaluates measured values against forecasts autonomously
and produces updated source node uncertainty metrics.
2.2 Discovery of environmental nodes
As described in the previous section, the first step realized by
the PESCaDO framework is the discovery of environmental
nodes. The huge number of the nodes, their diversity both in
purpose and content, as well as, their widely varying and a priori
unknown quality, set several challenges for the discovery and the
orchestration of these services [21].
The PESCaDO discovery framework combines the main two
methodologies of internet domain specific search: (a) the use of
existing search engines for the submission of domain-specific
automatically generated queries, and (b) focused crawling of
predetermined websites [23]. To support domain-specific search
using a general purpose search engine [12], two types of domain
specific queries are being formulated: the basic and the extended.
Basic queries are produced by combining environmental related
keywords (e.g. weather, temperature) with geographical data (e.g.
city names). Extended queries are generated by enhancing the
basic queries with additional domain-specific keywords, which
are produced using the keyword spice technique [14]. Both types
of queries are then submitted to Yahoo BOSS API search engine.
In parallel, a focused crawler is employed, built upon the
Apache Nutch -crawler and is based on [18]. This implementation
attempts to classify sites by using hyperlink and text information
(i.e. anchor text and text around the link) with the aid of a
supervised classifier. This approach is new in comparison to a
previously presented method for web-based information
identification and retrieval with the aid of a domain vocabulary
and web-crawling tools [2].
The output of both techniques is post-processed in order to
improve the precision of the results by separating relevant from
irrelevant nodes and categorizing and further filtering the relevant
nodes with respect to the types of environmental data they
provide (air quality, pollen, weather, etc.). The determination of
the relevance of the nodes and their categorization is done using a
supervised classification method based on Support Vector
Machines (SVM). The SVM classifiers are trained with manually
annotated websites and textual and visual features extracted from
the environmental nodes. The textual features are key phrases and
concepts extracted from the metadata and content of the
webpages using KX [15] and the vector representation is based on
the bag of words model. The visual features (MPEG-7, [17]) are
extracted from the images included in the discovered websites in
order to identify heatmaps that are usually present in air quality
forecast websites.
2.3 Orchestration of environmental nodes
and data extraction
Once the environmental nodes have been detected and indexed,
they are available as data sources or as active data consuming
services (if they require external data and are accessible via a web
service API).
To distil data from text, advanced natural language parsing
techniques are applied, while to transform semi-structured web
content into structured data, regular expressions and HTML trees
are used. Data extraction from images focused on heatmap
analysis using the AirMerge system, described in the following
section.
2.4 AirMerge subsystem
A significant portion of Air Quality (AQ) related information
(in particular, Chemical Weather forecasts) is published on the
Internet only in the form of colour-mapped geo-referenced
images. Such image-based information is impossible to be parsed
via usual text-mining and screen-scraping techniques used in web
mash-up-like services. It was thus important to provide
PESCaDO with a specialized service that allows accessing and
using CW forecast images as another source of data to use during
the Orchestration and Fusion phases. Such a system, called Air
Merge, has already been developed and described in [3], [1].
AirMerge is an open access system, which is currently
dedicated to the whole European continent (the coverage of
different territories is possible, accessing a wide number of
environmental nodes containing CW information, and can
automatically extract data from various data sources). These
images commonly have geographical spatial resolutions ranging
from 1x1 km to 20x20 km, and temporal resolutions from a
minimum of one hour to an entire day [10]. The reported values
usually are maximum or average air pollution concentration
values for the selected integration time.
Figure 2: Example of a PM2.5 forecast (produced by MACC)
conversion process using AirMerge. Bitmap data (a) is
transformed into numerical form by using the colour scale c).
The heatmap a) has been reproduced in b) using the
converted numeric grid.
In the context of PESCaDO, AirMerge apart from performing
image extraction, it acts as an autonomous web-crawling, parsing
and database-storage mechanism for CW forecasts, using its own
means and processes which are distinct from those of PESCaDO,
having been developed independently. The harvested data cover
most of Europe for a time period going back to August 2010
when it first became operational. Time resolutions range from one
hour to a day, depending on the capabilities of the sources used.
A typical set of CW models and the resulting images can be
found in the European Open-access Chemical Weather
Forecasting Portal described by [1], that has been developed in
33
the frame of COST Action ES0602 (www.chemicalweather.eu).
AirMerge is able to convert such image-based concentration maps
into numerical, geographically referenced data, accounting for
geographical projections, missing data, noise and the differences
in publishing formats between different model providers. The
result is the effective conversion of image data back into
numerical data, which is then made directly available for a
number of numerical processing applications.
It should be clarified that in the proposed system AirMerge has
two roles: a) it performs image data extraction and b) it is an
additional environmental node that provides environmental data
encoded in images.
3. FUSION OF EXTRACTED
INFORMATION
The fusion of information in an orchestrated service such as
PESCaDO, offers several advantages to the user. First, the output
of the system includes only one set of values instead of an
extensive collection of pieces of information that may not agree
with each other. Secondly, the fusion result will be of a better
quality with respect to the individual sources. Third, small
geographic and temporal gaps in the input data can be
extrapolated.
The above mentioned services for environmental node
discovery and data retrieval guarantee a large amount of relevant
input data which need to be fused with respect to the user defined
query. However individual competing pieces of information from
different nodes can seldom be regarded as equally relevant and
thus a general measure for information relevance and quality is
needed for data fusion.
In the fusion process, all pieces of meteorological and air
quality data correspond to a certain time and place. These pieces
of information can be regarded as statistical estimators
or in short, in which is distance and is time, for the
conditions governing the area and time of interest for the user:
(1)
where / is the coordinate vector for the location of interest /
location associated with the estimator, / is the time of
interest / estimator time and is the estimator error. For sensors
the estimator time is simply the time of measurement. The
algorithm that is used in calculating the fused value requires
information about the statistical properties of , namely the
expected variance of . Thus, a detailed description of the
evaluation of is given. The fusion service estimates an
aggregate statistical variance measure for each and these
variance measures are then used for the assignment of averaging
weights to each . Essentially a large estimated aggregate
variance causes the assigned weight to decrease, while the data
from the more accurate and relevant sources are assigned larger
weights and gain more emphasis in the fusion.
3.1 Variance estimation
The variance of , , is affected by the information
source’s capability to properly assess the phenomenon of interest.
In addition, information about air pollutant concentrations and
weather conditions loses accuracy rapidly as a function of the
temporal interval between the measurement time and the time of
interest defined by the user. Furthermore, a data point near
should always get a larger weight in the fusion in contrast to other
data points that describes the conditions in more remote locations.
Thus, we assume that the variance related to is the sum of these
three individual (independent and thus summable) components,
given by
(2)
where is the variance component as function of , is
the temporal variance component as a function of , in which
(3a)
(3b)
in Eq. 2 describes the information source’s
inherent quality in terms of variance, i.e., the capability to
estimate at point-blank range when and are equal to
zero. For the evaluation of , stored information
about the source’s prediction accuracy in past can be used,
evaluated by the Uncertainty Metrics Tool (see Fig 1). More
specifically, measurements and model forecasts are paired
together if they represent the same time and location and the
statistical variance is then calculated for the population of
evaluation pairs.
In the presented PESCaDO framework, the location for the
estimator may not have been defined exactly; this is
usually the case, for instance, with extracted weather forecasts for
cities. In these cases actually pinpoints the center of city while
information represents the conditions through-out the city. In such
cases the coordinates are flagged as approximations and set
, where is the radius of the city.
The variance models and can be formulated with
statistical methods. In the fusion service these have been
formulated individually for each air pollutant species using
regression analysis with historical measurement data. For the pilot
application of the method, these data represent 6 to 43
measurement stations across Finland, depending on the measured
values. More specifically, the following simple regression models
are employed:
(4a)
(4b)
where parameters and are defined with statistical
regression techniques. More complex regression models were also
studied but the added benefit for using more natural, logarithmic
regression models was negligible; the achieved correlation of
polynomial models is generally very high for the temporal
domain of interest (τ < 36h). In the formulation of , the
measurement station’s capability to predict the measured
phenomenon at a distance of (covariance of the two time series)
is evaluated.
3.2 Optimal weight calculation
Assuming all data sources to be independent and the estimators
to be non-biased ( ), an optimal fused value
can be calculated according to [16] given by:
(5)
where individual weights is given by
(6)
34
To assure statistical independence of .., only the most
relevant estimator per data source is selected for the fused
value calculation in Eq. 5. If a collection of estimators
{} is available from the same source, the
selected to represent the source is simply the one with the
lowest from the collection. In the particular case for
extracted time series from measurement stations, the estimator
which has the smallest is selected to represent the source, as
and the base variance are the same for all ...
Theoretically, it can be shown that the fused value is
the optimal estimator in terms of mean squared error and that the
prediction accuracy increases while the number of independent
data sources (n) is increased [16]. More importantly,
does not suffer from low quality input data, as long as in
Eq. 2 has been estimated reasonably well.
3.3 Bias correction
In the algorithm presented in section 3.2, it was assumed that
each is an unbiased estimator for the conditions in at the
time . Local air quality measurements from a different
environment, however, are usually significantly biased estimators
for the conditions in other nearby environments. Moreover, the
hour of day may even contribute to the bias (consider a
measurement station near a busy road during the morning traffic).
Thus, in order to use Eq. 5 effectively, the fusion service utilizes a
geographic profiling feature to detect and automatically remove
this kind of structural bias from the estimators. The fusion service
was incorporated with high-resolution land use and population
density masks for Finland (the selected domain for the PESCaDO
prototype). For land-use, a dataset from CORINE with a
resolution of 50m x 50m is being used. For population density
data (for 2010), the fusion service has the prototype domain
covered with a resolution of 250m x 250m. These two data
sources are used for profiling and comparing the differences
between the environments in and and ultimately, is
polished into a non-biased estimator for . The profiling
is done as follows:
- The surrounding land use (with evaluation radius of
200m) and population density (a wider evaluation
radius of 6km) for both and is evaluated.
- The evaluated environment is expressed as a collection
of selected land-use frequencies and population density.
This collection is referred to as a profile in this paper
(Fig 3).
After the evaluation of profiles, the difference between the
expected values is evaluated. Let be the estimator profile
and be the evaluated profile corresponding to the user
defined location and time. Then, a bias corrected estimator
is given by
(7)
where is the expected hourly concentration of the
pollutant at the estimator’s location at time and is
the expected pollutant concentration in the user defined location
at the time .
The evaluation of Eq. 7 requires yet another statistical model
(for each pollutant) to calculate the expected concentration as a
function of time and key land-use frequencies. Such a set of
statistical models has been implemented with the fusion service,
using the archived measurement time series in Finland as
calibration data: the environments around the stations were
evaluated and multi-variable regression was applied. The
regression was repeated with several different land-use and
population evaluation radii; the best correlation was achieved
with the abovementioned values (land-use with a 200m radius,
population density with a 6km radius). Nevertheless, this
mathematically intensive regression procedure is not discussed in
this paper further although for the NO2 pollutant, a demonstration
of the profiling method and its capability to predict the expected
hourly concentration is presented in section 4.1.
Figure 3: Profile evaluation with land use and population
density maps. The larger circle represents the area for local
population determination and the smaller red circle
represents the area for land use determination. Satellite image
provided by Google Earth.
As discussed in at the beginning of section 2.1 the fusion
service stores measurements as evaluation material for individual
service providers and models. Thus for another completely
different region other than Finland, the regression parameters for
profiling can be set without a fixed set of calibration material; the
stored measurements that have flown through the PESCaDO
system can be further exploited by setting up the regression
parameters for profiling automatically as the number of
measurements builds up over time. In this sense the profiling
feature within the Fusion Service is adaptive.
The presented bias correction method offers yet another
advantage: episodes that affect air quality on a major scale, such
as forest fires, are automatically accounted for if the input data
contains some measurements from the episode-driven locations.
For instance, if a background station has measured an
exceptionally high concentration of NO2, then the expected NO2
concentration at a nearby urban environment is going to be
reflected on the episode-affected background concentration.
4. RESULTS
The performance of the presented environmental information
fusion method was evaluated using temperature forecasts
provided by four well known weather service providers (FMI,
SMHI, Met Norway and Weather Underground). For 43 locations
around Finland weather forecasts were extracted from respective
online sites and stored during several months in 2012. Uncertainty
metrics in terms of for individual SPs were
evaluated by comparing measured temperature values against
individual stored forecasts for each SP; a total of 2500 forecasted
versus measured temperature -pairs for each SP were gathered in
order to get statistically meaningful estimates as
a function of forecasted period length. Then, fused forecasts
(temperature of the next 3 days) for the locations in August 2012
were produced on a daily basis for each of these locations using
the stored forecasts.
35
In Figure 4, the mean absolute error of temperature forecasts
and the fused forecast is presented. According to the figure fused
temperature forecasts have the lowest mean error with just four
different SPs providing forecasts simultaneously. This result goes
to show that the well-known benefits of forecast fusion can be
exploited within web services such as PESCaDO when the
performance of forecast providers is being monitored.
Figure 4: Mean absolute error of temperature (C) forecasts
and the fused forecast for different forecast time spans.
Forecasted and measured data for 43 different locations and
time periods in august was used.
4.1 Performance of the environmental
profiling feature
The environmental profiling feature of the fusion service was
calibrated using measurement time series from Finland during
2010. To test the performance of this novel feature, 8 different
NO2 measurement stations with varying environments were
selected in 2011, and the observed hourly concentrations were
compared against the values predicted with the aid of the profiling
feature. The profiling feature differentiates working days and
weekends and for this test, the working days were selected.
It can be seen from the figures 5a-h that the profiling feature is
able to predict the expected average NO2 concentration well in
various different environments. Background areas, urban and
rural, fare better in the comparison while the traffic-intense
environments are more difficult to predict. This is to be expected
as the actual traffic volumes have to be derived using only the
local population and road intensity. As a consequence, the
profiling feature inevitably underestimates the expected
concentration near large motorways that have a small surrounding
population.
4.2 Comparison of measured and predicted
NO2 time series
The performance of the fusion of air quality measurements
with the presented methodology was tested with NO2
measurements in Southern Finland. Measurement time series for
February 2011 from the available stations (n = 20) were used as
input data and fused NO2 concentrations were calculated for a
remote location for which comparison time series was readily
available. The domain for the test can be seen from Figure 6
which illustrates the fused concentration of NO2 at one of the
hours of interest.
Figure 5a-h: Predicted and observed hourly average
concentration of NO2 during working days (Monday to
Friday) in several measurement sites. Predicted values have
been obtained by evaluating the station’s environment with
the aid of the profiling feature.
Figure 6: Fused NO2 concentration in Southern Finland in
2011 at 07:00.
The highest concentration can be found at the centre of
Helsinki, which resides in the bottom-right corner of the figure.
The remote test area is a small city centre (Lohja), located
approximately 70 kilometres to the right of Helsinki – 50
kilometres away from the nearest measurement station. The fused
values were compared against the on-site measurements in the
test area and results are shown in Figure 7.
The comparison between fused and measured NO2
concentration at the test site (Figure 7) shows that the pollutant
concentration has been estimated fairly accurately with the
presented method.
During the study period the mean absolute error between
predicted and measured NO2 hourly concentration was of the
order of 7 µg/m3 (mean = 12µg, Var = 107 µ2g2). This error is
significantly less than the achieved mean error when a
conventional geographical extrapolation method would be used:
36
using inverse distance weighting (IWD), [11] the resulting mean
absolute error would be 14 µg/m3.
Figure 7: The observed and predicted NO2 concentration
during February 2011 at the test site, the centre of Lohja city.
Figure 8 illustrates a collection of mean absolute prediction
errors from calculations similar to the one presented in Fig 7. One
by one, the measurement stations were removed from the input
data and the removed time series was compared against the fused
time series which was produced using the remaining data.
According to Fig 8 if the locations for near-by measurements
represents similar environment than the location for IWD
extrapolation (Laune station, Tikkurila station of Fig 8), then the
IWD extrapolation may be able to predict the hourly
concentration fairly well. Otherwise, the IWD-method without
bias correction capabilities produces generally poor estimates in
terms of mean absolute error whereas the fusion service performs
well regardless of the collection of estimators used as input.
Indeed, Luukki station, a rural NO2 background measurement
station is an example of this; there are several urban measurement
stations nearby and thus the hourly concentration of NO2 in
Luukki cannot be extrapolated with conventional methods.
Figure 8: Comparison of IWD extrapolation and the
presented fusion method in terms of standard deviation.
Observed average describes the average hourly NO2
concentration at measurement site.
5. CONCLUSION
To provide timely meteorological and air quality related
information to citizens and administrative user alike, a prototype
service PESCaDO was developed. By combining the data
discovery, extraction and fusion methods, described in this paper,
it possible to produce accurate and personalized information to
the users. Unlike several search engines, the user is not confused
by the sheer amount of presented data and suggestions; instead,
the user is provided with a single, understandable yet precise
answer. This is also what separates PESCaDO from a
conventional, generic search engine. The self-maintaining design
of PESCaDO system facilitates the discovery and indexing of
new information sources. The source provider’s performance can
be evaluated and stored on a continuous basis and the stored
performance data can be used to guide the fusion of information.
Furthermore, the measured air quality and meteorological data
that flows through the system can be used in the calibration of the
fusion service’s various statistical models effectively allowing the
system to adapt into different regions.
The fusion method offers several advantages for the PESCaDO
system. For instance, it is not necessary to discard any extracted
information as the algorithm takes care that the irrelevant input is
not over-emphasized. In this paper, a demonstration of the fusion
of temperature forecasts was given. It was shown that the fused
temperature forecast in fact had the lowest margin of error, which
goes to show the benefits to be had in the fusion of information
even if the amount of service providers is small.
It was shown that the presented profiling feature of the fusion
service is able to predict hourly concentrations of NO2 in different
environments quite well. As a consequence, the fusion method
was able to outperform a conventional extrapolation method
(IWD). However, NO2 is strongly affected by urbanization and
road traffic and thus is an ideal phenomenon to be handled with
the proposed fusion method. Other pollutants however, such as
ozone and carbon monoxide are more difficult to handle with the
presented profiling feature. In fact, the static environment based
bias-removal needs to be more dynamic in the future. This could
be achieved by introducing meteorology in the fusion process. For
instance, the profile could be analysed from the wind’s direction.
Furthermore, the expected concentration could be a function of
several meteorological parameters such as rain, sky conditions
and wind speed. As a result, the PESCaDO system would be
orchestrated in another new level, where the extracted
meteorological data would be subject to fusion and used again in
the fusion of air quality pollutants.
6. ACKNOWLEDGMENTS
This work was supported by the European Commission under
the contract FP7-ICT-248594 (PESCaDO).
7. REFERENCES
[1] Balk, T., Kukkonen J., Karatzas, K., Bassoukos, A., and
Epitropou, V., European Open Access Chemical Weather
Forecasting Portal, Atmospheric Environment, 38(45),
6917–6922, 2011.
[2] Bassoukos A., Karatzas K., Kelemis A. (2005)
Environmental Information portals, services, and retrieval
systems, Proceedings of of “Informatics for Environmental
Protection- Networking Environmental Information”-19th
37
International EnviroInfo Cenference, Brno, Czech Republic,
pp. 151-155.
[3] Epitropou, V., Karatzas, K., Bassoukos, A., Kukkonen, J.
and Balk, T., A new environmental image processing
method for chemical weather forecasts in Europe,
Proceedings of the 5th International Symposium on
Information Technologies in Environmental Engineering.
Poznan: Springer Series: Environmental Science and
Engineering, 781–791, 2011.
[4] Hoek, G., Beelen, R., Hoogh, K., Viennau, D., Gulliver, J.,
Fischer, P. and Birggs, D. A review of land-use regression
models to assess spatial variation of outdoor air pollution.
Atmospheric Environment 42 (2008) 7561–7578,
doi:10.1016/j.atmosenv.2008.05.057. 2008.
[5] Janssen, S., Gerwin, D., Fierens, F. and Mensink, C. Spatial
interpolation of air pollution measurements using CORINE
land cover data. Atmospheric Environment,Volume 42, Issue
20, June 2008, Pages 4884–4903, 2008.
[6] Karatzas K. A fuzzy logic approach in Urban Air Quality
Management and Information Systems (UAQMIS),
Proceedings of the 4th International Conference on Urban
Air Quality Measurement, Modelling and Management (R.
Sokhi and J. Brexhler eds), Charles University, Prague,
Czech Republic, 25-27 March 2003, pp. 274-276, 2003
[7] Karatzas K. Informing the public about atmospheric quality:
air pollution and pollen, Allergo Journal 18, Issue 3/09, pp
212-217, 2009
[8] Karatzas, K. and Kukkonen, J., COST Action ES0602:
Quality of life information services towards a sustainable
society for the atmospheric environment, ISBN: 978-960-
6706-20-2, Thessaloniki: Sofia Publishers, 2009.
[9] Κlein Τh., Kukkonen J., Dahl Å., Bossioli E., Baklanov A.,
Fahre Vik Α., Agnew P., Karatzas, K., and Sofiev, M.,
Interactions of physical, chemical and biological weather
calling for an integrated assessment, forecasting and
communication of air quality, AMBIO,, 41(8), pp. 851-864,
2012
[10] Kukkonen, J., Olsson, T., Schultz, D.M., Baklanov, A.,
Klein, T., Miranda, A. I., Monteiro, A., Hirtl, M., Tarvainen,
V., Boy, M., Peuch, V.-H., Poupkou, A., Kioutsioukis, I.,
Finardi, S., Sofiev, M., Sokhi, R., Lehtinen, K. E. J.,
Karatzas, K., San José, R., Astitha, M., Kallos, G., Schaap,
M., Reimer, E., Jakobs, H., and Eben, K., A review of
operational, regional-scale, chemical weather forecasting
models in Europe, Atmos. Chem. Phys (12), 1-87,
doi:10.5194/acp-12-1-2012, 2012.
[11] Li, J. and Heap, A.D., 2008. A Review of Spatial
Interpolation Methods for Environmental Scientists.
Geoscience Australia, Record 2008/23, 137 pp, ISBN 978 1
921498 30 5.
[12] Moumtzidou, A., Vrochidis, S., Tonelli, S., Kompatsiaris, I.,
& Pianta, E. (2012). Discovery of Environmental Nodes in
the Web", Proceedings of the 5th IRF Conference, Vienna,
Austria, 2012.
[13] Moßgraber, J., Rospocher, M. Ontology Management in a
Service-oriented Architecture. Architecture of a Knowledge
Base Access Service. Proceedings of the 23rd International
Workshop on Database and Expert Systems Applications.
2012.
[14] Oyama, S., Kokubo, T., Ishida, T.: Domain-Specific Web
Search with Keyword Spices Awareness in Urban Areas. J.
IEEE Transactions on Knowledge and Data Engineering. 16
(1), 17—24, 2004
[15] Pianta, E., & Tonelli, S. KX: A Flexible System for
Keyphrase Extraction. Proceedings of SemEval, 2010.
[16] Potempski, S. and Galmarini, S., Est modus in rebus:
analytical properties of multi-model ensembles, Atmos.
Chem. Phys., 9, 9471–9489, doi:10.5194/acp-9-9471-
2009,2009,
[17] Sikora, T. The MPEG-7 visual standard for content
description-an overview. IEEE Transactions on Circuits and
Systems for Video Technology, 11(6), pp. 696-702, 2001
[18] Tang, T. T., Hawking, D., Craswell, N., & Sankaranarayana,
R. S. Focused crawling in depression portal search: A
feasibility study. Proceedings of the 9th Australasian
Document Computing Symposium, Melbourne, Australia,
2004.
[19] Wald, L., Some terms of reference in data fusion, IEEE
Transactions on Geosciences and Remote Sensing 37(3), pp.
1190-1193, 2001.
[20] Wanner, L., Vrochidis, S., Tonelli, S., Mossgraber, J.,
Bosch, H., Karppinen, A., Myllynen, M., Rospocher, M.,
Bouayad-Agha, N., Bügel, U., Casamayo,r G., Ertl, T.,
Kompatsiaris, I., Koskentalo, T., Mille, S., Moumtzidou, A.,
Pianta, E., Saggion, H., Serafini, L., and Tarvainen, V,.
Building an Environmental Information System for
Personalized Content Delivery. In (Hrebícek J., Schimak G.,
Denzer R. eds.): Environmental Software Systems.
Frameworks of eEnvironment - 9th IFIP WG 5.11
International Symposium, Proceedings. IFIP Publications
359, Springer, ISBN 978-3-642-22284-9, pp. 169-176, 2011.
[21] Wanner L., Vrochidis S., Rospocher M., Moßgraber J.,
Bosch H., Karppinen A., Myllynen M., Tonelli S., Bouayad-
Agha N., Casamayor G., Ertl Th., Hilbring D., Johansson L.,
Karatzas K., Kompatsiaris I., Koskentalo T., Mille S.,
Moumtzidou A., Pianta E., Serafini L. and Tarvainen V.
Personalized Environmental Service Orchestration for
Quality Life Improvement, 8th IFIP WG 12.5 International
Conference, AIAI 2012 Workshops, IFIP AICT 382 (L.
Iliadis et al., eds), Proceedings, Springer, pp.351-360., 2012
[22] Weigel, A.P, Liniger, M.A. and Appenzeller, C. Can multi-
model combination really enhance the prediction skill of
probabilistic ensemble forecasts?. QUARTERLY
JOURNAL OF THE ROYAL METEOROLOGICAL
SOCIETY. Q. J. R. Meteorol. Soc. 134: 241–260, 2008
[23] Wöber, K. Domain Specific Search Engines, In: Fesenmaier,
D. R., Werthner, H., Wöber, K. (eds.) Travel Destination
Recommendation Systems: Behavioral Foundations and
Applications, 205—226. Cambridge, MA: CAB
International, 2006.
