The amount of time evolving spatial data that is currently being generated by automatic observation processes is huge. In general, observation data consists of both heterogeneous spatio-temporal data and relevant observation metadata. The former includes data of Spatial Entities (cities, roads, vehicles, etc.) and data of temporal evolution of both properties of Spatial Entities (population of a city, position of a vehicle, etc.) and properties of space (temperature, elevation, etc.). Real uniform integrated management of all these types of data is still not achieved by current models and systems. The present paper describes the design of a data modeling and management framework for observation data warehouses. A hybrid logical-functional data model based on the concept of MappingSet and relevant language enables the specification of spatio-temporal analytical processes. The framework in currently being implemented.
According to [
Observation data has an inherent temporal nature.
Besides, in many cases FOIs are also spatial. Therefore,
systems devoted to observation data analysis should cope with
spatial and spatio-temporal data analysis. In particular,
they should support relevant functionality for the
management of Spatial Entities and Spatial Coverages, and their
evolution with respect to time [
The amount of data that is currently being obtained from automatic observation processes is huge and the estimated tendency is to have an exponential growth during the upcoming years. The analysis of all these data to support appropriate decision making is key challenge for future information systems. Many application domains exist that would benefit from innovative technologies in this area, including environmental observation and monitoring, natural disaster management, e-health, etc.
Based on the above, in the present paper a data modeling and management solution is proposed that enables spatiotemporal analysis in data warehouses of observation data. In particular, a proposed E-R extension enables the insertion of observation metadata in spatial models at a conceptual level. At a logical level, a new data model based on MappingSets enables the integrated management of any kind of spatial and temporal data. A MappingSet is a collection of Mappings, in the functional programming sense, defined on a common domain. Both Spatial Entities and Spatial Coverages and both Time-triggered and Event-triggered observation data are modeled uniformly with MappingSets.
The remainder of this paper is organized as follows. Section 2 describes other pieces of work related to the proposed solution. The MappingSet based spatio-temporal logical model is described in Section 3. The conceptual level E-R extension for observation data is described in Section 4, as it is also its translation to the MappingSet based logical model. Section 5 illustrates the spatio-temporal analysis capabilities of the model for the definition of Non-Physical spatio-temporal analytical processes. Finally, Section 6 concludes the paper and outlines lines of future work. 2.
The OGC defines an abstract specification of a data model
for Observations and Measurements [
In [
Observation data has always a temporal nature. Besides,
the spatial components of observation data and metadata is
centric to many application domains, such as those related
to environmental observation and monitoring. Spatial and
temporal extensions of conceptual and logical data models
have to be considered. Examples of spatio-temporal
conceptual models are [
Various different data management approaches are
possible to deal with spatio-temporal observation data
automatically generated by sensing devices. If we consider the
data generated by each sensor as a virtual temporal
relation, then the simplest approach is to consider Materialized
Views of such virtual relations. Automatic maintenance of
such views on the arrival of new data from sensors has to
be solved by the system [
A more sophisticated solution is to consider sensor data
streams and to enable the continuous execution of queries
on those input streams. Continuous query languages [
BASED DATA MANAGEMENT
This section introduces the MappingSet based data model that is the basis of the proposed framework. Temporal and spatial data types are first defined. Based on them Mappings and MappingSets are next formalized. Data management will be based on the intensional definition of MappingSets using both logical and functional paradigms.
Conventional data types include Boolean, CString (variable size character strings), Int16, Int32, Int64 (integers), Float32, Float64 (reals with floating point representation). Fixed point parametric type Numeric(P,D) consists of real numbers with a maximum of P digits, D of them are in the fractional part. In order to define temporal and spatial data types, 1D and 2D samplings are first formalized. Let R and I denote the set of real and integer numbers, respectively, then 1D and 2D samplings are defined as follows.
Definition 1. A 1D-sampling S with resolution r ∈ R and phase p ∈ R is defined as the infinite subset of R {x|x = i · r + p, ∀i ∈ I}
Definition 2. Let vr1, vr2, vp1 and vp2 be four vectors in R2 defined by respective directions D1, D2, D1, D2 ∈ (−π, π] and respective magnitudes r1, r2, p1 and p2. A 2Dsampling S with directions D1, D2, resolutions r1, r2 and phases p1 ∈ [−r1/2, r1/2], p2 ∈ [−r2/2, r2/2] is defined as the infinite subset of R2 {(x, y) ∈ R2| x = (i1r1 + p1) cos(D1) + (i2r2 + p2) cos(D2)∧ y = (i1r1 + p1) sin(D1) + (i2r2 + p2) sin(D2), ∀i1, i2 ∈ I}
An element s of a 1D-sampling (2D-sampling) S is called a 1D-sample (2D-sample). Integer i, i1, i2 are called the sampling coordinates of s. s(i), s(i1, i2) denote respectively the 1D-sample and 2D-sample with sampling coordinates i and (i1, i2). Figure 1 illustrates the above definitions with a geometrical representation.
Definition 3. TimeInstant(D) is defined as a finite subset of elements s(i) of a 1D-sampling S with resolution 10−D and phase 0 such that
−263 < i < 263 + 1 where each s(i) is interpreted as the time instant 1/1/1970+ s(i) seconds. Maximum allowed D is 6 (microsecond).
Definition 4. TimeInstantSample(D, R) is defined as a finite subset of elements s(i) of a 1D-sampling S with resolution R · 10−D and phase (R · 10−D)/2 such that −263 < i < 263 + 1 where each s(i) is interpreted as the time interval [1/1/1970+ s(i) seconds, 1/1/1970 + (s(i) + R · 10−D) seconds). Again, maximum allowed D is 6 (microsecond). (a) 1D Sampling p r y vp2 (0,0) s(0,0) vp1 (a) 2D Sampling s(0,1) 2 r v vr1
s(1,1) s(1,0) x
Definition 7. Point2DSample(P,D,R) is defined as the finite subset of elements s(i1, i2) of a 2D-sampling S with implementation dependent directions D1 and D2, resolutions r1 = r2 = K · R · 10−D and phases p1 = p2 = 0 such that 1. −10P < i1, i2 < 10P 2. K < max(¯¯cos( D2−D1 )¯¯ , ¯¯sin( D2−D1 )¯¯)
2 2
TimeInstant and Point2D data types provide discrete representations for both time and space, where the user has control over the supported precision. Types TimeInstantSample and Point2DSample provide representations for temporal and spatial samplings at user defined resolution. It is noticed that each time instant is approximated by its closest lower TimeInstantSample, whereas each 2D point is approximated by its closest Point2DSample. It is out of the scope of this paper to demonstrate that K factor above ensures that any 2D point is approximated by a sample at a distance lower or equal to R · 10−D. Type castings are available for the above data types.
If T is either a numeric or temporal type, then data type
Interval(T) is a new data type whose values are closed
intervals over data type T. If t1, t2 are two elements of data
type T, then [t1, t2] is used to denote the relevant closed
interval. Similarly, if S is spatial data type then the
following geometric data types are also supported, based on the
standard specification given by [
• Geometry(S): Abstract type. Represents any vector geometry or set of geometries defined with elements of S. • LineString(S): Vector polylines defined by sequences of elements of S. • Polygon(S): Vector polygons, possibly with holes, whose borders are defined by sequences of elements of S. • GeometryCollection(S): Heterogeneous collections of Geometries. • MultiPoint(S): Homogeneous collection of elements of S. • MultiLineString(S): Homogeneous collection of elements of LineString(S). • MultiPolygon(S): Homogeneous collection of elements of Polygon(S).
Definition 8. If ADT1, ADT2, . . . ADTn are not necessarily distinct data types, A1, A2, . . . , An are distinct names and RDT is a data type, then: 1. A Mapping with signature M () : RDT is defined as a value of type RDT 2. A Mapping with signature
M (A1 : ADT1, A2 : ADT2, . . . , An : ADTn) : RDT is defined as a partial function
M : ADT1 × ADT2 × ADTn → RDT
Operations are syntactic sugar for Mappings. Implicit castings between compatible data types are applied during Mapping invocations, enabling transparent transformation between temporal and spatial elements of different resolutions by applying constant interpolation. Various primitive mappings and operations are provided by the model. However, formalizing a complete set of them is out of the scope of the paper. Informal descriptions of required primitive mappings will be given throughout the paper.
A MappingSet is nothing but a set of Mappings that share a common domain defined as a n-ary relation over data types. Formalism is given below.
Definition 9. Let C1, C2, . . . , Cn be distinct names, ADT1, ADT2, . . ., ADTn be not necessarily distinct data types and RDT1, RDT2, . . ., RDTm be not necessarily distinct data types. Let also D be a n-ary relation with scheme D(C1 : ADT1, C2 : ADT2, . . ., Cn : ADTn) defined as a finite subset of ADT1 × ADT2 × . . . × ADTn. Then a MappingSet is defined in either of the three following forms: 1. A 1-tuple M S = hDi. 2. A m-tuple M S = hM1, M2, . . . , Mmi, where each Mi is a Mapping with signature Mi() : RDTi defined as a value of RDTi. 3. A (m+1)-tuple M S = hD, M1, M2, . . . , Mmi, where each Mi is a Mapping with signature Mi(C1 : ADT1, C2 : ADT2, . . . , Cn : ADTn) : RDTi defined as a partial function Mi : ADT1 × ADT2 × ADTn → RDTi.
The evolution with respect to time of spatial entities and spatial coverages may be modeled with appropriate MappingSets that contain both Domain and Mappings. n-ary relationships are also modeled with MappingSets, usually without Mappings. MappingSets without Domain are also useful to record short collections of key-value pairs that are common in the specification of configuration settings.
The Domains and Mappings of a MappingSet may be defined either extensionally or intensionally. If a extensional definition of the Domain is given, then both extensional and intensional definitions of Mappings are allowed. On the other hand, an intensional definition of the Domain may only be accompanied by intensional definitions of Mappings. Generally, an extensional definition is a sequence of all the elements of Domain and Mappings in some specific order. Both row-wise and column-wise orderings may be used. It is even possible to combine row and column-wise orders for different components and Mappings. If the data type of a Domain component is of some integer or sampling data type, then its extensional definition might be given in the form of a collection of sequence definitions. In general, a sequence definition has an start element, a size and a step. For example, for an integer data type, a sequence starting at 5, with size 4 and step 2 describes the following list < 5, 7, 9, 11 >. For a TimeInstantSample data types, a sequence starting at “2013 − 05 − 0215 : 00 : 45.06”, with size 2 and step 30.42 describes the following sequence of t“y2p0e13T−im05eI−n0st2a1n5t(:20,03:05402.)6¡4“”2¿0.113. −Fo0r5P−oi0n2t125D:S0a0m:p2le0.d2a2t”a, types, starting element is fo type Point2D and step has to be given by two pairs (direction, resolution).
Spatio-temporal analysis is enabled through the intensional definition of Mappings and MappingSets. Mappings may be intensionally defined with functional, conditional and aggregate expressions.
Functional expression. A Mapping M with signature M(D): DT may be defined by a expression of the form
M(D) := e where e is a functional expression of data type DT that may include variables referencing components of D, mappings, operations, constants and castings.
Conditional Expression. A Mapping M with signature M(D): DT may be defined by a expression of the form M(D) := CASE b1 THEN e1
CASE b2 THEN e2 . . .
CASE bn THEN en [OTHERWISE en+1] where each bi is a functional expression that yields a value of Boolean type and each ei is a functional expression that yields a value of type DT. The semantics are the obvious ones.
Aggregate Expression. A Mapping M with signature M(D): DT may be defined by a expression of the form
M(D):= agge
OVER {P} where P is a domain relational calculus predicate and agge is an functional expression where variables bounded to MappingSet domains in P must be used as arguments of aggregate functions. Various aggregate functions are provided 1Notice that the start instant of the sequence is automatically adapted to match the underlying time representation for type TimeInstant(2, 3042) keyProperty property
GeoProperty
(a) Spatial Entities
EO
TO (c) Observed Entities property1 property2
C
(b) Spatial Coverages
EO relat.
TO relat.
(d) Observed Relationships simpleProperty TO component1
component2 simpleProperty EO complexProperty EO
multiValued TO (e) Observed Properties by the system including both statistical and rank functions. MappingSet domains may also be intensionally defined. Intensional Domain. Let e be a functional expression that yields a value s of either Interval(T) or Geometry(S) data type, whose base type T, S is either some integer type or some sample type. Then, SAMPLING(e) yields all the elements of type T or S contained in s. Based on this, the domain D of a MappingSet M may be defined by an expression of the form
{(e1, e2, . . . , en)|P } where P is a domain relational calculus predicate and each ei is either a functional expression or an expression of the form SAM P LIN G(e), where e is also a functional expression. Expressions e and ei may include variable names bounded to MappingSet domain components in P. Given that nested structures are not allowed in the model, if an expression SAM P LIN G(e) is used then the result relation has to be unnested. 4.
MODELING OBSERVATION DATA WAREHOUSES
The data model described in this section captures observation data semantics and integrates them with spatial entities and coverages. An E-R extension is proposed in Subsection 4.1 to model observation metadata. The translation of such a conceptual model to the MappingSet based logical model is explained in Subsection 4.2. 4.1
Contrary to conventional metadata that is recorded at the level of entity and property types, some observation metadata has to be recorded at the level of entity and property instances, i.e., combined with the data itself. This is the case for example of observation time instants and observation processes.
An extension of the E-R model is next proposed to incorporate spatial and observation data semantics in conceptual models. Spatial Entity types are represented in diagrams as conventional entities (see Figure 2(a)). Spatial Coverage Types are represented as entities tagged with the symbol name comfort are also tagged with the same TO and EO symbols, as it is shown in Figure 2(e) for simple, complex and multivalued properties.
To illustrate the use of the above notation the E-R diagram of a reduced running application example is given in Figure 3. Spatial Entity Type ICES records fishing zones defined by the International Council for the Exploration of the Sea (ICES). Spatial Coverage SST records Sea Surface Temperature at each location of the sea, daily produced by the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor installed in the Terra and Aqua NASA satellites. Entity type Vessel records data of fishing vessels, including an identifier (vesId ) and its engine power. Vessels incorporate CTD sensors that enable obtaining triples of water conductivity, water temperature and depth. Every time a ctd observation is performed a Non-Physical Process is executed that computes the difference with the value given by MODIS and provides it as a derived property difTemp. Vessels also incorporate GPS sensors from which locations are obtained every 30 seconds. Entity type Species records data of fishing species, including an identifier specId, species name and an interval of temperature values where the fish feels comfortable (property comfort ). The derived property comfGeo records the geometry of the area of the sea where comfortable temperatures for the fish are located. This property is obtained by a Non-Physical Process from the SST data. Property load of relationship catches records the values measured by the vessel bascule for each species. The authorized fishing capacity of a vessel is given by two parameters. The Fishing Effort gives a measure of the number of days weighted by the vessel engine power that the vessel may stay in each zone. Relationship Effort records both the initial quota and the available one (property stock ). Available Fishing Effort stock is obtained by a Non-Physical Procress using the quota and the vessels GPS information. The Fishing Capacity gives the kilograms of each species that the vessel may get from each zone. Again both quota is recorded and stock is computed by a Non-Physical Process.
The translation of the above model to the MappingSet logical model of the framework is explained in the following subsection. 4.2
To support the implementation of the conceptual model of the previous section, observation metadata has to be added to the frameworks catalog. Thus, the catalog contains metadata of the defined Mappings and MappingSets and metadata related to the various observation processes, including observation properties and features of interest. The E-R diagram of such catalog structures is given in Figure 4.
Entity types MappingSet, DomComponent and Mapping record general metadata of the MappingSets. Entity type FOI records metadata of Features of Interest, and it references the MappingSet that records its data. FOIs that are fully generated by observation processes are registered in ObsFOI. The remainder FOIs, i.e., those that combine observed with non observed properties are represented by entity type NonObsFOI. Each observed property of such a FOI is represented by a weak entity of type ObsProperty, which references the MappingSet that records its data. Finally, ProcessType records metadata of the various types of observation processes registered in the framework. Metadata of each specific instance of each process type is recorded in weak entity type Process. Notice the difference between the process type “Vessel Bascule” that obtains values of load property of relationships catches and the specific bascule installed in each vessel that must be referenced from each observation.
The rules that enable the transformation of the conceptual model of the previous section to MappingSets are now given next. Each Entity Type, either Spatial or not, generates a relevant MappingSet, whose domain is defined by key properties and whose Mappings are defined by the remainder properties. See for example Entity Types Vessel, Species and ICES in Figure 3 and relevant MappingSets in Figure 5. Each Spatial Coverage generates a MappingSet, whose domain has just one component of some Point2DSample type and whose Mappings are generated from coverage properties. Each Relationship Type with cardinalities various to
DOMAIN vesId: CString MAPPINGS power(vesId:CString):Numeric(6,2)
DOMAIN obsTime: TimeIntantSample(0, 30), vesId: CString MAPPINGS loc(phenTime: TimeIntantSample(0, 30), vesId:CString):Point2D process(obsTime: TimeIntantSample(0, 30), vesId:CString):CString
DOMAIN obsTime: TimeIntant(0), vesId: CString MAPPINGS cond(obsTime: TimeIntantSample(0, 30),
vesId:CString):Numeric(4,1) condUOM(obsTime: TimeIntantSample(0, 30),
vesId:CString):CString temp(obsTime: TimeIntantSample(0, 30),
vesId:CString):Numeric(5,2) tempUOM(obsTime: TimeIntantSample(0, 30),
vesId:CString):CString depth(obsTime: TimeIntantSample(0, 30),
vesId:CString):Numeric(5,2) depthUOM(obsTime: TimeIntantSample(0, 30),
vesId:CString):CString process(obsTime: TimeIntantSample(0, 30),
vesId:CString):CString geo(zoneId:CString):Polygon(Point2D(9,2)) various generates a MappingSet whose domain is defined from the key properties of the participating Entity Types. Properties of those Relationship Types generate Mappings in such a MappingSet. See for an example Relationship Types capacity and effort in Figure 3 and MappingSets Capacity and Effort in Figure 5. If an Entity, Coverage or Relationship Type is tagged with the symbol TO , then a component named obsTime of some TimeInstantSample(D,R) data type is added to the MappingSet Domain to enable the recording of observation time.2 Besides, a Mapping named process is also added to obtain the id of the process used to produce the observation. See for example Spatial Coverage Type SST in Figure 3 and relevant MappingSet SST in
If a simple or complex property is tagged with symbol TO then such property is not added as a Mapping to the relevant MappingSet. Instead, a separate MappingSet is created for the property whose domain has components to reference the key of its Entity Type (FOI of the relevant observation) and has a component named obsTime of some TimeInstantSample(D,R) type to record observation time. The property itself is added as a Mapping to the MappingSet and an additional Mapping named process is added to record the id of the process that generates the observation. An example is loc property of Entity Type Vessel in Figure 3 and relevant Vessel loc MappingSet in Figure 5. If symbol EO is used instead then the transformation is exactly the same except for the fact that Domain component obsTime is of some TimeInstant(D) type. For an example see ctd property of Vessel Entity Type in Figure 3 and relevant Vessel ctd MappingSet in Figure 5. In any of the above cases an entity of type ObsProperty is added to the catalog, with appropriate references to its MappingSet, ProcessType and NonObsFOI.
Once the MappingSets are created and the required metadata are added to the catalog, the insertion of observation data may be started. ETL tasks are continuously executed to maintain the data warehouse updated with latest observation data, using extensional MappingSet definitions. Each observation is appended to the appropriate MappingSet with its observation time and reference to its process and FOI.
DEFINITION OF SPATIO-TEMPORAL ANALYTICAL PROCESSES
The capabilities provided by the framework for the intensional definition of MappingSets enable the specification of spatio-temporal analytical processes. These capabilities are now illustrated with some examples.
Example 1. Define a Non-Physical Process that obtains a derived observed property that computes the difference between the temperature measured by the CTD and the sea surface temperature produced for the same location by 2Currently we restrict to phenomenon time semantics, however, it can be extended with result time and other required metadata.
MODIS (see difTemp derived property of Vessel in Figure 3).
MAPPINGSET Vessel difTem DOMAIN
{(obsTime, vesId) | Vessel ctd(obsTime, vesId)} MAPPINGS difTem(obsTime, vesId):=
SST.temp(Vessel loc.loc(obsTime, vesId), obsTime) − Vessel ctd.temp(obsTime, vesId) difTemUOM(obsTime, vesID):=
Vessel ctd.tempUOM(obsTime, vesId) process(obsTime, vesID):= “difTemProcess”
In the expression above it is noticed that automatic castings of spatial and temporal types are performed during the evaluations of Mappings Vessel loc.loc and SST.temp.
Example 2. Define a Non-Physical Process that detects when a vessel leaves an ICES zone to enter a new one (see enters derived relationship in Figure 3).
ICESFromLoc(loc):= singleton(zone)
OVER {ICES(zone) ∧ within(loc, ICES.geo(zone))} MAPPINGSET enters DOMAIN {(vesId, ICESFromLoc(Vessel loc.loc(obsTime, vesId)), obsTime) | Vessel loc(obsTime, vesId) ∧ ICESFromLoc(Vessel loc.loc(obsTime, vesId)) <>
ICESFromLoc(Vessel loc.loc(predecessor(obsTime), vesId))} MAPPINGS
process(vesId, zoneId, obsTime):= “entersProcess”
In the above expression, Mapping within(g1, g2) yields true if geometry g1 is within geometry g2. Aggregate function singleton(S) yields the element contained in the unitary set S. Finally, Mapping predecessor (ts) yields the time sample that precedes time sample ts in its data type.
Example 3. Define a Non-Physical Process that produces a measure of the remaining fishing effort for each vessel and ICES zone for each of the preceding 60 days. Consumed fishing effort is obtained from the temporal evolution of vessel location data and ICES zone geometries (see derived property stock of relationship type Effort in Figure 3).
ICESFromLoc(loc):= singleton(zone)
OVER {ICES(zone) ∧ within(loc, ICES.geo(zone))} consumed effort(vesId, zoneId, obsTime) := ((count(obsTime2)*30)/86400)*Vessel.power(vesId) OVER {Vessel loc(obsTime2, vesId2) ∧ obsTime2 < obsTime ∧ vesId2 = vesId ∧ ICESFromLoc(Vessel loc.loc(obsTime2, vesId2)) = zoneId } MAPPINGSET Effort stock DOMAIN {(vesId, zoneId,
SAMPLING([cast(difTime(now(), 60 Days) AS Date), cast(now() AS Date)])) | Effort(vesId, zoneId)} MAPPINGS stock(vesId, zoneId, obsTime):=
Effort.quota(vesId, zoneId) − consumed effort(vesId, zoneId, obsTime) stockUOM (vesId, zoneId, obsTime):=
Effort.quotaUOM(vesId, zoneId) process(vesId, zoneId, obsTime):= “EffortStockProcess” In the above expression, Mapping now () yields the current system time instant. Mapping difTime(t, i) subtracts time interval i from time instant t.
Example 4. Define a Non-Physical Process that obtains the evolution with respect to to time during the last 7 days of the geometry of the comfort zone for each species. Comfort zone is obtained from the temperature interval defined for each species and the sea surface temperature generated by MODIS (see derived property comfGeo of entity type species in Figure 3).
MAPPINGSET Species comfGeo DOMAIN {(speId,
SAMPLING([cast(difTime(now(), 7 Days) AS Date), cast(now() AS Date)])) | Species(speId)} MAPPINGS comfGeo(speId, obsTime):= vectorize(loc) OVER { SST(loc, obsTime2) ∧ obsTime = obsTime2 ∧ within(SST.temp(loc, obsTime2), Species.comfort(speId))} process(vesId, zoneId, obsTime):= “ComfortZoneProcess” In the above expression, aggregate function vectorize(loc) obtains the vector geometry that surrounds the set of sample locations loc. Mapping within(e, i) yields true if element e is within interval i. 6. CONCLUSIONS AND FURTHER WORK
A data model and data management framework has been proposed spatio-temporal analysis of data in data warehouses of spatial observation data. The approach consists of an E-R extension for observation data to be used at a conceptual level and a new logical level model that combines logical and functional paradigms. The advantages of the approach can be summarized as follows: • General purpose observation data and metadata coexist with application specific Spatial Entities and Coverages, enabling efficient analysis over the whole set. • Few primitive Mappings combined with general purpose logical and functional expressions enable the integrated management of any kind of spatial and spatio-temporal data. Besides, both data and analytical processing is unified under the well known mathematical concept of function. • Parametric temporal and spatial types enable the user to have control over the precision and resolution of underlying time and space representations. • Specific constructs for the specification of sampled and nonsampled domain components together with the absence of nested structures simplifies efficient implementation.
Further work is mainly related to efficient implementation structures and algorithms and the extension of the framework to deal with continuous queries on sensor data.
This work has been partially supported by the Spanish Ministry of Science and Innovation (TIN2010-21246-C02-02).