Ethical Decision

U o b

S G

a

A A

K E S I N N

I

h ( a a G l d r 2

o t w c ( p f d

h 0

International Journal of Applied Earth Observation and Geoinformation 37 (2015) 133–141

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and Geoinformation

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / j a g

sing ontological inference and hierarchical matchmaking to vercome semantic heterogeneity in remote sensing-based iodiversity monitoring

imon Nieland ∗, Birgit Kleinschmit, Michael Förster eoinformation in Environmental Planning Lab, Technische Universität Berlin, Straße des 17. Juni 145, 10623 Berlin, Germany

r t i c l e i n f o

rticle history: vailable online 16 October 2014

eywords: arth observation emantic reasoning nteroperability

a b s t r a c t

Ontology-based applications hold promise in improving spatial data interoperability. In this work we use remote sensing-based biodiversity information and apply semantic formalisation and ontological infer- ence to show improvements in data interoperability/comparability. The proposed methodology includes an observation-based, “bottom-up” engineering approach for remote sensing applications and gives a practical example of semantic mediation of geospatial products. We apply the methodology to three different nomenclatures used for remote sensing-based classification of two heathland nature conserva-

ature conservation atura 2000

tion areas in Belgium and Germany. We analysed sensor nomenclatures with respect to their semantic formalisation and their bio-geographical differences. The results indicate that a hierarchical and trans- parent nomenclature is far more important for transferability than the sensor or study area. The inclusion of additional information, not necessarily belonging to a vegetation class description, is a key factor for the future success of using semantics for interoperability in remote sensing.

ntroduction

Semantic interoperability’s importance for data harmonisation as often been discussed in geographic information integration Janowicz, 2012; Hess et al., 2007; Kavouras et al., 2005; Rodriguez nd Egenhofer, 2003; Visser et al., 2002; Lutz et al., 2009). In ddition to syntactic interoperability, thoroughly defined by Open eospatial Consortium (OGC) standards, heterogeneities of under-

ying semantics represent an unsolved barrier for data integration, ata discovery and knowledge sharing – especially in a variety of emote sensing-based applications (Arvor et al., 2013; Blaschke, 010).

Although remote sensing products and classification procedures ften implicitly use semantics for developing rule-sets or indica- ors there is a lack of structured, computer-readable formalisation ithin the given classification approaches. Remote sensing-based

lassification conceptualises a real world object or phenomenon entity) and produces its mapping (symbol). When trying to com-

are classification results, naming conflicts (different descriptions or the same conceptualisation or one ambiguous description for ifferent conceptualisations) and conceptual conflicts (different

∗ Corresponding author. Tel.: +49 30 314 72601. E-mail address: simon.nieland@tu-berlin.de (S. Nieland).

ttp://dx.doi.org/10.1016/j.jag.2014.09.018 303-2434/© 2014 Elsevier B.V. All rights reserved.

© 2014 Elsevier B.V. All rights reserved.

conceptualisations for the same mapping) occur, leading to seman- tic heterogeneity (Kuhn, 2005).

The described heterogeneities hamper the examination of remote sensing output information which is especially problem- atic when it is required for multi-national legal processes as is the case for the EU Habitats Directive (Council Directive92/43/EEC, 1992) (HabDir) and the Water Framework Directive (Council Directive2000/60/EEC, 2000). A comparable thematic depth is needed for the subsequent decision making process. Due to the semantic diversity of remote sensing results, such products are either not considered useful or in the early stages of develop- ment (Manakos and Hellas, 2013). Several approaches (Lutz et al., 2009; Visser et al., 2002; Rodriguez and Egenhofer, 2003; Durbha et al., 2009; Mena and Illarramendi, 2000; Kavouras et al., 2005; Schwering and Raubal, 2005) were applied to achieve seman- tic interoperability of spatial data by using ontologies based on the Resource Description Framework (RDF) or Web Ontology Language (OWL). What these approaches all have in common is the matchmaking process; a technique used to find equiv- alent information that fit to the particular subject of interest. Matchmaking between spatial datasets can be generated by using similarity values. Often based on dictionaries, thesauri, other

RDF/OWL-based data structures (Hess et al., 2007; Rodriguez and Egenhofer, 2003; Kavouras et al., 2005; Fonseca et al., 2006) or geospatial concepts and their geometrical models (Schwering and Raubal, 2005), similarity values indicate the degree of

1 rth Ob

c b c l c t m a s 2

o s p r t o c o s e a (

t C t S b c

r fi s d m t i s c h s

R b i r o r

•

•

•

A f

modules (Frick, 2006). Therefore, image data has to be structured and relative values (NDVI, texture, spatial arrangement) have to be derived.

34 S. Nieland et al. / International Journal of Applied Ea

orrespondence between entities. Additionally, matchmaking can e achieved by reasoning about the formalised concepts of one spe- ific domain ontology (Decker et al., 1998) and its aligned upper evel ontology (Cruz and Sunna, 2008). Multi-ontology systems in ombination with query rewriting techniques have also been used o generate matchmaking (Mena and Illarramendi, 2000). Further-

ore, a “hybrid ontology” approach where a shared vocabulary is pplied for the formalisation of concepts and inter-ontological rea- oning was used in several systems (Lutz et al., 2009; Visser et al., 002; Durbha et al., 2009).

Several studies have proposed using observations for geo- ntologies (Janowicz, 2012; Couclelis, 2010; Frank, 2003). In o-called observation-driven engineering approaches ontological rimitives (classes in the ontology that are not conceptualised) epresent elementary concepts that can be derived from observa- ions. Therefore, included primitives are restricted to observations r derived by aggregation of observed phenomena. Conceptualised lasses in the developed ontology can be assigned to upper level ntologies to foster a broader interoperability. Starting with the emantic descriptions of observations in a bottom-up ontology ngineering approach preserves the benefit of semantic diversity nd local conceptualisations without giving up interoperability Janowicz, 2012).

Recent approaches of remote sensing classification are bound o semantic web standards proposed by the World Wide Web onsortium (W3C) and therefore allow the utilisation of seman- ic reasoning in the classification process (Andres et al., 2012; di ciascio et al., 2013; Belgiu et al., 2014). These approaches are not roadly used in the remote sensing community, despite recent dis- ussions touting their benefits (Arvor et al., 2013).

Heterogeneities in classes based on remote sensing analysis are esulting from the fact that classification procedures are speci- ed through electromagnetic signals, whereas indicators of field urveys and nomenclatures reflect the composition of parameters efined in the particular area of research or cognitive interest. In any cases remote sensing classification techniques are adapted

o classes which are optimised for manual interpretation of aerial magery or fieldwork through the aggregation of the primitive clas- ification results to the target classes. Consequently, these primitive lassifications conceal information because users or customers only ave the generalised mapping without its underlying conceptuali- ation.

A better conceptualisation of remote sensing outputs in DF/OWL-based structured metadata would not only lead to a etter re-usability and exchangeability, but would additionally

mprove spatial information retrieval (Arvor et al., 2013). Inferring elations between data requirements or products and existing data r nomenclatures is a benefit that is already broadly used in other esearch areas (Bard and Rhee, 2004).

The main objectives of this work are to

propose an observation-based, bottom-up ontology engineering approach for remote sensing applications, which will be used for solving semantic heterogeneity problems in remote sensing classification results by taking into account ontology-based auto- matic reasoning in combination with matchmaking processes based on generalisation, give a practical example of semantic mediation of geospatial data in the field of remote sensing-based biodiversity monitoring and analyse certain criteria of selected areas (similarity of sensors, number of classes, similarity of geographical region) in regard to their influence on used indicators and subsequently their effect

on data interoperability.

big future challenge of remote sensing research is to trans- orm local or regional classification outputs into interoperable,

servation and Geoinformation 37 (2015) 133–141

comparable information. Since there are existing interoperability approaches in other research domains, we contribute to the exist- ing research by addressing the need for interoperability with a novel semantic approach that is based on ontological subsumption.

Methodology

This section proposes a bottom-up, observation-based ontol- ogy engineering approach and shows how it can be used for data interoperability in a prototype application.

Study sites and existing habitat data

We analyse remote sensing classification results of Natura 2000 heathland areas and corresponding nomenclatures for this study. The Natura 2000 sites are heathland and grassland habitats in Flan- ders (Belgium) and Brandenburg (Germany).

Kalmthoutse Heide1 (abbr. FL), located in northern Belgium is mainly covered with dry and wet heathland, inland dunes, water bodies and forests (Chan et al., 2012). The 6 broader habitat classes at level 1 are gradually arranged into subcategories that reflect the definitions of the habitat structure as well as the structures and functions that are crucial for the assessment of habitat quality (Thoonen et al., 2013).

The second study site, Döberitzer Heide,2 located in eastern Germany is characterised by heathland and grassland vegetation, humid meadows and woods on predominantly dry and sandy soils.

For the Döberitzer Heide, two classification hierarchies are avail- able (see Table 1). The nomenclature for multi-temporal, high resolution (HR) and hyper-spectral analysis (abbr. BB-HyMap) extends the federal nomenclature towards specific plant commu- nities.

These plant communities are used as indicators for the evalua- tion of habitat conservation status (Schuster et al., 2015). Additional class attributes were already included in the federal nomenclature. Since only parts of the developed and well-formalised BB-HyMaP nomenclature have been classified within this study, a synthetic dataset was created to acquire more significant information about the quality of the semantic transformation process. It uses the extent and pixel-size of BB-VHR and includes one band with cor- responding class values. The multi-temporal classification was performed by Schuster et al. (2015) with 21 RapidEye scenes cover- ing dates from March to October (Schuster et al., 2015). The classes were created using federal habitat descriptions and field-based mapping in Brandenburg.

Habitat classification with VHR imagery was realised using a knowledge-based classification approach. The development of the classification procedure can be divided in the following steps. Ini- tially, suitable indicators are selected, which are limited to those that can be derived from very high spatial resolution (VHR) remote sensing imagery. In the next step, these indicators can now be used to develop a hierarchical classification schema. To validate the sep- arability of the determined classes by using a discriminant analysis, each class has to be associated with representative ground truth areas. The developed hierarchical schema is the basis for the multi- level, pixel-based classification procedure. To be able to analyse the variability in the imagery statistically, the classification procedure uses a hybrid system of supervised and unsupervised classification

1 http://natura2000.eea.europa.eu/Natura2000/SDF.aspx?site=BE2100015 2 http://natura2000.eea.europa.eu/Natura2000/SDF.aspx?site=DE3444303

S. Nieland et al. / International Journal of Applied Earth Observation and Geoinformation 37 (2015) 133–141 135

Table 1 Description of classification hierarchies and datasets used for ontological conceptualisation.

Nomenclature Hierarchy levels

Number of classes (total)

Number of classes per hierarchy level (top to most detailed)

Nomenclature created by (date of publication)

Developed for sensor (spatial resolution (m)/number of bands)

Dataset/classification produced by

Brandenburg (abbr. BB-HyMap)

4 61 6 Förster et al. (2012) RapidEye part of this study

17 Schuster et al. (2015)

(6,5×6.5/5)

18 HyMap 20 (5×5/126)

Brandenburg (abbr. BB-VHR)

5 33 2 Frick and Weyer (2005)

QuickBird Frick (2006)

2 (2.4×2.4/4) 9 20

Flanders (abbr. FL) 4 51 6 Thoonen et al. (2013)

AHS-160 -Airborne Thoonen et al. (2013)

H

b L d r t g t L

i b s a o s i i t t a s

“ c

D

r o t C d

c i b ( t T d

t

formalising classification keys with shared vocabulary in DL. Fig. 1 shows a schematic representation of subsumption (1), equivalence (2) and disjointness reasoning (3) and illustrates the construction of a combined concept hierarchy (ab).

Table 2 Concept constructors of the developed ontology.

(A|) (A|) (atomic concept) top* � (universal concept) bottom* ⊥ (bottom concept) (not E) ¬ E (complement)

12 18 15

ybrid ontology model

The basis of this work is a hybrid ontology model, which com- ines single and multiple ontology approaches (Wache et al., 2001; utz and Klien, 2006; Lutz et al., 2009; Visser et al., 2002). The fun- amental concept of this approach is to describe the classification esults of different regions, each with its own ontology. In contrast o multiple ontology approaches, the concepts of the “local” ontolo- ies are based on primitives and properties of a shared vocabulary hat are stored in upper level domain ontologies (Guarino, 1998; utz et al., 2009).

The main advantages of this methodology are to keep the flex- bility of a multi-ontological approach and preserve comparability y using a shared vocabulary. An important characteristic of the hared vocabulary is its independence from existing classification pproaches and sensors. Compared with the observation-driven ntology engineering approach proposed by Janowicz (2012), the hared vocabulary’s classes describe an observed phenomenon nstead of merely conceptualising possible spectral signatures or ndices of sensors. In this work the shared vocabulary stores quali- ative attributes of the remote sensing classification process, while he local ontologies formalise the classification outputs. Semantic nnotation of classification products are realised in the prototype oftware using a PostGIS database back-end.

Furthermore we describe the shared vocabulary ontology as a domain ontology”, while ontologies that formalise specific con- epts of a respective region are referred to as “local ontologies”.

escription logics and reasoning

In this work Description Logic (DL) is used for formal knowledge epresentation, providing the formal basis for the web ontol- gy language (OWL) (Schmidt-Schauß and Smolka, 1991). For he ontological modelling the Attributive Concept Language with omplements (ALC) is used, which includes so-called elementary escriptions and concept descriptions (see Table 2).

Basic syntactic elementary descriptions represent atomic con- epts (unary predicates), atomic rules (binary predicates) and ndividuals (constants), whereas complex concept descriptions are uilt from these elementary descriptions with concept constructors union, intersection, complement, universal restrictions, existen- ial quantifications) (Baader and Werner, 2003; Lutz et al., 2009).

he universal concept includes the set of all individuals in the omain. The bottom concept defines an empty set.

A DL knowledge base can be divided into a TBox containing he terminology (description of relations between concepts and

Hyperspectral (2.5×2.5/63)

roles) and an ABox containing assertions about named individuals and concepts. Since this work focuses on relations between com- plex concept descriptions, further developments refer to the TBox language features.

Implicit relations between complex concepts and atomic concepts can be inferred automatically by performing logical rea- soning. The first task is to determine whether a newly defined concept makes sense or is inconsistent. The second reasoning task is to decide whether a complex concept A is subsumed by a concept B. Subsumption in a TBox T is identifiable in every model of T if the set of concepts denoted by A is a subset of the set denoted by B (Donini, 2003) (see Fig. 1(1)).

If AT ⊆ BT then A � Bt (1) Equivalence between concept A and B with respect to T can be

proven if the set of concepts denoted by A equates the set denoted by B for every model I of T (Donini, 2003) (see Fig. 1(2)).

If AT ⊆ BT then A ≡ Bt (2) Disjointness between concept A and B with respect to T can be

proven if the union between the set of concepts denoted by A is null for every model I of T (Donini, 2003) (see Fig. 1(3)).

AT ∩ BT = ∅ (3) Since reasoning over big ontologies is often time-consuming

we use the reasoner Hermit, which is based on a hypertableau and hyperresolution calculi providing fast and efficient reasoning capabilities and making this task technically feasible (Motik et al., 2009).

Hence, it is possible using the reasoner to build up one remote sensing classification hierarchy of two regions (a and b) by

(and E) E1 ∩ E2 (intersection) (or E) E1 ∪ E2 (union) (only E) ∀ R.E (universal restriction) (some E) ∃ R.E (existential quantification)

136 S. Nieland et al. / International Journal of Applied Earth Ob

F s

t n g

H

s m f a

i s i a s t s 2

A a

cross-checked with the developers of the regarded classification hierarchies. Therefore, the shared vocabulary represents a hierar- chy of indicators, which are detectable by using remote sensing

ig. 1. Schematic representation of subsumption, equivalence and disjointness rea- oning.

However, the constructed hierarchy’s structure needs to be aken into account as it is not an undirected graph in which every ode only has one super-element, but is instead a directed acyclic raph, whose elements can have multiple super-elements.

ierarchical matchmaking

Since the main focus of this work is to find relations in remote ensing outputs classified in different regions and with differing ethodologies, we developed an algorithm which is able to per-

orm equality tests between the concepts of the two regions on scending levels of the inferred hierarchy.

Since remote sensing-based nomenclatures are often described n a hierarchical way, matchmaking based on “ontological sub- umption” (furthermore referred to as hierarchical matchmaking) s a practical technique for achieving comparability. We decided gainst using a similarity model because of the hierarchical tructure of the remote sensing outputs, the very detailed descrip- ions and the minimal differences between the classes in remote ensing products (Cruz and Sunna, 2008; Rodriguez and Egenhofer, 003).

lgorithm 1. Finding matches between concepts C of region a nd b in iteratively extending generalisation levels.

Data: A finite set of concepts of region A (Ca ), which includes Concepts (Ca1 , Ca2 , Ca3 ,.., Can ) and region B (Cb ), which includes the concepts(Cb1 , Cb2 , Cb3 ,. . ., Cbn )

Result: A set of matching pairs

def getMatchingPairs (Ca ):

for i in Ca : if getequivalentClasses(Cai ) != null:

matchingPairs[Cai ] = getequivalentClasses(Cai )

servation and Geoinformation 37 (2015) 133–141

else: numberOfSuperClassLevels = getNumberOfSuperClassLevels(Cai ) for z in range 0, numberOfSuperClassLevels:

if testLevelEquivalence(Cai , z) != null matchingPairs[Cai ] = testLevelEquivalence(Cai , z) break

return matchingPairs

Inputs to the algorithm (see Algorithm 1) are sets of con- cepts of region A (Ca) and region B (Cb) (see Fig. 2). Region A and B include complex concept descriptions (Ca1, Ca2, Ca3,. . ., Can/ Cb1, Cb2, Cb3,. . . Cbn) which formalise the region’s classifi- cation results by using elementary descriptions of the shared vocabulary. The function getequivalentClasses(Cai) returns equiv- alent classes of Cai that are situated in region B. If any direct equivalences are given back from the reasoner the destination con- cepts can directly be assigned to the origin concept (Cai). If not, the number of hierarchy levels above concept Cai have to be identi- fied (getNumberofSuperClassLevels(Cai))(that means the number of potential upscaling processes until reaching the universal con- cept). The function testLevelEquivalence() returns concepts of the hierarchy level z that are equivalent to the origin concept Cai and are situated in region B. Therefore searching for the correct stage in the classification schema is iteratively extended to the next gen- eral level until a match can be found. Thus, the return value of the algorithm is a set of pairs including the origin concepts (Cai) and corresponding destination concepts (Cbx). Since the algorithm is able to generate equivalence checks between all concepts included in the local ontologies, all possible transformations between the nomenclatures have been realised (see Table 4).

Formalisation of classification outputs

Well-formalised semantics of classification keys are the funda- mental basis of this work.

The Dolce Ultra-Light (DUL)3 ontology provides the basis for the Semantic Sensor Network Ontology (SSN) 4 (Compton et al., 2012), developed by the W3C’s Semantic Network Incubator Group, and the Stimulus-Sensor-Observation ontology design pattern (Janowicz and Compton, in press). We chose DUL as top-level ontology mainly to provide broader interoperability to applications realised with SSN and respectively DUL, as this is a recommended framework to model monitoring services based on sensor observa- tions.

Furthermore, we formalise certain heathland and grassland habitats protected under the HabDir. The basic concepts, which are stored in a shared vocabulary and used to describe important indicators for these habitats, have been adopted from the devel- oped classification schemes. From an ecological point of view, the so-called “indication” is the most suitable principle to generate sci- entifically correct and comprehensive examination of objectives regarding nature conservation (Frick, 2006). The derived indicators should be particularly sensitive to changes of relevant environ- mental factors. Indicators are parameters which can be measured or derived to determine and evaluate a complex ecological phenomenon that cannot be described directly (Niemeijer, 2002). To develop correct and consistent indicators, the indicators were

3 http://www.loa.istc.cnr.it/ontologies/DUL.owl 4 http://purl.oclc.org/NET/ssnx/ssn

S. Nieland et al. / International Journal of Applied Earth Observation and Geoinformation 37 (2015) 133–141 137

on a a

c t r t

r i s i i u u u i k b “ v s

c t d

w b

B “ i s

V

f

I

Fig. 2. Finding matches between concepts C of regi

lassification techniques. This is stored in an OWL2 RDF/XML syn- ax based ontology, including implicit concept constructors such as elations between classes (e.g. disjointness), equality and charac- eristics of properties (e.g. symmetry).

Fig. 3 shows a fragment of the domain ontology describing a emote sensing-based biodiversity indicator entity. Concepts are llustrated as ellipses, black arrows denote inheritance relation- hips while grey arrows show concept constructors. The hierarchy ncludes examples of implicit relations between indicators. For nstance, domination of common rushes (Juncus effusus, L.) can be sed as a wetness indicator, and fallow land is an area not currently sed by humans. Indicators which are mutually exclusive (e.g. sed and unused or wet and dry), can be described as disjoint. By

nferring this simple ontology fragment there is already implicit nowledge revealed. For example, habitats which are dominated y common rushes cannot be dry, or habitats with the attribute Natural” cannot be used for agriculture. Currently, the shared ocabulary includes 120 concepts describing indicators for remote ensing-based Natura 2000 habitat monitoring.5

In the local ontologies, basic concepts for remote sensing-based lassification can now be described through the analysed indica- ors. For instance a concept “Gt” (dry grassland) of region A can be escribed as:

Gt: Gt ⊆ G Gt ≡ ((∃hasQuality.Dry) ∩ (∃hasQuality.SandDominated)) ∪ ((∃hasQuality.Dry)) ∩ (∃hasQuality.Grassdominated))

hile a concept natural permanent dry grassland in region B can e formalised as:

Gpnd: Gpnd ⊆ Gpn Gpnd ≡ ((∃hasQuality.Grazing) ∪ (∃hasQuality.Mowing) ∩ (∃hasQuality.Natural)) ∪ ((∃hasQuality.SemiNatural) ∪ (∃hasQuality.Dry)) ∩ (∃hasQuality.Grassdominated))

y inferring these classes “Gpnd” can be identified as a subclass of Gpn” (natural permanent grassland) and “Gt”. That means, “Gpnd” ncludes all attributes of “Gt,” therefore “Gt” is a potential up- caling/transfer target class.

alidation

In order to achieve an objective validation, a reference trans- ormation was created by taking into account ecological expertise

5 http://www.user.tu-berlin.de/simon.nieland/ontologies/MS MONINA nteroperabilityOntology.owl

nd b in iteratively extending generalisation levels.

from test sites and classification hierarchies. First, all the classes of the test sites were listed in a table and related indicators have been added by analysing detailed class descriptions. Secondly, indica- tors have been compared and used as a basis for finding equivalent classes in the other nomenclatures. For example, the class “dry grassland rich in cryptogams” (BB-VHR) can be described as dry grassland dominated by cryptogams (>30%) with potential parts of open soil. This class was developed for the assessment of class 2330 (inland dunes with open Corynephorus and Agrostis grasslands). By analysing the nomenclature of FL class Sfm, we can identify it as a corresponding category since it includes low, open vegetation on (partly) fixed sand dunes, predominantly covered by mosses and lichens (≥60%).

To achieve an objective validation of the semantic transforma- tion in terms of thematic accuracy, efficiency and practicability, the results were compared to the manual transformation for the two test sites. To evaluate the correctness of the results we adopted the concept of recall and precision, which is mainly used in infor- mation retrieval (Korfhage, 1997). In our case, precision describes the percentage of found valid equivalence relations on all relations, whereas recall illustrates the portion of found valid equivalence relations on all valid relations.

Results

Table 3 shows the result for the test case Flanders to BB-VHR nomenclature. Note that not all of the classes of the nomenclature are present in the dataset and that the classification procedures use classes at different levels of the hierarchy. Therefore the total num- ber of classes in Table 3 does not have to match one level of the hier- archy (see Table 1). The results shown in Table 3 indicate that there is just one equivalent class in the two regions; consequently, the- matic up-scaling was applied for all other classes in order to achieve comparable datasets in regard to their content. Therefore the gener- alisation level is expressed by the number of up-scaling procedures necessary to generate equivalent geospatial datasets of the two exemplary test cases. Table 4 gives an overview of necessary up- scaling processes in the performed semantic transformations.

Fig. 4 illustrates the classification result of study area FL (level 1), which was derived from an airborne hyperspectral scanner. Level 2 shows the same dataset visualised in the nomenclature of classification BB-VHRS, whereas level 3 demonstrates the nomen- clature used for classification of BB-HyMap imagery. Due to the necessary semantic upscaling processes the nomenclatures of level 2 and level 3 contain fewer classes, but the main categories (heath- land, arable land, grassland, open soil, forest and water) can be

differentiated in all levels. Moreover, it shows that nearly all classes and a high percentage of spatial coverage could be transferred.

Table 5 shows that for some classes more than one relation is correct. For example “Grassland Intensive” in region BB can either

138 S. Nieland et al. / International Journal of Applied Earth Observation and Geoinformation 37 (2015) 133–141

bulary

b t u u t d r e

Fig. 3. Ontology fragment of the shared voca

e transferred to Gp (Grassland permanent) or to Gt (Grassland emporary) in nomenclature FL. Both classes describe intensively sed agricultural grassland and the difference is only in degree of sage. Whereas Gp is permanent grassland used for hay and/or pas- ures, Gt is periodically used for crops. If there is no equivalence

etectable and the manual transformation also does not show any elation class, the discovered relation “no equivalence” is consid- red to be correct.

Fig. 4. Semantic transformation of datasets FL (Level

representing biodiversity indicator entities.

All results have precision values between 81.0 and 94.7 and respectively recall values from 72.1 to 87.5. Transfers from Flan- ders to Brandenburg nomenclatures seem to perform slightly better than transfers from BB-HyMap nomenclature. For BB-VHRS nomenclature, transfer results are diverse, with one very good

result from BB-VHR to BB-HypMap and one slightly worse result from BB-VHRS to FL. Generally, transfers from Flanders to Bran- denburg achieve better results than the reverse, whereas the

1), BB-VHR (Level 2) and BB-HyMap (Level 3).

S. Nieland et al. / International Journal of Applied Earth Ob

Table 3 Result semantic transformation from Flanders (FL) to Brandenburg (BB-VHR). Classes of the origin region Flanders are named in their original description. The first letter represents the broad habitat category (H - Heathland, A - Arable land, S - Sand dunes, G - Grassland, F- Forest, W - Water). The remaining letters indicate the attributes of the classified habitat (t - temporary, b- bare, c - crops/ or Calluna dominated (heath classes), o - other, w - wet, e - Erica dominated, d - dry, a- adult/ or agriculture (grassland), y - young, m - mixed age/ or molinia encroached (heath- land)/or maize (arable land), g - grass encroached, j - Juncus Effuses dominated, n- natural or seminatural, f - fixed).

Origin class (FL) Destination class (BB - VHR)

Number of thematic up-scaling processes

Acm SpeciesOfArableLand 2 Aco SpeciesOfArableLand 2 Cloud Fcpc SpeciesOfArableLand 2 Fcps Wood 2 Fdb Vegetated 2 Fdqz Wood 2 Gpap GrasslandIntensive 1 Gpar GrasslandIntensive 1 Gpj SpeciesOfWetGrassland 0 Gpnd SpeciesOfDryGrassland 1 Gt GrasslandIntensive 1 Hdca DrySandHeath 2 Hdcm DrySandHeath 2 Hdco DrySandHeath 2 Hdcy DrySandHeath 2 Hgmd SpeciesOfDryGrassland 1 Hgmw Grassland 2 Hwe Heathland 3 Sfgm SpeciesOfDryGrassland 2 Sfmc MossDominatedAreas 1 Sfmp MossDominatedAreas 1

i t

D

r h i b s

T N

T R

Unclass Wou Water 1 Wov Water 2

ntra-regional transformation has no clear tendencies in terms of he quality of the transformation process.

iscussion

In this study we presented an ontology engineering approach for emote sensing applications (in this case heathland and grassland

abitat classification), and introduced a hierarchical matchmak-

ng algorithm that found direct relations between nomenclatures ased on different remote sensing classification methodologies and ensors in different regions.

able 4 umber of necessary up-scaling processes per transformation.

Number of up-scaling processes

Number of classes BB(VHR)- Flanders

Number of BB-HyMap- Flanders(%)

Number of classe Flanders- BB-HyMap(%)

0 8 (38.1%) 13 (21.3%) 3 (12.5%) 1 8 (38.1%) 20 (32.8%) 14 (58.3%) 2 – 7 (11.5%) 6 (25%) 3 – 10 (16.4%) – Not transferable 5 (23.8%) 11 (18.0%) 1 (4.1%) Classes (total) 21 61 24

able 5 esults of semantic transformation.

Transformation type Total correct relations Discovered relati

FL → BB-HyMap 24 23 FL → BB-VHR 24 23 BB-HyMap → FL 61 57 BB-HyMap → BB-VHR 61 51 BB-VHR → FL 21 21 BB-VHR → BB-HyMap 21 19

servation and Geoinformation 37 (2015) 133–141 139

Generally, the presented approach of using ontology engi- neering and matchmaking for interoperability issues in remote sensing-based monitoring is feasible. The results show that in most cases the prototype application produces equivalent outcomes to manual work.

The outcomes indicate that the transfer between nomenclatures designed for very similar sensors (like FL and BB-HyMap) do not show better results than transferring between sensors with very different spectral and spatial resolutions (BB-VHR and FL, BB-VHR and BB-HyMap). Since we have only analysed three nomenclatures this trend cannot be clearly confirmed. Moreover, the quality of the formalisation as well as the geographical region and the effects of sensor similarity (in terms of spatial and spectral characteris- tics) seem to have an impact on interoperability. Nomenclatures that are used for similar or equal sensors, such as the ones in FL and BB-HyMap, produced good results in both directions, whereas nomenclatures of the same region (BB-VHR and BB-HyMap) had varying results. Also, results of the transfer between nomenclatures from different sensors and regions (BB-VHRS and FL) produced mostly (BB-VHR to FL) successful results. Therefore the influence of class formalisation seems to be higher than regional and sensor differences.

The described indicators and methodologies of image analysis for biodiversity monitoring reflect the challenge of evaluating and designating nature conservation areas predominantly from remote sensing information. Since at present most of the monitoring data is generated by manual field work, the results had to be evalu- ated by taking into account ecological health criteria. We see two main reasons for the gap between ecological and remote sensing perspectives. Problems in the matchmaking process occur if the integrated logic of the hierarchies, which are mainly ecological con- ceptualisations, do not match the logic of the underlying ontology.

Another problem occurs when the remote sensing-based indi- cator does not correspond with one from an ecological perspective, or a class does not contain all indicators which are necessary for an accurate evaluation. In this case it would be necessary to include indicators that cannot be assessed by remote sensing into the shared vocabulary. Addition of ecological indicators in the shared vocabulary could improve interoperability with more field-based nomenclatures and observations aiding in transferability. From a

remote sensing perspective, the definition of a classification hier- archy is often neglected and poorly documented. Before starting an actual classification, the remote sensing and ecology experts should be more detailed in class descriptions by giving more

s Number of classes Flanders- BB-VHR(%)

Number of classes BB-HyMap - BB-VHSR(%)

Number of classes BB(VHR)- BB(HyMap)

1 (4.1%) 13 (21.2%) 9 (42.3%) 8 (33.3%) 22 (36%) 8 (38.1%)

13 (54.1%) 11(18%) - 1 (4.1%) 5 (8.2%) – 1 (4.1%) 10 (16.3%) 4 (19%)

24 61 21

ons (%) Correct relations (%) Precision Recall

21 91.3 87.5 21 91.3 87.5 51 89.5 83.6 44 86.3 72.1 17 81.0 81.0 18 94.7 85.7

1 rth Ob

m c c n

o 2 u A p fl t a R r s c t s a a a

e S c s c c e

a o r a f m e

C

t c e c r a e a i f i o t d

n f r t

A

t

40 S. Nieland et al. / International Journal of Applied Ea

etadata and additional attributes to better substantiate class hoice. With these descriptive variables a better transferability ould be implemented without using the same nomenclature for ature conservation classes.

In comparison to other studies on ontology alignment and inter- perability issues (Visser et al., 2002; Rodriguez and Egenhofer, 003) the proposed matchmaking approach is focused on the nique requirements of remote sensing classification semantics. pproaches based on similarity measures, as used in many exam- les for semantic mediation (Cruz and Sunna, 2008), are more exible in regard to the structure of the ontology, but due to heir complexity (weighting of relations, symmetry/asymmetry) nd calibration difficulties lead to unacceptably high false matches. egarding the varying challenges and demands of remote sensing esearch, standardisation does not seem realistic. In fact, the diver- ity in methods and implied semantics reflects the variety of onceptualisations in different regions, which are based on cul- ure and social conventions (Vanden Borre et al., 2011). Therefore, emantic heterogeneity of habitat objects should not be understood s a “problem”, but rather a challenge on how to structure, describe nd store information in a way that allows inference on relations nd comparability.

The main problem in modelling ontological concepts is that xperts have to include as much implicit knowledge as possible. olving the implicit knowledge problem would then require fewer riteria for successful reasoning (Klien, 2008). Therefore it is neces- ary to request the expert to include certain attributes for classified oncepts. In order to achieve completeness in formalisation, criti- al attributes for differentiation have to be identified and included xplicitly.

Although the approach might be influenced by slight regional daptations (e.g. different plant communities of a more Atlantic r Continental influenced heathland), we show that a transfer of esults is possible without a time-consuming adaptation and re- pplication of the various classification algorithms. Since resources or remote sensing tasks are limited, the introduced hierarchical

atchmaking is an appropriate solution for the still-existing het- rogeneity of remote sensing products.

onclusion

The proposed methodology shows potential in terms of the ransferability of remote sensing output products. The storage of lassification meta-information in OWL/RDF ontologies based on xisting upper ontologies (DUL) leads to a better usability and omparability of remote sensing products. Using knowledge-based easoning for interoperability issues in remote sensing seems to be n obvious step and is increasingly supported by experts of differ- nt research fields (Arvor et al., 2013; Janowicz, 2012). This work is

first step in developing capable methodologies for this challeng- ng task. The results of the exemplary case showed, that a good ormalisation of classes is crucial for good interoperability. Giv- ng domain experts the chance to formalise the semantics of the bject of study in a computer readable way is a significant step owards the interoperability of remote sensing-based monitoring ata.

Furthermore, the developed ontology represents a basis for a umber of possible applications. Using the presented ontology

or cross-regional, semantic-based generalisation and information etrieval of remote sensing output data would be a further step owards comparable reporting in monitoring activities.

cknowledgements

The research leading to these results received funding from he Belgian Science Policy Office in the frame of the European

servation and Geoinformation 37 (2015) 133–141

Community’s 7th Framework Programme (FP7/2007-2013) under grant agreement on 263479 (MS.MONINA project). The authors greatly appreciate the support provided by Anett Frick (Luftbild Umwelt Planung Gmbh - LUP), Jeroen Vanden Borre, Toon Span- hove (Research Institute for Nature and Forest - INBO) and Birgen Haest (Flemish Institute for Technological Research - VITO).

References

Andres, S., Arvor, D., Pierkot, C., 2012. Towards an ontological approach for classify- ing remote sensing images. In: 2012 Eighth International Conference on Signal Image Technology and Internet Based Systems, pp. 825–832.

Arvor, D., Durieux, L., Andrés, S., Laporte, M.A., 2013. Advances in geographic object- based image analysis with ontologies: a review of main contributions and limitations from a remote sensing perspective. ISPRS J. Photogramm. Remote Sens. 82, 125–137.

Baader, F., Werner, N., 2003. Basic description logics. In: The Description Logic Hand- book. Theory, Implementation and Applications, pp. 47–95.

Bard, J.B.L., Rhee, S.Y., 2004. Ontologies in biology: design, applications and future challenges. Nat. Rev. Genet. 5, 213–222.

Belgiu, M., Tomljenovic, I., Lampoltshammer, T., Blaschke, T., Höfle, B., 2014. Ontology-based classification of building types detected from airborne laser scanning data. Remote Sens. 6, 1347–1366.

Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J. Pho- togramm. Remote Sens. 65, 2–16.

Chan, J.C.W., Beckers, P., Spanhove, T., Borre, J.V., 2012. An evaluation of ensem- ble classifiers for mapping Natura 2000 heathland in Belgium using spaceborne angular hyperspectral (CHRIS/Proba) imagery. Int. J. Appl. Earth Obs. Geoinform. 18, 13–22.

Compton, M., Barnaghi, P., Bermudez, L., García-Castro, R., Corcho, O., Cox, S., Gray- beal, J., Hauswirth, M., Henson, C., Herzog, A., Huang, V., Janowicz, K., Kelsey, W.D., Le Phuoc, D., Lefort, L., Leggieri, M., Neuhaus, H., Nikolov, A., Page, K., Pas- sant, A., Sheth, A., Taylor, K., 2012. The SSN ontology of the W3C semantic sensor network incubator group. Web Semant. 17, 25–32.

Couclelis, H., 2010. Ontologies of geographic information. Int. J. Geogr. Inform. Sci. 24, 1785–1809.

(Council Directive) 2000/60/EEC, C., 2000. Establishing a framework for Community action in the field of water policy.

(Council Directive) 92/43/EEC, C., 1992. The conservation of natural habitats and of wild fauna and flora.

Cruz, I.F., Sunna, W., 2008. Structural alignment methods with applications to geospatial ontologies. Trans. GIS 12, 683–711.

Decker, S., Erdmann, M., Fensel, D., Studer, R., 1998. Ontobroker: ontology based access to distributed and semi-structured information. In: DatabaseSemantics: Semantic Issues in Multimedia Systems., pp. 351–369.

Donini, F.M., 2003. Complexity of reasoning. In: Baader, F., Calvanese, D., Mcguinness, D.L., Nardi, D., Patel-Schneider, P. (Eds.), The Description Logic Handbook. The- ory, Implementation and Applications. Cambridge University Press, pp. 96–136.

Durbha, S., King, R., Shah, V., Younan, N., 2009. A framework for semantic rec- onciliation of disparate earth observation thematic data. Comput. Geosci. 35, 761–773.

Fonseca, F., Camara, G., Miguel Monteiro, A., 2006. A framework for measuring the interoperability of geo-ontologies. Spat. Cognit. Comput. 6, 309–331.

Förster, M., Schmidt, T., Schuster, C., Kleinschmit, B., 2012. Multi-temporal detection of grassland vegetation with RapidEye imagery and a spectral-temporal library. In: IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp. 4930–4933.

Frank, A., 2003. Ontology for spatio-temporal databases. In: Spatio-Temporal Databases: The CHOROCHRONOS Approach, pp. 9–77.

Frick, A., 2006. Beiträge höchstauflösender Satellitenfernerkundung zum Entwick- lung eines wissensbasierten Klassifikationsverfahrens und Anwendung in Brandenburg, TU Berlin (Ph.D. thesis).

Frick, A., Weyer, G., 2005. A knowledge-based approach to vegetation monitor- ing with Quickbird imagery. In: ISPRS Workshop 2005 – High Resolution Earth Imaging for Geospatial Information, Hannover.

Guarino, N., 1998. Formal ontology and information systems. In: Proceedings of FOIS98, pp. 3–15.

Hess, G., Iochpe, C., Castano, S., 2007. Geographic ontology matching with iG-match. Adv. Spat. Temporal Databases, 185–202.

Janowicz, K., 2012. Observation-driven geo-ontology engineering. Trans. GIS 16, 351–374.

Janowicz, K., Compton, M.,. The Stimulus-Sensor-Observation Ontology Design Pat- tern and its Integration into the Semantic Sensor Network Ontology.

Kavouras, M., Kokla, M., Tomai, E., 2005. Comparing categories among geographic ontologies. Comput. Geosci. 31, 145–154.

Klien, E.M., 2008. Semantic Annotation of Geographic Information. Universität Mün- ster, Ph. D. thesis.

Korfhage, R.R., 1997. Information Storage and Retrieval. John Wiley & Sons, New

York.

Kuhn, W., 2005. Geospatial semantics: why, of what, and how? J. Data Semant. III, 1–24.

Lutz, M., Klien, E., 2006. Ontology based retrieval of geographic information. Int. J. Geogr. Inform. Sci. 20, 233–260.

rth Ob

L

M

M

M

N

R

S

S

graphic information processing. Comput. Geosci. 28, 103–117.

S. Nieland et al. / International Journal of Applied Ea

utz, M., Sprado, J., Klien, E., Schubert, C., Christ, I., 2009. Overcoming semantic heterogeneity in spatial data infrastructures. Comput. Geosci. 35, 739–752.

anakos, I., Hellas, T., 2013. Remote sensing in Europe: status analysis and trends focusing on environment and agriculture. J. Aeronaut. Space Technol. 6, 1–5.

ena, E., Illarramendi, A., 2000. OBSERVER: an approach for query processing in global information systems based on interoperation across pre-existing ontolo- gies. Distrib. Parallel Databases 8, 223–271.

otik, B., Shearer, R., Horrocks, I., 2009. Hypertableau reasoning for description logics. J. Artif. Intell. Res. 36, 165–288.

iemeijer, D., 2002. Developing indicators for environmental policy: data-driven and theory-driven approaches examined by example. Environ. Sci. Policy 5, 91–103.

odriguez, A., Egenhofer, M.J., 2003. Determining semantic similarity among entity classes from different ontologies. IEEE Trans. Knowl. Data Eng. 15, 442–456.

chmidt-Schauß, M., Smolka, G., 1991. Attributive concept descriptions with com- plements. Artif. Intell. 48, 1–26.

chuster, C., Schmidt, T., Conrad, C., Kleinschmit, B., Förster, M., 2015. Grassland habitat mapping by intra-annual time series analysis Comparison of RapidEye and TerraSAR-X satellite data. Int. J. Appl. Earth Obs. Geoinform. 34, 25–34.

servation and Geoinformation 37 (2015) 133–141 141

Schwering, A., Raubal, M., 2005. Measuring semantic similarity between geospatial conceptual regions. In: GeoSpatial Semantics: First International Conference, pp. 90–106.

di Sciascio, C., Zanni-Merk, C., Wemmert, C., Marc-Zwecker, S., de Beuvron, F.d.B., 2013. Towards a semi-automatic semantic approach for satellite image analysis. Proc. Comput. Sci. 22, 1388–1397.

Thoonen, G., Spanhove, T., Vanden Borre, J., Scheunders, P., 2013. Classification of heathland vegetation in a hierarchical contextual framework. Int. J. Remote Sens. 34, 96–111.

Vanden Borre, J., Paelinckx, D., Mücher, C.a., Kooistra, L., Haest, B., De Blust, G., Schmidt, A.M., 2011. Integrating remote sensing in Natura 2000 habitat monitoring: prospects on the way forward. J. Nat. Conserv. 19, 116–125.

Visser, U., Stuckenschmidt, H., Schuster, G., Vögele, T., 2002. Ontologies for geo-

Wache, H., Voegele, T., Visser, U., Stuckenschmidt, H., Schuster, G., Neumann, H., Hübner, S., 2001. Ontology-based integration of information-a survey of existing approaches. In: Proceedings of the IJCAI-01 Workshop: Ontologies and Informa- tion Sharing, pp. 108–117.