Review
Modelling with knowledge: A review of emerging semantic approaches to
environmental modelling
Ferdinando Villa a,*, Ioannis N. Athanasiadis b, Andrea Emilio Rizzoli b
a Ecoinformatics Collaboratory, Gund Institute for Ecological Economics and Department of Plant Biology, University of Vermont, 617 Main Street, Burlington, VT, USA
b Istituto Dalle Molle di Studi sull’Intelligenza Artificiale, Lugano, Switzerland
a r t i c l e i n f o
Article history:
Received 24 August 2007
Received in revised form 2 September 2008
Accepted 6 September 2008
Available online 11 December 2008
Keywords:
Semantic modelling
Ontologies
Semantic annotation
Model and data integration
Conceptual design
Model-based query
a b s t r a c t
Models, and to a lesser extent datasets, embody sophisticated statements of environmental knowledge.
Yet, the knowledge they incorporate is rarely self-contained enough for them to be understood and used –
by humans or machines – without the modeller’s mediation. This severely limits the options in reusing
environmental models and connecting them to datasets or other models. The notion of ‘‘declarative
modelling’’ has been suggested as a remedy to help design, communicate, share and integrate models. Yet,
not all these objectives have been achieved by declarative modelling in its current implementations.
Semantically aware environmental modelling is a way of designing, implementing and deploying envi-
ronmental datasets and models based on the independent, standardized formalization of the underlying
environmental science. It can be seen as the result of merging the rationale of declarative modelling with
modern knowledge representation theory, through the mediation of the integrative vision of a Semantic
Web. In this paper, we review the present and preview the future of semantic modelling in environ-
mental science: from the mediation approach, where formal knowledge is the key to automatic inte-
gration of datasets, models and analytical pipelines, to the knowledge-driven approach, where the
knowledge is the key not only to integration, but also to overcoming scale and paradigm differences and
to novel potentials for model design and automated knowledge discovery.
! 2008 Elsevier Ltd. All rights reserved.
1. Introduction
Many environmental models relate to the structure of the
systems they represent only in partial, hardly understandable ways.
In fact, very fewmodels can serve as an explanation of the modeled
processes: the understanding of the system is usually implicit and
typically resides outside the model specification and implementa-
tion. As a result, the purpose of models is typically restricted to the
specific application they have been developed for; the potential for
reuse, communication and integration with data and other models
is limited.
Inspired by the declarative programming approaches that
became popular in computer science in the 1980s, declarative
modelling has been suggested as a remedy for the ‘‘black box’’
nature of self-contained models (Robertson et al., 1991; Wenzel,
1992; Keller and Dungan, 1999; Muetzelfeldt, 2004). Declarative
models describe environmental processes in the form of concise
mathematical statements that bear a closer resemblance to the way
environmental scientists conceptualize systems. Declarative
modelling approaches have been incorporated in graphical
languages that can produce readable and fairly self-explanatory
model statements, e.g. Simile (Muetzelfeldt and Massheder, 2003)
and STELLA (Richmond, 2001), or in specialized conventional
languages such as Modelica (Tiller, 2001). Such approaches have
become popular enough with environmental scientists to make
libraries of reusable model components conceivable (Salles et al.,
2006). Yet, truly modular model repositories, capable of meeting
large-scale integration goals, have largely remained wishful
thinking. One reason is that declarative approaches to modelling,
while greatly enhancing readability of model components, have
mostly focused on syntactic aspects rather than semantics:
declarative models use convenient abstractions that are relevant to
the modelling process (such as ‘‘stocks’’ and ‘‘flows’’) but differen-
tiate processes and environmental entities merely by means of
variable names, and remain incapable of incorporating any formal
statement about what those names mean. It is close to impossible
to ensure that the requirements of two models match with suffi-
cient precision to allow their mergingwithout ‘‘conceptual’’ error, if
the semantics is not fully explicit. Model integration in software
terms does not guarantee sound integration of the model logics
(Athanasiadis et al., in press), and declarative modelling is no
exception on this principle. While declarative modelling makes
* Corresponding author.
E-mail addresses: ferdinando.villa@uvm.edu (F. Villa), ioannis@athanasiadis.info
(I.N. Athanasiadis), andrea@idsia.ch (A.E. Rizzoli).
Contents lists available at ScienceDirect
Environmental Modelling & Software
journal homepage: www.elsevier .com/locate/envsoft
1364-8152/$ – see front matter ! 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.envsoft.2008.09.009
Environmental Modelling & Software 24 (2009) 577–587

equations readable and easy to simulate within appropriate
computing environments, the environmental knowledge itself has
not been made explicit to the full extent.
In recent years, large-scale visions such as the Semantic Web
(Lee et al., 2001) have started to investigate the problem of
communicating shared resources (such as web documents) inways
that make meanings explicit and meaningful, automatic associa-
tions possible. Fueled by such developments, attention to the
semantic nature of shared data and resources has begun to perco-
late down to the natural sciences. Semantic annotation of datasets
and models (Kiryakov et al., 2003; Athanasiadis, 2006, 2007; Parr
et al., 2006; Khatri et al., 2006; Rizzoli et al., 2008; Villa, 2007; Villa
et al., 2007; Chen et al., 2007; Lee et al., 2007; Madin et al., 2007) is
now a recognized topic in modelling in general, and in environ-
mental modelling in particular. The rationale of such developments
is easing integration and reusability of environmental datasets,
models and processing dataflows. Yet, many of the potential
benefits of a semantically explicit environmental modelling remain
unexplored.
This paper discusses and previews the road ahead of conven-
tional declarative modelling. We discuss the current approaches to
environmental modelling that exploit the formalized semantics of
natural systems to unify representations of data and metadata,
improve their usability in scientific workflows, and ease the defi-
nition of dynamic models. The approaches we discuss link the
notions of declarative modelling, metadata, scientific workflow,
data integration and environmental databases in a synthesis of the
most recent advances in ecoinformatics and knowledge
representation.
Because the key tool to allow this unification is the use of
structured knowledge (organized in ontologies) to inform data and
metadata compilation, model conceptualization, and simulation,
we start with a brief introduction to ontologies geared to envi-
ronmental modelling applications. In Section 3 we shall discuss the
ways in which formal knowledge is being incorporated in envi-
ronmental data and model design, and how environmental
sciences are moving towards more formal statements of meaning
as an integral part of the activity of modelling. In Section 4 two
approaches to semantic modelling are presented. The ‘‘mediation
approach’’ where semantically enriched datasets and declarative
models facilitate integration and reuse will be discussed first. The
knowledge-driven modelling approach, discussed next, makes the
knowledge explicit and the model logics implicit, only requiring
a declaration of the modeled system; we discuss how a simulation
strategy can be inferred from it by means of machine reasoning,
and how the declaration of amodel is merely an extension of that of
a dataset. In the last section (Section 5), we shall discuss novel
scenarios and scientific approaches that are made possible by the
adoption of knowledge-based perspectives in environmental
modelling, and the corresponding remaining challenges and
uncertainties.
2. Formalizing knowledge: ontologies
The term ontology originates in philosophy, and refers to the
study of being, but nowadays it is also used more widely and with
eminently practical meanings. In artificial intelligence, an ontology
is any formal description of a conceptualization of a domain of
interest (Gruber, 1993, 1995). In computer science, an ontology is
a formal data structure that describes a conceptual domain, usually
consisting of a set of statements (axioms) that define concepts and
relationships between concepts (Wand et al., 1999). Languages,
formalisms, and tools to create, store and communicate ontologies
have proliferated in recent years (e.g. OWL (Guinness and van
Harmelen, 2004)). It is now common to identify an ontology with
a web-accessible document that can be created, edited and
validated using ad hoc tools. Ontologies are used as references to
annotate resources with concepts in standardized ways, e.g. in RDF
(Beckett, 2004).
In ontologies, a concept (or class) is the statement (definition) of
an entity, usually including at least a textual description and a short
label. Concept names (IDs) and descriptions do not definemeanings
of an entity formally: the actual meaning of concepts is always
defined through properties, which relate concepts to one another.
For example, the identifier ‘‘person’’ means little by itself; what
makes a ‘‘person’’ identifiable as such is having a name, an address,
relatives and other personal data. By making these properties
explicit, ontologies allow an automated process to recognize the
formal description of a person from that of a different object by
checking its properties and the logical constraints that accompany
them.
A property is the statement of an association between concepts.
The need for mathematical and computational tractability of the
resulting conceptualizations constrains most ontology frameworks
to handling binary properties, where a concept is associated to one
other only. Relationships are statements that link an instance to the
value of a specified property, which can be a concept, an instance or
a literal value such as a numeric or textual value. Ternary statement
such as ‘‘John birth date 10–10–1972’’ can be used to state rela-
tionships, which must comply with the model specified by the
concept and its properties. Properties may define a cardinality
constraint, which limits the number of allowed values (e.g. a pop-
ulation must have at least one individual and only one species).
Properties allow building the bones of knowledge representa-
tion structures: notably, the subsumption property (which can be
stated as ‘‘is-a’’, ‘‘kind-of’’, or ‘‘is-type-of’’) allows constructing
generalization–specialization (classification) hierarchies. Such
structures are often called taxonomies and allow specializing
meanings into less and less general concepts. In a very simple
example, the concept ‘‘Ecological assemblage’’ may be specialized
into ‘‘Population’’ and ‘‘Community’’; the latter may in turn be
specialized into ‘‘Floral’’ and ‘‘Faunal’’ and so on. Note that the
subsumption of concepts requires the inheritance of relationships:
all properties of an ecological assemblage (such as being composed
of one or more individual organisms) must also be properties of
populations and communities, although their meaning may be
restricted in specialized concepts to define narrower logical models
(e.g. floral communities may only include individuals of plant
species). Another common relationship in natural system ontol-
ogies is ‘‘part-of’’, used for example to guide collection and classi-
fication of plant specimens (Paterson et al., 2004). Properties can be
generalized or specialized just like concepts: e.g. ‘‘economic-
dependency’’ is-a ‘‘dependency’’.
Instances (also called Individuals or Objects) are the statement of
a real-world entity and ultimately the incarnation of a concept
defined in an ontology. Compared to a concept, an instance cannot
be further specialized and may not have additional or restricted
properties compared to those of the concept it incarnates. A
collection of instances that conform to an explicit ontology is often
called a knowledge base, although the term is not rigorous. The
records of a database with a formally specified relational schema
can always be considered a set of instances of the correspondent
ontology; the database paradigm provides an interface and opti-
mized methods for storage, search and retrieval of instances based
on matching property values to user-defined constraints.
Ontologies can be used at different degrees of internal
complexity and expressive power (Fig. 1). The simplest use is as
controlled vocabularies, consisting of a loosely structured set of
concepts with no properties, whose names define allowed terms
for purposes of harmonization of terminology. This usage is
becoming common in environmental applications (Baker et al.,
2000; AGROVOC, 2006; Batzias and Siontorou, 2006; Duepmeier
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587578

and Geiger, 2006). Terms of a controlled vocabulary often include
concept name translations in several languages. Yet, a useful
ontology typically contains more than just a list of terms and their
definitions. A step further is a taxonomy, where concepts are
arranged in a generalization (is-a) hierarchy as seen above. Other
common ontologies establish relationships like ‘‘broader/narrower/
related’’ between concepts: such ontologies are often referred to as
thesauri. The full power of ontologies is exploited when domain-
specific properties are defined along with concept classes.
Controlled vocabularies, thesauri and taxonomies can all be
evolved into full ontologies by defining custom properties. Exam-
ples of full ontologies in the natural sciences are abundant. For
example, the SWEET ontologies by NASA (Raskin and Pan, 2005;
NASA, 2006) propose a formalization of common knowledge about
the physical world and the biosphere, designed to facilitate coor-
dination and communication of applications and agents in earth
and space sciences.
These normative uses, where ontologies provide guidance for
conceptualization, can be thought of as the generalization of
formalisms that are commonly adopted in natural system infor-
mation management and modelling, such as metadata standards
(Jones et al., 2001; ISO, 2004), database schemata or object-
oriented models. The main rationale for the existence of ontologies,
however, is to enable machine reasoning on natural systems, either
by human actors or by a computer program (reasoner). The main-
stream ontology formalisms are carefully designed to only allow
axioms that preserve decidability and therefore guarantee their
usability for machine reasoning. The main operations in automated
reasoning are subsumption (inferring that concept A is more
specific that concept B) and classification (inferring that instance X
is an incarnation of concept A). A common framework for
reasoning, Description Logics (Baader et al., 2003) provides ways to
state necessary and sufficient conditions for instances of an
ontology to belong to concepts, so that logical inference (reasoning)
can be performed on them to attribute concepts and instances to
independently defined classes.
Reasoning is the key operation to enable sophisticated integra-
tion of datasets and models. For example, appropriately defined
concepts may accompany variables in a model and columns in
a tabular dataset. This in fact makes an instance out of the data.
When that is done, datasets (or model outputs) can be deemed
suitable to be used as the value of an input variable if and only if
a reasoner can classify them as an instance of the same class as the
variable’s. In the following sections, we will describe applications,
problems, and perspectives in applying ontologies and machine
reasoning to different sectors of environmental modelling and
management.
3. Knowledge models in ecology and environmental science
Ecological and environmental modelling activities are obviously
related to knowledge representation and management. Natural
systems knowledge is formalized in various ways, varying from
scientific publications, taxonomic classifications, data collections,
modelling theories, to simulation models and results. Ontologies
are increasingly used to support natural system modelling and
enhance its rigor and consistency. For example, Brilhante (2005)
proposes Ecolingua, an ontology for ecological quantitative data
that attempts to enable the synthesis of conceptual ecological
models from data descriptions through the reuse of existing model
structures; Ceccaroni et al. (2004) used an ontology to augment
and improve the formalization of knowledge to be used in case-
based and rule-based reasoning decision support system for
wastewater management. Scholten et al. (2007) use ontologies to
implement a methodology to support the development of multi-
disciplinary models for water management, providing formal
guidelines in the model development process. Currently, several
projects are trying to address the need for a pool of shared concepts
that natural system scientists can refer to for annotation of data
and models. The SEEK project (SEEK, 2004) is eliciting ontologies
from a large community of ecologists to enable semantic annota-
tion of datasets and processing steps in scientific workflows (Wil-
liams et al., 2007). Similar efforts are underway in other
disciplinary domains, e.g. in genomics (Ashburner et al., 2000),
atmospheric physics (McGuinness et al., 2007), water quality
(Chau, 2007) and agriculture (AGROVOC, 2006). Not surprisingly,
providing conceptual frameworks to express the vast realm of
natural systems knowledge is far from trivial: natural systems
sciences need to address the most complex domains and are still
far from the general acceptance that much simpler, single-scaled
conceptualizations, such as the Gene Ontology (Ashburner et al.,
2000) can enjoy today.
In this section, we relate the vision of a knowledge-based
approach to ecological and environmental modelling to the
conventional notions of data-driven and declarative modelling.
This discussion paves the way for that of explicit semantic model-
ling introduced in the next section.
3.1. Data-driven modelling
Datasets can be considered ‘‘static models’’ of systems: they
always conform to conceptual models, and depend on (usually
implicit) assumptions and world views just as much as dynamic
models do (Villa and Costanza, 2000; Villa, 2001, 2007). Conven-
tional data management practice separates ‘‘raw data’’ – actual
numeric or categorical values – from ‘‘metadata’’, the information
that allows a user to make sense of the numbers, providing needed
spatial and temporal contexts, unit of measurement, method and
accuracy information to complement the raw information.
In an ontology-informed framework, the starting point is always
the accurate formalization of the concept that is quantified in the
data, including the definition of its characteristics represented as
properties. Whether the link to the ontology is made after the
measurement (annotation) or beforehand, datasets can be seen as
instances of an explicit, formal concept. Data are associated to
concepts through one property: e.g. in a trivial case, a ‘‘numeric-
value’’ property could link a concept to the raw data, represented as
Fig. 1. A pictorial representation of common knowledge representation paradigms
ordered by increasing expressivity and complexity.
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587 579

textual (literal) values. The remaining properties incorporate the
‘‘metadata’’, which in this case are defined by and connected to an
explicit knowledge model. The benefits of this approach lie in the
richness and logical consistency gained by using a formal ontology
as the founding knowledge model for data and metadata. The latter
are specified in forms that allow mediation, alignment and
consistency checking across several sources; alignment between
different knowledge models adopted can be performed by auto-
matic reasoners if appropriate bridging information is supplied.
The ontology paradigm in its simplest incarnation offers a field-
by-field substitution for the tabular and relational schemata
conventionally used to represent data. Schema information con-
tained in normative ontologies can for example be seen as an add-
on to an existing database management system, implemented as
a software ‘‘wrapper’’ for a relational database engine or other
suitable technology (such as an XML database). In previous work,
we have investigated how to automatically transcribe logical
models expressed as ontologies into object-oriented models and
relational schemata (Athanasiadis et al., 2007a, b).
In addition to that, ontologies offer a powerful synthetic way to
specify both the data schema and the structure of the knowledge
behind it. While a relational schema can be seen as expressing the
structure (the ‘‘what’’) of the knowledge, ontologies also allow
specifying the how and the why at the same time. The database
paradigm that includes a rich knowledge model such as that
specified by ontologies is often called deductive or intelligent data-
base (Bertino et al., 2001), whose applications in the environmental
domain are beginning to appear (Villa et al., 2007).
Many common metadata standards have been formalized as
ontologies: e.g. the ISO 19115 standard for spatial information (ISO,
2004). Yet, these efforts typically stop at the level of controlled
vocabularies: they capture syntax (names and information
containment hierarchies) rather than meaning. Providing a set of
ontologies to define the actual meaning of the metadata properties
is a much more sophisticated activity, and lends itself to different
interpretations that reflect the complexity of the domain, including
the existence of different viewpoints that can sometimes be
complementary and other times competing. The ontologies chosen
ultimately depend on (and define) the applications; the choice
becomes crucial in determining the scope of potential integration
with other resources. Semantic mediation (Ludaescher et al., 2001)
is the attempt to integrate data and models using sets of ontologies
(see Section 3.2). Projects such as SEAMLESS (SEAMLESS, 2005),
SEEK (Kepler, 2004; SEEK, 2004; Madin et al., 2007), ARIES (Villa
et al., 2008) and IMA (Villa, 2001, 2007; Villa et al., 2007) are
endeavoring to develop sets of ontologies that best fit large-scale
application domains. Other projects have been relatively successful
in more restricted domains, e.g. food web theory (Webs on the Web
(URL)).
3.2. Systems theory and declarative modelling
Dynamic models, like data, always conform to a conceptualiza-
tion, and no philosophical difference needs to exist between
specifying data and models when this is done using ontologies
(Villa, 2001, 2007). Any model that can be ‘‘declared’’ can also be
specified as a set of instances of appropriate ontologies. The main
difference between data and models is the increased conceptual
richness necessary to describe how things are created, destroyed,
and modified in time and space. This requires at least notions of
linkage between concepts with causative or dependency relation-
ships that are normally not necessary to specify data. It also
requires developing ways to interpret this causality. The set of
abstractions (concepts) that allows conceptualizing and expressing
those cause–effect relationships and their results is the adopted
modelling paradigm, of which there are many examples
(e.g. ordinary differential equations, stock-and-flow, or individual-
based). A modelling paradigm, like any consistent conceptualiza-
tion, is amenable to being at least partially captured into the logical
constructs of an ontology. In fact, most existing modelling software
systems (e.g. STELLA: (Richmond, 2001); SIMILE: (Muetzelfeldt and
Massheder, 2003)) conform to one implicit, usually simple,
ontology, which defines their notion of entities familiar to the user,
such as state variables, flux variables, etc. Advanced integrative
systems (Kepler, 2004; Villa, 2007) can manipulate different
ontologies, which, supplemented by the necessary software, enable
them to orchestrate models that adopt heterogeneous modelling
paradigms. Such systems are in the best position to enable inte-
gration of independently developed models adopting different
paradigms into a higher-level, multiple-paradigm model.
In declarative modelling, model specification is based on the
attributes and semantics of the natural systems, rather than on the
algorithm (numerical integration procedure) that calculates their
results. For example, the graphical equivalent of a simple predator–
prey model (Fig. 2) shows how the algorithms that integrate the
relative difference equations have been substituted by the state-
ment of relevant variables and their mutual dependencies. Such
declarative specifications use an implicit ontology of modelling
entities that embodies a chosen paradigm (e.g. the stocks and flows
that reinterpret state variables and rates of change), but the
knowledge (meaning) about the environment must still be ‘‘infer-
red’’ from the names of the variables and the reconstruction of the
process.
Ontologies can support declarative modelling by providing, at
the same time, schemata for model declaration and meaning for
these schemata. Instances of ontologies represent declaratively
expressed models that refer to concepts laid out in the ontologies.
Such declarations contain enough information to enable a software
infrastructure to simulate the behavior of the systems represented
over a user-defined temporal and spatial extent. Thanks to the rich
meaningmade possible by ontologies, aworkflow environment can
properly connect models to data, and feed quantities calculated by
simulation to other models in the same environment.
4. Semantic environmental modelling
In semantic modelling, all concepts used to model a natural
system are explicitly defined by ontologies. The semantics of
a model is the intersection of two languages: one that describes the
natural knowledge, with concepts such as population, individual or
growth rate, and one that represents the modelling process, with
concepts such as variable, stock or flow. The declaration of a model
in a semantically enriched way, using ontologies, can be achieved
by specifying:
Fig. 2. Simple predator–prey system modeled according to common conventions of
graphical stock-and-flow languages. Squares represent state variables, circles represent
parameters and functions, and the remaining symbols represent rates. Influence
arrows represent dependence of a rate or variable value on the current value of others.
The ecological knowledge is represented implicitly in the variable names and model
structure.
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587580

a. the modeled entities, by identifying the relevant concepts and
properties and creating instances of them for each model
component;
b. the underlying relationships among these entities, capturing
the structure of causality in the system as understood by the
modeller.
Software implementations can automatically check that
a semantically enriched model specification is self-consistent, and
with appropriate support information, can go as far as ensuring
matching contexts of both space and time for all entities before the
model can be accepted or calculated. This preliminary semantic
validation prevents users from defining inconsistent models and
helps to retrieve compatible data sources from databases when the
model is applied. This kind of inference can be used to delegate to
an ontology-aware software system much of the difficulties of
properly designing and deploying a complex simulation model,
while facilitating its application by non-scientists such as decision
makers, and, at the same time, ensuring proper, correct use of both
model and data and the soundness of the results. Ontologies are
also valuable for adding ‘‘meta-information’’ related to both model
equations and variables, maximizing the potential for reuse. This
can be achieved by decoupling model interface elements and
equations, which allows specifying model interfaces elements
(variables or parameters) as measurements (Athanasiadis et al.,
2006). This way amodel exposes as inputs not only a set of numeric
values, but also a clear specification of dimensions, units and spa-
tio-temporal context. Because semantic links among models can be
logically verified, using ontologies in model design defines a clear
path to specifying reusable model components (Athanasiadis,
2006).
Two distinct approaches to incorporating ontologies in model
design can be identified. In the mediation approach, concepts from
ontologies supplement conventional data and models to facilitate
integration and reuse of the information. In the knowledge-driven
approach, datasets and models are directly represented as
instances of ontologies and embody a statement of the system
conceptualization, enabling machine reasoning about the system
structure that can lead to more sophisticated applications. We
describe both approaches in the rest of this section.
4.1. The mediation approach: data integration in
analytical workflows
The mediation approach for environmental knowledge repre-
sentation consists of the enrichment of existing data and environ-
mental models with formal semantics to connect measurements
and variables to the identity of the observable entities they quan-
tify. This approach is gaining ground as a mean to overcome
significant obstacles to the reuse of available computerized legacy
models or archived data sources. As discussed earlier, the main
reason for this situation is the typically poor semantics for repre-
senting environmental information and the absence of a standard-
ized framework to properly captures the represented knowledge in
the forms of annotations. In such a situation, managing and pro-
cessing environmental information require ever-increasing efforts.
Scientific workflows define information flows among existing
simulation models and data sources, and typically require intensive
pre-processing activities for data manipulation, scaling, formatting,
and so on. Without appropriate semantic information, the knowl-
edge that makes such tasks possible must be supplied externally, at
the cost of considerable human effort, only little of which can be
reused in a different context.
A semantically aware framework using the mediation approach
is typically data-centric, as it utilizes ontologies for describing the
representational contexts of environmental data, either residing in
databases or produced by models. From a representational point of
view both models and databases can be considered as ‘‘data nodes’’
that provide environmental data, operating either as data sources
(databases, files, datasets, etc.) or as data converters that include
both environmental models and pre- and post-processing algo-
rithms (such as statistical data reduction or calculation of indicators
from model results). The only difference between the two kinds is
that in the first case data are permanently archived and are avail-
able via querying, while in the second are dynamically produced,
and are made available via computation.
Defining the representational context of knowledge captured is
an important activity for annotating properly data sources and data
converters, and is critical for reusing them by constructing scientific
workflows (Ludaescher et al., 2005; Athanasiadis, 2007). A scientific
workflow is a pathway between two or more processing steps,
along which data are transformed until a desired result is reached.
Users assemble workflows, by connecting data nodes together and
linking the results to storage or visualization facilities. A workflow
environment is thus a ‘‘blackboard’’ for users to assemble the
information flow needed to address a particular problem. In func-
tional terms, a dynamic model can often be seen as a workflow,
because in the end all models process input information and
produce a set of output results. The similarity, however, does not
hold when models and workflows are considered in semantic
terms: the meaning of a model is not to produce outputs, but to
describe a natural process. In fact, a more correct generalization
sees workflows as special cases of models, whose ‘‘paradigm’’
entails data transfer and transformation along the connections of
an artificial system.
In this sense, a workflow is amenable to the same ontology-
based description as any other model: concepts of ‘‘input’’,
‘‘output’’, ‘‘processing step’’ can be specialized as needed using
ontologies that can describe all steps of any workflow environment.
The most important use of ontologies in workflow environments,
however, is another: to allow the system to enforce meaningful,
correct connections between inputs and outputs, and – if necessary
and possible – insert transformation steps in the workflow that
guarantee a proper match. The operation of enforcing and sup-
porting semantic consistency along data paths in workflows is
usually called semantic mediation (Kohler et al., 2000; Ludaescher
et al., 2001) and it is essential to guaranteeing correct results
(Athanasiadis, 2006; Villa, 2007), particularly when processing
steps are heterogeneous and users are not domain experts. To allow
semantic mediation, all inputs and outputs must come tagged with
concepts from ontologies that are known to the workflow envi-
ronment; the latter needs to use a reasoner program to ensure the
consistency of concepts along each connection made by the user.
The operation of associating real-world entities (e.g. input and
output ‘‘ports’’ of workflow components) to concepts from ontol-
ogies is usually called semantic annotation (Kiryakov et al., 2003),
and it is done by the same actors that have developed themodels or
processing steps. Conceptual compatibility of data nodes interfaces
(inputs and outputs) is tested with a reasoning operation that
checks if the data being fed from an output to an input represent
the same concept. In a simple example, a model X is ‘‘packaged’’ as
a workflow component and all its inputs are semantically anno-
tated by its developer according to a set of commonly understood
ontologies. The semantic annotation operation requires that all the
conceptual details of each ‘‘port’’ (or interface element) are
understood and appropriately defined. As an example, an input Xi
representing temperature at the earth’s surface may require that
the temperature is expressed as monthly data over the simulated
timespan, and the model has only been calibrated for temperatures
in the 19–30 !C range, so it should not accept data outside of these
boundaries. Semantic annotation is a way to express such condi-
tions, which commonly are only expressed verbally in the model’s
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587 581

documentation, in a formal and machine-readable way. In order to
do so, an ontology is created to define model X, with concept
defined for each exposed ‘‘port’’. Using restrictions and concepts
from appropriately linked ontologies, the concept definition asso-
ciated with input Xi may look similar to that shown in Text box 1.
When a semantically annotated model is used in a workflow,
inputs and outputs are connected by the user. For example, a time
series of temperature data retrieved from a database may be con-
nected to input Xi. Upon connection, a semantically aware work-
flow environment can ensure the appropriate match between the
input and the output by feeding the respective semantic annota-
tions to a reasoner and ask if they describe the same concept (a
classification operation). A reasoner can make the necessary
inferences to deduce the equivalence of types that have different
names, based on their properties. In some cases, a verdict of
conceptual compatibility may be sent back to the processing
environment to further check if a transformation needs to be
inserted in the dataflow to make the input and output numerically
compatible. A typical case is conversion of compatible numeric
values that are expressed in different units of measurement. In
a more complex example, requiring a higher sophistication of the
processing environment, consider a data source of weekly data
rather than monthly. A sophisticated workflow environment can
understand that the data need to be aggregated into amonthly time
scale before being passed to the model that uses them, and direct
the workflow environment to create a transformation step to
perform the aggregation and insert it between the data source and
the model. Although decisions of this complexity are beyond the
capabilities of first-order reasoning alone, software systems can be
assisted by reasoners and ontologies to determine the proper
operations to apply in order to mediate different, compatible
contexts for the information exchanged in dataflows.
The mediation approach can be successfully applied to
conventional declarative modelling. A dynamic system can be seen
as described by a set of state variables whose numeric state changes
under the influence of phenomena defined in instantaneous terms
(rates) through differential equations. A commonly accepted
declarative modelling interpretation of such systems, incorporated
in several visual modelling environments such as Stella (Richmond,
2001) and Simile (Muetzelfeldt and Massheder, 2003), uses the
abstractions of stocks and flows to refer to state variables and rates
(Fig. 2). The use of these abstractions has become commonplace in
ecological and environmental modelling, to the point that they
constitute a good starting point for conceptualizing a modelling
ontology that can be widely understood.
A declarative modelling language defining phenomena in terms
of stocks and flows is highly enriched if these dynamic aspects can
be related to the definition of the actual physical entities where
they take place. So for example, instead of simply providing the
stock and the two flows that describe change in population size, we
can adopt a language that allows us to define the population entity,
and include in its definition the fact that the ‘‘population numer-
osity’’ variable represents its state, and the phenomena of birth and
death are the processes that influence it. By specifying the pop-
ulation in terms of its actual semantics (including, e.g. the property
of being a population of hares) we greatly facilitate the process of
connecting this model to existing data for initialization, and of
ensuring that any connection of its output to other components in
aworkflow is meaningful. Model design is also greatly facilitated by
the fact that the semantics of a population captured in the ontology
can serve as a blueprint for the design of the model.
4.2. Knowledge-driven modelling
If the mediation approach enriches an existing representation
with formal knowledge, in knowledge-driven modelling the formal
knowledge structure is the model. This is accomplished by defining
environmental models (either static or dynamic) directly as
instances of adequately expressive ontologies. Because different
ontologies can be freely combined in a conceptualization, knowl-
edge about the natural domain and the modelling paradigm can be
pulled together to completely define a natural system, pairing
knowledge about model dynamics with the definition of the rele-
vant ecological entities. When the model is stated as instances in
a knowledge-based framework, a reasoner-supported system can
be employed to render models declaratively and pass them to an
execution environment for simulation. A model is at this point no
longer just an annotated data node, semantically equivalent to
a data source; at the same time, it does not operate as a black box
whose interface only (inputs and outputs) is documented, as in the
mediation approach. On the contrary, a model expressed entirely
using ontologies becomes a statement of the logics that constitutes
the understanding of the modeled system, and captures the
processes involved and its interpretation on behalf of the modeller
in full semantic detail.
To explain how this approach can be realized, we use the
common ‘‘hare and lynx’’ predator–prey system. Typically, a pred-
ator–prey system evokes a nonlinear differential equation system
or a discrete ‘‘stock and flow’’ model in the mind of most ecologists.
Figs. 3 and 4 explore this model in a knowledge-based framework:
in Fig. 3, the ‘‘identity’’ of the system is captured in a particular
instant of time, while in Fig. 4 enough knowledge is added to enable
a suitable software system to infer a dynamical model by using
machine reasoning alone.
In both cases, the ‘‘hare-lynx’’ system uses concepts from
biodiversity ontologies that describe populations and communities.
In the same way, it can be coupled with a taxonomy ontology that
allows defining species unambiguously by referring to species
identifiers from known repository services such as that provided by
GBIF (GBIF, 2004). Using taxonomic resolution services can be key
to integrating models with species data from remote datasets that
have been annotated with the same conventions.
The knowledge model in Fig. 3 is essentially a semantically rich
dataset: it states that at a certain time two populations coexist in
a community, and have specified numerical abundances. Yet, it’s
much richer than a typical dataset with conventional metadata, or
a conventional model component with the associated documen-
tation, because of the extra knowledge layer expressed by linking to
ontologies. The system as defined this way is essentially the non-
dynamic model of a two state variables system.
The switch between a static and a dynamic model is operated by
providing information on the causal connections among the
different entities, and on how these connections influence the
evolution of state variables through time. Causality in conventional
models is usually expressed through equations associated with the
value of variables. Equations, by referring to the values of other
Text box 1. Semantic annotation of a required model input.
I :: ¼
is-a: Temperature,
vertically-distributed-in: PlanetarySurface,
has_unit: Fahrenheit,
max-value: (is-a: Temperature, has-value: 30.0, has-
unit: Celsius)
min-value: (is-a: Temperature, has-value: 19.0, has-
unit: Celsius)
distributed-in: (is-a: TimeSpan, step: 1, has-unit:
Month).
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587582

variables, implicitly define causal relationships that correspond to
computational dependencies. An ontology-based framework can
make these dependency relationships explicit, and add semantics
to them by further specializing these dependencies. For example,
a generic depends-on relationship can be specialized into a flows-
into relationship between a state variable and a flux variable (rate).
The presence of this relationship in the knowledge-rich definition
of a model informs the underlying software architecture that the
flux must be integrated over time in order to assess its contribution
to the value of the state variable. In other words, equations in
a model definition can be seen as statements of specialized
dependency relationships. The notion of variable, so central to
conventional approaches, can similarly be enriched and made
dependent on the modeled entity. For example, in an individual-
based paradigm, variables describe quantitative traits of modeled
individuals, but maintain the link to the individual which is the
main entity considered. No conflicts need exist between paradigms,
whose conceptual boundaries often become blurred when
a explicit knowledge-based approach is used, particularly if notions
of scale are formally defined (Villa, 2007).
The system conceptualization sketched in Fig. 4 adds enough
information to the system of Fig. 3 to define how changes in time
are enacted. To this purpose, the specification of the population
abundances is extended to inform the system of how they change in
time. The abundance of the lynx population is now not only
a numeric value exposed in the model interface, but it is also made
Fig. 3. Portions of a possible conceptualization of a predator–prey community. The system is conceptualized in a static way, meaning that the temporal variability is captured as
multiple data points and no attempt is made to describe the causal factors that make the abundances vary with time.
Fig. 4. The same system as in Fig. 3, with the addition of causal relationships from a modelling ontology. Such details allow an appropriate software system to infer a simulation
workflow for the system.
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587 583

a stock, by defining it as an instance of a ‘‘stock’’ concept from
a modelling ontology that has been added to the conceptualization.
Similarly, its time specification declares that while its current value
is an hourly measurement equal to the initial value of the system in
Fig. 3, there is an extent now defined, that expresses the fact that
such hourly values will exist over the span of one year, determining
a multiplicity of states for the system. Corresponding flows are
added and linked to the existing model to express the factors that
influence the change in the stocks. A knowledge representation
framework for environmental modelling can analyze such decla-
ration and decide that in order to know the numeric state of the
system over the given extent, the stock and flow identities require
difference equations to be defined, and the initial state must be
extended over the time extent by integrating the flows.
There are limits to the declaration of models in current ontology
frameworks, particularly if reasoning capabilities must be
preserved. These limitations stem from the fact that any depen-
dency other than linear requires higher-order logics statements to
be fully formalized, and the handling of higher-order logics is
beyond the capabilities of existing reasoning systems. The domi-
nant paradigm, Description Logics (Baader et al., 2003) can only
operate on a subset of first-order statements that guarantee
decidability and computability in finite times. For example, in
description logics it is possible to express the fact that the birth rate
of the hare depends at any given time on both the abundance of the
hare and that of the lynx; it is, however, impossible to express the
nonlinear dependency that translates the notion of predation.
Therefore, machine reasoning cannot be used to assess facts rela-
tive to the species interaction unless additional information is
given. The mechanism of interaction between the species can only
be captured with externally processed information such as an
equation, whose syntax can be expressed through a literal property
and is easily parsed by software, but whose logical underpinnings
remain obscure to a reasoner.
Even within the constraints of first-order logics, reasoning can
be profitably used to enforce sound designs and consistent defini-
tions of models. As an example, a biodiversity ontology used to
model an ecological community can state that coexistence of
populations in a community (e.g. as captured in the coexisting-
population relationship) also implies coexistence in space and time,
a constraint that can be checked by an appropriately configured
semantic modelling system which will refuse, for example, to
simulate a systemwhere lions prey on dolphins simply by checking
population distribution data.
A knowledge-driven specification enables an appropriately
configured system to automatically generate an algorithm (the
actual simulation model) capable of generating a simulated dataset
from given values of parameters and initial conditions. In this
specific case, the system in Fig. 4 is easily compiled into a system
similar to that of Fig. 2, which can later be calculated by numeric
integration. A semantically enabled environment can target
different compilation or execution environments, for example
compiling the simulation in a high-level programming language or
into a declarative workflow system.
In the knowledge-driven approach, what has conventionally
been called the modelling paradigm can be incorporated to
a certain degree in the reasoning strategy that ‘‘resolves’’ the
specification into a running simulation. As a result, the choice of
modelling paradigm may become less relevant to the modeller. The
real paradigm in such a system is captured in the ontology used to
lay out the conceptual definitions of the system. Finer-grained
paradigm choices such as how to interpret the differential equa-
tions composing the system (e.g. continuous time vs. discrete time)
can be absorbed in the knowledge base of the framework; users
only need to concentrate on accurately describing the system using
known, well-documented concepts and their relationships. To this
end, knowledgemodelling tools such as GrOWL (Krivov et al., 2007)
can be used effectively. For example, an individual-based solution
for the system of Fig. 4 is also possible; decisions can be taken
automatically based on the ontologies themselves and on the
comparison of efficiency metrics. For example, Villa (2007)
describes a hybrid model where reasoning is used to decide the
best strategy to model a coupled system only after independent
conceptual models have been merged into a higher-level one.
No matter what strategy has been chosen to simulate the
system, the result of simulating a system will produce a semanti-
cally enriched result that preserves the original identities. For
example, simulating the system of Fig. 4 may produce a result very
similar to the static system of Fig. 3, only with multiple abundance
values per each time step. The system will automatically define
a process in terms of variables, inputs, outputs, concepts that users
won’t need to manipulate unless they want to, and use it to
‘‘resolve’’ the dynamic formulation into numeric states. The
dynamic specification, shown partially in Fig. 4, is considerably
more verbose, but still recognizable as an extended version of the
previous ‘‘static model’’. It is noteworthy how metadata, units, and
other information are the same: storage, query or inference can use
the same infrastructure for both the static and dynamic model. The
boundaries of ‘‘dataset’’ vs. ‘‘model’’ can be effectively blurred by
a full semantic specification, reducing to a choice of level of detail
and representational framework for a similarly conceptualized
system.
5. Discussion and perspectives
Advanced knowledge-based systems, e.g. IMA (Villa, 2001,
2007) and Kepler (Kepler, 2004), are being prototyped and have
been applied to selected case studies (Pennington et al., 2007; Villa
et al., 2007, 2008). Such systems are typically not committed to
a particular set of concepts, with the possible exception of a small
core ontology, carefully designed for generality, paradigm
neutrality, and extensibility. The uncoordinated extensibility of
such environments allows domain experts to produce knowledge
describing specific disciplinary contexts without having to employ
specific tools. Users can adopt the necessary concepts to produce
representations of natural systems; the software architecture,
informed by the ontologies, can resolve the system model into
numeric states. Such prototypes are laying the groundwork tomake
a purely semantic approach possible, where all technological
details related to the calculation of the model are inferred by
machine reasoning based on the logical assertions that define the
model, while remaining hidden from the user. The rationale for this
approach is the notion that an accurate description of nature’s
entities needs also to be complete: if enough knowledge has been
given to allow a full description of the system, the description also
embodies enough information to allow a system to calculate the
corresponding model.
The advances discussed so far introduce potentials that were
obviously not available to ecologists and environmental scientists
before ontologies were available. Even the ‘‘lighter’’ semantic
approach, which uses formal concepts to enrich conventionally
specified model and data structures, adds a powerful dimension to
modelling because of the novel integration potential. Adding to
that, initiatives in environmental planning (Villa et al., 2008) and
ecology (SEEK, 2004) are being paralleled by others in agriculture
(SEAMLESS, 2005), geosciences (GEON, 2005) as well as genomics
and other branches of natural sciences. Most of these initiatives are
coordinating in order to make their core ontologies eventually
converge towards commonly accepted standards. At the same time,
obstacles to widespread adoption of semantic approaches remain
that may reduce or postpone the benefits resulting from these new
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587584

opportunities. In this section, we briefly review the most relevant
potential benefits and the main challenges to adoption.
5.1. New opportunities
5.1.1. Multi-paradigm modelling
Modelling at the conceptual level allows users to employ
a language that’s tailored to the knowledge domain of reference,
adding the necessary dynamic information to the definition so
obtained, and letting appropriate infrastructures define the corre-
sponding computing workflows. Because each modelling paradigm
can be described by a set of ontologies handled by matching soft-
ware, a system can be extended so that, for example, the flow/stock
model can coexist with others where the behaviors of organisms
are modeled individually, instead of relying on community-level
state variables. The details of the scheduling and the interactions
between different calculation workflows can be sorted out auto-
matically. For example, a knowledge-explicit approach can greatly
ease the specification of hybrid models, such as those that are most
necessary in decision making, e.g. landscape models (best modeled
as a spatially explicit process based model) coupled with human
component models that react to changes in the landscape and
influence it in turn (best modeled as individuals moving on the
landscape and reacting to its change). Ontologies can identify
the common ancestor concepts that allow both the modeller and
the infrastructure to represent the coupled model consistently or to
seamlessly merge the independently developed components.
5.1.2. Automated contextualization in space and time
Models are necessary because the phenomena they describe
vary in time and space. The property of being distributed in time
and space causes a multiplicity of states for the variables of
a system. When we model the temporal or spatial heterogeneity in
a system, what is being modeled is not the system itself, but more
accurately the context of its observation. Temporal and spatial
scales of observation can be changed, so that a dynamic model
appears constant, or so that what appears to be a static variable
reveals fine-grained internal dynamics. By virtue of their relative
conceptual independence, time and space can be modeled inde-
pendently from the abstract conceptualization of natural systems;
the definition of the contexts of time and space can be connected to
that of the entities by coupling the definition of the temporal and
spatial contexts of interest with that of the modeled entities and
their behavior.
An important consequence of adopting a knowledge-explicit
approach to modelling is that when space and time become part of
the allowed semantics, there is no need for specially tailored
knowledge or tools for basic spatially explicit modelling, because
such functionalities can be invoked as necessary by the knowledge-
based system, and the paradigms necessary to enact it are auto-
matically integrated into the specification. It is for example
conceivable to make a non-spatial model spatially explicit by
simply describing one or more of its components as distributed in
space (Villa, 2007). A properly configured system can propagate the
notion of space in one concept to the whole conceptual network, or
mediate competing representations by operating transformations,
e.g. to propagate coarse polygon data over a fine-resolution grid.
As a generalization of the contextualization mechanism in time
and space, it is possible to imagine other reasons for a multiplicity
of states in the simulation of a model. As an example, it is possible
for the behavior of the system to depend on conditions or param-
eters whose values are not known exactly, but can be assumed to
have different values according to different likely scenarios or
interpretations. It is useful to be able to model this situation as
a ‘‘classification’’ context, which can be included in a modeled
system to cause the automatic replication of the simulation as
many times as the number of competing explanations of its
behavior. This becomes very useful inwhat-if prediction, sensitivity
analysis and optimization.
5.1.3. Model discovery in databases
In a distributed database context, semantically aware modelling
opens novel and exciting perspectives such as ‘‘model-driven
query’’ (Villa, 2007). An appropriately general version of a model
can become a powerful discovery tool that can be used as
a constraint over a distributed knowledge base or semantic web in
order to discover new knowledge in a fully automated way. As an
example, the concept of a species–area relationship can be defined
as a property of any set of coexisting populations of the same taxon.
A semantically aware infrastructure can automatically translate this
logical statement definition into a query, and launch an iterative
process that identifies all possible instances of species–area rela-
tionships represented by the population data stored in a database.
A foodweb can be described in a similar way. It becomes possible to
match an abstract model structure to a distributed database that is
semantically annotated. By describing patterns of interest in terms
of ontologies, repositories of environmental knowledge with
sufficient semantic information can be searched to automatically
discover patterns and relationships that have traditionally taken
lengthy investigations to find, even when only the necessary data
are present in a database.
5.2. Challenges and barriers to adoption
Despite the clear potential offered by semantically aware tech-
nologies applied to environmental modelling, only limited case
studies are available, and as a consequence the practicality of large-
scale adoption remains to be fully understood. While it is clear that
the advantages of semantic specification of natural systems will
play a role in the future of environmental modelling, the extent to
which this will happenwill depend on both technical feasibility and
acceptance by researchers.
5.2.1. Technical feasibility
While by no means trivial, the engineering aspects of higher-
level knowledge-based systems are well within reach, and proto-
types of knowledge-basedmodelling systems are in use today (Villa
et al., 2007). Ontology frameworks and knowledge editors allow
convenient specification of knowledge and provide standardized
platforms for sharing it. High quality reasoner programs are avail-
able as open source and can be easily run on ordinary computers.
Indeed, at the time of this writing, natural systems science seems
well on its way to the adoption of the mediation approach. Projects
such as ARIES, SEAMLESS, SEEK and GEON strongly emphasize the
role of ontologies to describe data products and analytical steps and
enable their composition.
Yet, Description Logics limits the complexity of what can be
obtained through reasoning quite severely, preventing for example
the analysis of the consequences of nonlinear interactions in
models. These limitations, inherent to the theoretical aspects of
knowledge representation and unlikely to be solved in the short
term, prevent a true ‘‘logical modelling’’ from existing: any
nonlinear model will need to be generated from the first-order
description of the systems, but no automated inference can inform
us about their likely behaviors without simulating the systems first.
A larger, if less clearly identified, problem has to do with the
lifecycle of shared knowledge and the alignment of annotated
content with meanings that shift as the knowledge evolves. First of
all, there is no ‘‘right’’ ontology for any domain; the choice of
conceptualization depends on the purposes as well as on often
rather personal viewpoints and assumptions. In fact, Goguen
(2005) refers to ontologies as theories, and highlights the need and
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587 585

the possibility of developing bridges between different ontologies
that respect the diversity of viewpoints without preventing cross-
reasoning. Unfortunately, knowledge alignment algorithms remain
to this day experimental and complex, requiring significant human
supervision, and one-to-one concept alignment is often impossible.
So the usefulness of semantic approaches to modelling for inte-
gration and harmonization can be limited by the degree of sharing
of the viewpoints expressed in the ontologies used.
Even when ontologies are agreed upon, they are likely to evolve
in time as the understanding of the natural systems changes. The
problem of keeping existing annotations and conceptualization
aligned with an evolving knowledge base is similarly complex. The
management of the lifecycle of knowledge is a highly researched
area whose future results will decide much of the practicality of
adopting semantic approaches in modelling. The current state of
the art in knowledge-driven modelling is mostly using conceptu-
alizations developed ad hoc to reflect the needs of the project, and
is by consequence relatively immune to these problems. Alignment
issues will become more significant as the adoption of semantic
annotation increases, or when knowledge-driven modelling
approaches become mainstream.
5.2.2. Adoption
At this stage, it is hard to ascertain whether the adoption of
ontology-driven approaches by the larger community will be
smooth. This factor becomes even more important in the semantic
approach, where ontologies incorporate paradigms for specifica-
tion and reasoning as opposed to the simpler role of ‘‘tagging’’
information for validation and mediation of workflows, and are
therefore bound to create more controversy.
As a consequence, acceptance problems may hamper the large-
scale usage of semantic approaches. Even when the difficulties of
developing, storing and maintaining large ontologies are success-
fully addressed, their recognition and acceptance will likely remain
difficult for some time. For now, knowledge-based approaches are
a little-understood black box in the minds of many scientists.
Knowledge-based computing will require a paradigm shift in the
way environmental scientists think about modelling, whose feasi-
bility will depend on scientists feeling that the knowledge incor-
porated in models is their knowledge. Bridging technologies are
available that allow us to merge, to a certain extent, concepts from
different ontologies into larger-scale ones. Yet, integration remains
difficult and confidence in new technology is typically low at first.
The progress of environmental modelling in the next decade will
probably clarify many aspects of the feasibility of these new
approaches.
Acknowledgements
This work is supported by the US National Science Foundation,
award DBI 0640837 (ARIES) for Villa, and by the European Union,
award 010036-2 (SEAMLESS), for Villa, Athanasiadis and Rizzoli.
Three anonymous reviewers contributed greatly to clarifying and
improving the manuscript.
References
AGROVOC, 2006. The AGROVOC multilingual dictionary of the United Nations food
and agriculture organization. www.fao.org/agrovoc/.
Ashburner, M., Ball, C.A., Blake, J.A., Botstein, D., Butler, H., Cherry, J.M., Davis, A.P.,
Dolinski, K., Dwight, S.S., Eppig, J.T., Harris, M.A., Hill, D.P., Issel-Tarver, L.,
Kasarskis, A., Lewis, S., Matese, J.C., Richardson, J.E., Ringwald, M., Rubin, G.M.,
Sherlock, G., 2000. Gene ontology: tool for the unification of biology. Nature
Genetics 25, 25–29.
Athanasiadis, I.N., 2006. An intelligent service layer upgrades environmental
information management. IT Professional 8, 34–39.
Athanasiadis, I.N., 2007. Towards a virtual enterprise architecture for the environ-
mental sector. In: Protogeros, N. (Ed.), Agent and Web Service Technologies in
Virtual Enterprises. Idea Group Inc., pp. 256–266.
Athanasiadis, I.N., Villa, F., Rizzoli, A.E., 2007a. Ontologies, JavaBeans and Relational
Databases for enabling semantic programming. In: Proceedings of the Thirty-
First IEEE Annual International Computer Software and Applications Confer-
ence (COMPSAC), vol. 2. IEEE, pp. 341–346.
Athanasiadis, I.N., Villa, F., Rizzoli, A.E., 2007b. Enabling knowledge-based software
engineering through semantic-object-relational mappings. In: Proceedings of
Third International Workshop on Semantic Web-Enabled Software Engineering,
Fourth European Semantic Web Conference, pp. 16–30.
Athanasiadis, I.N., Rizzoli, A.E., Donatelli, M., Carlini, L., 2006. Enriching software
model interfaces using ontology-based tools. In: Third Biennial Meeting of the
Int’l Environmental Modelling and Software Society. Burlington, VT, USA.
Baader, F., Calvanese, D., McGuinness, D., Nardi, D., Patel-Schneider, P. (Eds.), 2003.
The Description Logic Handbook: Theory, Implementation and Applications.
Cambridge University Press.
Baker, K., Benson, B., Henshaw, D.L., Blodgett, D., Porter, J.H., Stafford, S.G., 2000.
Evolution of a multisite network information system: the LTER information
management paradigm. Bioscience 50, 963–978.
Batzias, F.A., Siontorou, C.G., 2006. A knowledge-based approach to environmental
biomonitoring. Environmental Monitoring and Assessment 123, 167–197.
Beckett, D., 2004. RDF/XML syntax specification (revised). http://www.w3.org/TR/
2004/REC-rdf-syntax-grammar-20040210.
Bertino, E., Catania, B., Zarri, G., 2001. Intelligent Database Systems. Addison-Wes-
ley, New York.
Brilhante, V., 2005. Ecolingua: a formal ontology for data in ecology. Journal of the
Brazilian Computer Society 11 (2), 61–78.
Ceccaroni, L., Corte´s, U., Sa`nchez-Marre`, M., 2004. OntoWEDSS: augmenting envi-
ronmental decision-support systems with ontologies. Environmental Modelling
& Software 19 (9), 785–797.
Chau, K.W., 2007. An ontology-based knowledge management system for flow and
water quality modeling. Advances in Engineering Software 38, 172–181.
Chen, Z., Gangopadhyay, A., Karabatis, G., McGuire, M., Welty, C., 2007. Semantic
integration and knowledge discovery for environmental research. Journal of
Database Management 18, 43–68.
Duepmeier, C., Geiger, W., 2006. Theme park environment as an example of envi-
ronmental information systems for the public. Environmental Modelling &
Software 21, 1528–1535.
GBIF, 2004. Global biodiversity information facility. http://www.gbif.org.
GEON, 2005. GEON cyberinfrastructure for the geosciences. http://www.geongrid.
org.
Goguen, J., 2005. Information integration, databases and ontologies. http://www.cs.
ucsd.edu/~goguen/projs/data.html Available from:
Gruber, T.R., 1993. A translation approach to portable ontology specifications.
Knowledge Acquisition 5, 199–220.
Gruber, T.R., 1995. Toward principles for the design of ontologies used for knowl-
edge sharing. International Journal of Human–Computer Studies 43, 907–928.
McGuinness, D., Fox, P., Cinquini, L., Benedict, P.W.J., Garcia, J., 2007. Current and
future uses of OWL for scientific data frameworks: successes and limitations. In:
Golbreich, C., Kalyanpur, A., Parsia, B. (Eds.), Proceedings of the OWLED 2007
Workshop on OWL: Experiences and Directions, p. 258.
Guinness, D.L.M., van Harmelen, F. (Eds.), 2004. OWL Web Ontology Language
Overview. http://www.w3.org/TR/owl-features/.
ISO, 2004. ISO/TC211 Geographic Information Metadata Standard.
Jones, M.B., Berkley, C., Bojilova, J., Schildhauer, M., 2001. Managing scientific
metadata. IEEE Internet Computing 5, 59–68.
Keller, R.M., Dungan, J.L., 1999. Meta-modeling: a knowledge-based approach to facil-
itating process model construction and reuse. Ecological Modelling 119, 89–116.
Kepler, 2004. The Kepler project. http://www.kepler-project.org.
Khatri, V., Ram, S., Snodgrass, R.T., 2006. On augmenting database design-support
environments to capture the geo-spatio-temporal data semantics. Information
Systems 31, 98–133.
Kiryakov, A., Popov, B., Ognyanoff, D., Manov, D., Kirilov, A., Goranov, M., 2003.
Semantic annotation, indexing, and retrieval. Lecture Notes in Computer
Science 2870, 484–499.
Kohler, J., Lange, M., Hofestadt, R., Schulze-Kremer, S., 2000. Logical and semantic
database integration. IEEE International Symposium on Bio-Informatics and
Biomedical Engineering, 77–80.
Krivov, S., Williams, R., Villa, F., 2007. GrOWL, visual browser and editor for OWL
ontologies. Journal of Web Semantics: Services and Agents on the World Wide
Web 5 (2), 54–57.
Lee, T.B., Hendler, J., Lassila, O., 2001. The semantic web. Scientific American 284,
28–37.
Lee, S., Wang, T.D., Hashmi, N., Cummings, M.P., 2007. Bio-STEER: a semantic web
workflow tool for grid computing in the life sciences. Future Generation
Computer Systems 23, 497–509.
Ludaescher, B., Altintas, I., Berkley, C., Higgins, D., Jaeger, E., Jones, M., Lee, E.A.,
Tao, J., Zhao, Y., 2005. Scientific workflow management and the Kepler system.
Concurrency and Computation: Practice and Experience 18, 1039–1065.
Ludaescher, B., Gupta, A., Martone, M.E., 2001. Model-based mediation with domain
maps. Seventeenth International Conference on Data Engineering (ICDE) Hei-
delberg, Germany.
Madin, J., Bowers, S., Schildhauer, M., Krivov, S., Pennington, D., Villa, F., 2007. An
ontology for describing and synthesizing ecological observation data. Ecological
Informatics 2, 279–296.
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587586

Muetzelfeldt, R., 2004. Declarative modelling in ecological and environmental
research. In: European Commission Directorate-General for Research.
Muetzelfeldt, R., Massheder, J., 2003. The Simile visual modelling environment.
European Journal of Agronomy 18, 345–358.
NASA, 2006. Semantic web for earth and environmental terminology (SWEET).
http://sweet.jpl.nasa.gov.
Parr, C., Parafiynyk, A., Sachs, J., Ding, L., Dornbush, S., Finin, T., Wang, T.,
Hollender, A., 2006. Integrating ecoinformatics resources on the semantic web.
In: Proceedings of the Fifteenth International World Wide Web Conference.
ACM Press, pp. 1073–1074.
Paterson, T., Kennedy, J.B., Pullan, M.R., Cannon, A., Armstrong, K., Watson, M.F.,
Raguenaud, C., McDonald, S.M., Russell, G., 2004. A universal character model
and ontology of defined terms for taxonomic description. Lecture Notes in
Computer Science 2994, 63–78.
Pennington, D., Higgins, D., Peterson, A.T., Jones, M.B., Ludaescher, B., Bowers, S.,
2007. Ecological niche modelling using the kepler workflow system. In:
Workflows for Science: Scientific Workflows for Grids. Springer Verlag, New
York, pp. 91–108.
Raskin, R.G., Pan, M.J., 2005. Knowledge representation in the semantic web for
earth and environmental terminology (SWEET). Computers and Geosciences 31,
1119–1125.
Richmond, B., 2001. An Introduction to Systems Thinking: STELLA Software. High
Performance Systems Inc.
Rizzoli, A.E., Donatelli, M., Athanasiadis, I.N., Villa, F., Huber, D., 2008. Semantic links
in integrated modelling frameworks. Mathematics and Computers in Simula-
tion 78, 412–423.
Robertson, D., Bundy, A., Muetzelfeldt, R., Haggith, M., Uschold, M., 1991. Eco-logic:
Logic-Based Approaches to Ecological Modelling. MIT Press, Boston, MA.
Salles, P., Bredeweg, B., Araujo, S., 2006. Qualitative models about stream ecosystem
recovery: exploratory studies. Ecological Modelling 194, 80–89.
Scholten, H., Kassahun, A., Refsgaard, J.C., Kargas, T., Gavardinas, C., Beulens, A.J.M.,
2007. A methodology to support multidisciplinary model-based water
management. Environmental Modelling & Software 22 (5), 743–759.
SEAMLESS, 2005. SEAMLESS IP home page. http://www.seamless-ip.org.
SEEK, 2004. Enabling the Science Environment for Ecological Knowledge. http://
seek.ecoinformatics.org.
Tiller, M., 2001. Introduction to Physical Modelling with Modelica. Springer, New
York, NY.
Villa, F., 2001. Integratingmodelling architecture: a declarative framework for multi-
paradigm, multi-scale ecological modelling. Ecological Modelling 137, 23–42.
Villa, F., 2007. A semantic framework and software design to enable the transparent
integration, reorganization and discovery of natural systems knowledge. Jour-
nal of Intelligent Information Systems 29, 79–96.
Villa, F., Ceroni, M., Krivov, S., 2007. Intelligent databases assist transparent and sound
economic valuation of ecosystem services. EnvironmentalManagement 39, 887–899.
Villa, F., Ceroni, M., Krivov, S., Johnson Jr., G.W., Bagstad, K., Batker, D., Portela, R.,
Honzak, M., 2008. The ARIES project: artificial Intelligence for Ecosystem
Services. Ecoinformatics Collaboratory white paper, UVM. Available from:
http://esd.uvm.edu/uploads/media/ARIES.pdf.
Villa, F., Costanza, R., 2000. Design of multi-paradigm integrating modelling tools
for ecological research. Environmental Modelling & Software 15, 169–177.
Wand, Y., Storey, V.C., Weber, R., 1999. An ontological analysis of the relationship
construct in conceptual modelling. Acm Transactions on Database Systems 24,
494–528.
Webs on the Web (URL). Webs on the Web project: home page. http://www.
foodwebs.org/index_page/wow2.html.
Wenzel, V., 1992. Semantics and syntax elements of a unique calculus for modelling
of complex ecological-systems. Ecological Modelling 63, 113–131.
Williams, R.J., Martinez, N.D., Goldbeck, J., 2007. Ontologies for ecoinformatics. Web
Semantics, Science, Services and Agents on the World Wide Web 4, 237–242.
F. Villa et al. / Environmental Modelling & Software 24 (2009) 577–587 587