Postharvest diseases are a common problem in the cultivation of apples. Since determining the exact disease is often difficult and time-consuming, an ontology-based decision support system is being developed to support this diagnosis process. In this paper a reference ontology is presented that captures the lifecycle of apples from the orchard to the storage phase. This ontology focuses mainly on postharvest diseases including their symptoms, characteristics, and other important factors. This reference model is represented using the ontology-driven conceptual modeling language OntoUML and it is then encoded as an OWL specification enriched with SWRL rules. The latter will then serve as a basis for supporting automated reasoning in the aforementioned decision support system.
The cultivation and export of apples is a big economic factor for many countries around
the world with the US, China, and Italy leading the list. In 2014, such exports amounted
to a total value of 7.5 trillion dollars [
To address this problem, the decision support system DSSApple [
The remainder of this paper is organized as follows: First, section 2 describes the approach employed to develop our reference ontology. In that section, we discuss both the origin of our disease data, as well as adopted approach for engineering the ontology. Section 3 presents the main contribution of this paper, i.e., the DSSApple Ontology. Section 4, presents a preliminary evaluation of our model. Finally, section 5 discusses some final considerations and summarizes the work presented in this paper.
We started the project by collecting and categorizing the necessary apple disease-related
information2. Most of the data was provided by the FrudiStor project - a web
application that lists more than 40 different post-harvest diseases and physiological disorders
of apples [
The data we used from the Frudistor application mainly consisted of plain text descriptions of the symptoms of diseases and their development progress. Similar to human diseases, the characteristics of a fruit disease at an early stage may be different from those at a later moment, which needs to be taken into account when determining the exact disorder. Typical symptoms are: spots on the surface of the fruit, different types of rots, discolorations, development of spores, etc. Further information includes the life cycle of diseases (the points in time when fruits are infected and when symptoms become visible), susceptible varieties, possible causes, and similar diseases that can be easily confused.
For the development of the ontology presented in section 3, we have considered the data
gathered about the domain (see section 2.1), the expert opinion of domain experts, as
well as existing ontologies in the field [
Despite considering a number of ontological choices put forth by these existing ontologies, reusing their structure directly posed some significant difficulties. Besides not covering all aspects that are needed for supporting our decision-support system, the most salient problem with these existing ontologies is that frequently they conflate multiple (albeit independent) phenomena dealing with plant diseases. This tends to result in an combinatorial proliferation of classes (e.g.,). To avoid this problem, we decided to model 2Note that unless specified otherwise, this paper may use the terms disease, disorder, and defect interchangeably for a better reading. the scope defined by some of these existing models but to evolve and organize it as a new full-blown reference ontology addressing specific aspects of the apple pathology domain. In particular, by applying the methodological approach of separation of concerns, we organize our ontology in 5 modules, each one addressing a different type of phenomenon in reality (see section 3).
As a second methodological principle, we separated the conceptual modeling phase
of the engineering process from the codification (or implementation) phase [
In the sequel, we elaborate on the 5 modules constituting our reference ontology: apple characterization, apple handling, apple pathologies, apple dispositional factors, and apple representation.
This module addresses the apple itself, its possible types, its parts as well as its qualities and features.
An Apple is composed of a number of (functional) Apple Parts [
Apples and their parts (here termed Apple Substances) can be characterized by
Apple Aspects, including Apple Qualities or Apple Features. The former
includes Taste, Color, but also Texture, Odor (not in the diagram), among others. In
UFO and, hence, in OntoUML, qualities are full-fledged endurants [
As with all aspects [
There are certain qualities that characterize the fruit as a whole. In particular, the Apple Nutritional Factor. For this reason, for example, a redefined relation of characterization is created between Apple Nutritional Factor and Apple. Finally, since aspects are endurants in UFO, they can bear their own aspects, which, in turn, can change in time. For instance, a hole can have a diameter, which can vary in time.
This second module addresses the parts of an apple handling activities as the characteristics of these handling processes may have a significant influence on the health of whole apple batches.
Apple Handling activities here include Apple Harvest and Apple Storage
activities. Since all apple handling activities are subtypes of Event, they occur in a
particular time interval (delimited by time points). As discussed in [18], from these time
points, we can infer the derived relations between events. Here, use the a historical
dependence relation (see [
In all Apple Handling Activities, we have the participation of an Apple Batch, which is a collective whose members are individual Apples (see the member of relation in [19]). In Apple Storage, we also have the participation of a Storage Facility, which is a role played by a Building.
Storage facilities and spatial regions are Spatial Entities. All spatial entities can be characterized by Physical Qualities, such as Humidity, Gas Concentration and Temperature. Since we are interested in how these qualities change their values through time, we, once more, employ here one of the truthmaking patterns proposed in [20]. In this way, we explicitly represent the Physical Quality States of Physical Qualities. Each of these states (which are events and, hence, occur in a particular time interval) can be mapped to a value in a suitable quality space. Moreover, we define a particular type of Physical Quality State called a Physical Quality State of Spatial Entity During Apple Handing. As explicit in its name, this is a quality state of either a store facility involved in an Apple Storage activity or the quality state of a spatial region on which an Apple Handing Activity occurs. As such, the relation of occurs during represented in this model implies the Allen relation of temporal mereological overlapping between the two events it relates [18].
By doing this, we can capture in different times points, for example, the temperature, humidity and gas concentration of particular facilities or regions during storage and harvest, respectively. Moreover, since both Physical Quality States as well as Apple Handling Activities are events, which occupy a certain spatial region, we can infer, for example, the origin of the fruits in a batch, the age of the fruits at the moment they enter the storage facility, storage duration, etc. Different storage conditions can favour the spread of a disease, but in some cases they may also heal certain symptoms. We address these aspects in the next module of this ontology.
Characteristics of these handling processes may have a significant influence on the health of whole apple batches. Handling processes include the actions during harvest as well as the storage characteristics. Main parameters here are the storage temperature, humidity, gas concentration, the storage duration, and the age of the fruits at the moment they enter the storage facility. Different storage conditions can favour the spread of a disease, but in some cases they may also heal certain symptoms. These concepts are represented in the second module. Additional classes connected to the harvest allow to draw conclusions about the origin of the fruits.
This third module addresses the core of this ontology, dealing with Apple Pathologies
and their manifestations (Apple Pathology Manifestation). In this model, an
Apple Substance can eventually be in a Pathological Apple Substance phase
(see semantics of phase in [
Again, as it is always the case with dispositions [21], when manifested, an Apple Pathology is manifested through the occurrence of an event, which in this case it is instances of Apple Pathology Manifestation. This even, in turn, is composed of a number of sub-events termed Apple Symptom Manifestation. A symptom manifestation is considered here to be an aggregation of states of particular relevant qualities (Apple Quality State) of the Apple Substance (the apple itself and its parts). Symptom manifestations are instances of Apple Symptom Type, which are correlated with particular types of Apple Pathologies (Apple Pathology Type).
The fourth Module represents pathology enabling factors Apple Pathology Enabling Factor (sometimes called dispositional factors). These are factors that are associated with Apple Pathology Types in the sense that they can enable the onset of apple pathologies3, but also in the sense that they can activate the manifestation of these pathologies (see discussion on mutual activation partners for dispositions in [21]). Apple Pathology Enabling Factor is a cross-categorical types since its instances can be features (Apple Features such as Insect Sting in Apple or Pressure Mark in Apple), Apple Qualities (e.g., Apple Low Nutritional Factor - a contingent phase of a particular Apple Nutritional Factor quality), or even events such as Inadequate Physical Quality State of Spatial Entity During Apple Handling (e.g., a storage activity having a storage facility with an inadequate temperature, humidity or gas concentration).
This final module addresses the image processing part of the decision support system. To support the operation the system is working, our model needs to be able to relate prototypical pictures associated with symptom types (Prototypical Apple 3For a discussion on how enabling factors can activate the manifestation of vulnerabilities, the reader can refer to [23]. Symptoms Photograph), as well as prototypical pictures associated with types of (Prototypical Apple Pathogen Sign Photograph). The model also recognizes the existence of photographs of particular symptom manifestations (Apple Symptoms Photograph) of particular apples, as well as of pathogen signs (Apple Pathogen Infection Photograph). Photographs of these two latter kinds can deliberately chosen to play the role of prototypes, i.e., Prototypical Apple Symptoms Photograph and Prototypical Apple Pathogen Sign Photograph are roles played by photographs of individual apples
In order to support reasoning with actual disease instances according to the descriptions in the specialist literature, we implemented parts of the model in section 3 as an ontology in the OWL language enriched with SWRL rules [24].
This approach not only gave us the required flexibility to represent the myriad of possible combinations of aspects and aspect values that can be used to characterize specific pathologies. So, the disease data in our model would not just include information like “there are spots on the skin of the apple”, but rather provide detailed descriptions like “the spots on the apple are brown with irregular shapes and diffuse margins”, which may increase the reasoning power of a decision support system. As a concrete example, a description of Alternaria Rot in FrudiStor amounts to “Early infestation usually occurs in the calyx area of the fruit. Typically, there are small (up to 0.5 mm large) dark brown or black lenticel spots, which are often surrounded by a brown ring.”. A description like this can be described as a conjunctive SWRL rule using the predicates of our proposed ontology, which, in this case, refers to pathogen signs, apple parts, features, and qualities of features having specific values.
In order to get a first experience on how the decision support system would work, we developed a simple diagnostic system for a subset of 5 diseases (Alternaria Rot, Bitter Rot, Botrytis, Penicillium Rot, and Mucor Rot). Through an online survey [24], a nonexpert user was given 15 images of bad apples, from which each one was infected by one of the five diseases. The user had to specify which kinds of symptoms occurred (rots, spots, lesions, etc.), on which parts of the fruit they appeared, and finally describe the characteristics of the symptoms in more detail (including color, shape, texture, etc.). In order to evaluate the prediction accuracy, we split each rule into multiple simpler SWRL rules, where - for simplicity sake - a single rule only contained a single symptom of a disease, which resulted in 8-19 one-symptom rules per disease. This approach follows the recommendation for preference-based reasoning with soft rules or constraints as outlined in [25]. Requiring the conjunction of all symptoms as a precondition for deriving a disease would actually lead to empty result sets in all cases. However, when ranking the diseases for each of the 15 test cases based on the share of identified symptoms by the total number of symptoms of the respective disorder, i.e. assuming equal weights for each rule and ranking items like described in [25], the experiment revealed the results presented in next section.
In most cases and for a given disease the user identified about 50% of the symptoms as described in the literature. Although this number may appear rather small at first, it has to be taken into account that tests have been conducted only with images of apples and not with real infected fruits. Since it is very unlikely that all possible symptoms of a disease will be present on a single image, identifying 100% of all symptoms is extremely unlikely.
In addition, several symptoms have been wrongly identified (false positives). Furthermore, several diseases have some symptoms in common, which also leads to high scores for multiple diseases.
As a measure for the predictive accuracy of our results, we report Precision at 1,
which means that based on the identified symptoms the correct disease was most highly
ranked. Top-1 Precision: actually reached 0.533% (8 out of 15). In contrast, in an earlier
user study in which randomly selected images of diseased apples were presented to users,
who clicked on these images based on the perceived similarity with their target image
[
In this paper we propose a reference ontology to model the life cycle of apples, so that it can be used in a decision support system to support the diagnosis of postharvest defects. Reusing existing ontologies as base models showed limited results, as they provided a useful base structure, but were lacking the expressivity to represent more complex relations. To handle this complexity, we designed our new model in OntoUML and divided it into 5 modules, where each describes a specific aspect of an apple’s life cycle, while focusing on the disease-related aspects. To be able to include all information from the specialist literature in our ontology, we added SWRL rules that allowed us to model the symptom descriptions very precisely. Finally, first tests of these rules in a simple diagnostic system already showed some promising results for decision support of users.