Expert Systems With Applications

allow advanced SPARQL querying. Although the proposal has been designed to consider remote sensing data products in general, the current study is mainly focused on the Sentinel 2 satellite mission from the Copernicus Programme of the European Space Agency (ESA). Four different use cases are showcased to check potentials of the proposed semantic model in terms of ontology integration, federated querying, data analysis and reasoning.


Introduction
The development of new sensors and the growing ease of access to data generated with remote sensing techniques are pushing datadriven research and the development of new innovative algorithms for its analysis. In this context, Earth Observation's satellite systems are continuously generating a great quantity of data, which are nowadays essential for applications in diverse areas, such as: climate change monitoring (Plummer, Lecomte, & Doherty, 2017), precision agriculture (Weiss, Jacob, & Duveiller, 2020), smart urban design (Reba & Seto, 2020), and many others. This promotes the continuous generation of data-driven software facilities oriented to advanced processing, analysis and visualization, which often offer enhanced computing capabilities. However, while promising, such new-generation applications still require the symbolic representation of scenes and objects in images, as well as the setting of threshold values of computed vegetation indexes for the generation of knowledge rule systems (Belgiu, Drǎguţ, & Strobl, 2014). In this regard, the development of knowledge-driven and linked open data (meteorological, plant phenotype, etc.). Therefore, there is a clear need of studying integration aspects of existing ontologies in the context of remote sensing (Arvor et al., 2019), as well as to show the potentials of such integration for feeding advanced analysis. To this end, the use of ontologies and web semantic technologies has shown high success in many fields. In the specific domain of remote sensing, there are several distinct attempts of ontologies, although they still constitute local prototypes that illustrate the potential in their use (Andrés, Arvor, Mougenot, Libourel, & Durieux, 2017;Arvor et al., 2019).
On the basis of this necessity, this article proposes a semantic model for the consolidation, integration, reasoning and linking of data (and meta-data), in the context of satellital remote sensing products for EO. With this objective, an OWL (Web Ontology Language) (Group, 2012) ontology has been developed and an RDF repository has been generated to allow advanced SPARQL querying. The proposed ontology, called RESEO (REmote SEnsing Ontology), has been designed to consider remote sensing data and meta-data products in general, including satellite constellations, unmanned aerial vehicles (UAVs), airborne, etc. For the sake of better understanding, the current study is mainly focused on the Sentinel 2 satellite mission of the Copernicus Programme of the European Space Agency (ESA), due to the growing popularity it is exhibiting among the scientific community since its launch in 2015 (Pahlevan, Sarkar, Franz, Balasubramanian, & He, 2017). Several use cases are showcased to check potentials of the proposed semantic model in terms of ontology integration, federated querying, data analysis and reasoning.
The main contributions of this article are summarized as follows: -A new ontology called RESEO 1 is proposed for the semantic modelling of remote sensing data and meta-data, produced in the scope of Earth observation. This ontology is developed for the first time (to the best of our knowledge) to cover multiple kinds of data products of multi/hyper-spectral images and meta-data collected from well-known satellite EO programs, UAVs, etc. RESEO is indeed linked with other existing ontologies in the field of Earth observation, as well as with ontologies devoted to meteorological open data, so an enriched knowledge framework is obtained as a result. -In terms of materialization, the proposed ontology is developed in OWL 2 and it has been linked with related external ontologies according to the same standard. Then, a series of mapping functions have been developed for data consolidation in RDF (Resource Description Framework) standard, including automatic storage in common RDF repository and Endpoint service. From this, a series of advanced SPARQL queries are set in form of API service to promote the use from the research community. -For validation purposes, a series of use cases have been worked that comprise: time series analysis, multiple satellital data product consolidation (Sentinel 2 and Landsat 8), data integration for analysis enrichment, and semantic reasoning for land-cover classification. RESEO is then shown to be useful to provide a knowledge framework for the data integration and enriched analysis, in the scope of remote sensing.
This article is organized as follows. In Section 2 background concepts and literature overview are given. Section 3 describes the semantic approach, focusing on the OWL Ontology. The procedure to validate this approach is described in Section 4. Section 5 provides discussions. Finally, Section 6 presents concluding remarks and future works. 1 Available at URL https://github.com/KhaosResearch/RESEO.

Background and related work
This section describes required background concepts about semantic web and remote sensing. A review of related articles in the specialized literature is also provided to clarify the contributions or our proposal with regards to the current state of the art.

Background concepts
-Ontology. An ontology defines a simplified representation of the world, so that it can be represented for some purpose (Gruber, 1993). Ontology languages define a set of representational primitives which are used to model a body of knowledge. The main elements of an ontology are classes (or concepts), properties (or attributes), instances (or class members) and relationships. The Web Ontology Language (OWL) is a semantic markup language used to define and publish ontologies. OWL its build on top of RDF and it is a standard by the W3C. To formalize the proposed ontology in this work, a description logic syntax OWL-DL is used as summarized in Table 1.
-RDF. Resource Description Framework is also a W3C standard for the representation of information in the Web (Schreiber & Raimond, 2014). W3C encourages the use of RDF in applications where the data are going to be processed by other applications instead of only being shown to users. RDF offers a common framework where information can be shared between applications without losing their meaning, identifying each resource by an URI (Uniform Resource Identifier). To define resources, RDF uses statements in the form of triples, which contain a subject, a predicate and an object.
-SPARQL. SPARQL is a query language for RDF graphs (Harris & Seaborne, 2013), allowing the execution of queries between several graphs in different repositories (federated queries). SPARQL queries use RDF patterns to retrieve the set of triples in the RDF repository that match.
-SWRL. This language incorporates mechanisms to identify semantic relationships between individuals (Horrocks, Patel-Schneider, Bechhofer, & Tsarkov, 2005), hence providing OWL-based ontologies with extra inference capabilities. SWRL is based on rule expressions in form of ''Antecedent ⇒ Consequent'' to represent semantic relationships. Antecedent, as well as consequent, can be formulated as conjunctions of elements, which are associated to one or more attributes defined by a question mark and a variable (e.g., ? ) in the rule. SWRL is being used to perform reasoning tasks for object classification in remote sensing imagery (Andrés et al., 2017;Gu, et al., 2017).
-Data Product. Compressed file that includes both, images from spectral bands and meta-data files needed to work with them. In Earth observation remote sensing, data products are collected from satellites, UAVs, or airborne devices, among others. Multi-spectral imagery data products generally refers to 3 to 13 bands regarding to channels, e.g., red, green, blue, near-infrared and short-wave infrared, whereas hyper-spectral imagery consists of much narrower bands (10-20 nm), with hundreds or thousands of them.
-Sentinel 2 Data Product. This kind of products are distributed at several levels of processing, in this work level 2 A with Bottom-ofatmosphere reflectance, correcting the atmosphere distortion effect on the image. Sentinel 2 products cover an area of 100 × 100 2 and have a spatial resolution of 10, 20 or 60 metres per pixel depending on the band. This satellite takes 5 days to scan the globe, so it is the frequency to revisit a scene and to generate the corresponding data product. These data are distributed through the Copernicus Open Access Hub. 2 Fig. 1 illustrates the data product folder structure of Sentinel 2, including the ranges of spatial resolutions (% reflectance and wavelength) of the twelve bands. Table 1 Basic OWL-DL semantic syntax used to formally define the proposed ontology.

Descriptions
Abstract syntax DL syntax Operators -Vegetation indices. A vegetation index (Abdou, Morin, Bonn, & Huete, 1996) is a value calculated from a set of channels (bands) from satellital sensors that quantifies the intensity of a complex phenomenon. Each channel (band) of a satellite image represents a different part of the electromagnetic spectrum, not limited to the visible light. An example of this type of indices is the Normalized Difference Vegetation Index (NDVI), which quantifies the liveliness of green vegetation in an area. It is calculated by Eq. (1), where (B8 in Sentinel 2) is the near infrared channel of an image, and (B4 in Sentinel 2) is the red channel of an image.
There exist many other indices that can be calculated from popular remote sensing satellite products (Sentinel 2, Landsat 8, MODIS, Word-View, etc.), such as: Soil Adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI), Shadow Index (SI) or Normalized Difference Water Index (NDWI). All these can be obtained for the same sensed area and with different satellites, although showing different characteristics and resolutions (depending on the specific physical features of each sensor instrument). Therefore, data product harmonization is an important task in remote sensing, since it allows to complement information, hence enhancing the image analysis (Roy, et al., 2019). The RESEO ontology proposed in this work aims at covering this issue from a knowledge-driven perspective.

Related works
As commented before, a promising research line in Earth Observation consists in the development of knowledge-based solutions (Chen, et al., 2016). This will support domain experts to perform advanced analysis where context knowledge and the interpretation of remote sensing images are involved. In this sense, mechanisms like GEOBIA allows the classification of groups of pixels that share several properties in satellite images, by analysing them using knowledge from experts (Blaschke, 2010). However, GEOBIA is still limited to objects, so a complete contextual framework is required to capture the semantics of such observations, including their relationships in different levels, to better represent them in form of knowledge base.
A first attempt in this direction is the OBOE ontology (Madin, et al., 2007), which is oriented to represent ecological observations and to map real-world geographic entities with their corresponding objects in images. OBOE includes an extensive set of unit definitions and can facilitate automatic unit conversions, but is intended as a broadly applicable ontology, missing characteristics from specific research domain. OBOE can be aligned with the O&M ontology (Cox, 2013) by means of a property of equivalence between the classes Measurement (in the former) and Observation (in the latter). O&M is an OWL ontology that follows the ISO/OGC standard for Observations, as well as for other standard geographic information schemes.
From a different perspective, the Semantic Sensor Network (SSN) ontology (Compton, et al., 2012) comprises a contextual framework to represent sensor meta-data and observations, including remote sensing. In turn, SSN can be aligned with OBOE and O&M to compose a high level integration scheme, hence considering ecological and sensors knowledge domains. In this regard, the Semantic Web for Earth and Environmental Terminology (SWEET) (Raskin & Pan, 2005) is actually a collection of OWL ontologies considering such different domains (space, time, biological realms, physical quantities, etc.) and science knowledge concepts (phenomena, reactions, chemical processes, events, etc.). SWEET has been extended in some works to cover other domains, such as hydrogeology (Tripathi & Babaie, 2008) and Earth systems sciences in general (DiGiuseppe, Pouchard, & Noy, 2014).
From an orthogonal viewpoint, web semantic technologies can be also used to discover and integrate remote sensing services, which  (2015), an ontologyenabled framework is proposed for enabling collaboration among service providers and applications to semantically discover remote sensing services. This framework combines the use of ontologies and processing workflows. Another interesting semantic model is the standard GeoSPARQL (Battle & Kolas, 2012). It supports the semantic representation and querying of geospatial data. GeoSPARQL defines a specific ontology for representing geospatial data in RDF, including an extension to the SPARQL query language for dealing with this kind of data. GeoSPARQL allows qualitative spatial reasoning and computations, so it is also suitable for being used in the context of remote sensing analysis (Viqueira, Villarroya, Mera, & Taboada, 2020).
All these semantic approaches can be extended to consider specific elements to the remote sensing domain, which lead to interpret EO images (e.g. spectral bands, indices, product's meta-data, etc.). Recent studies in this direction can be found in Andrés et al. (2017) and Gu, et al. (2017), although they are focused on the specific case of ontology-driven image classification.
The RESEO ontology aims at covering this gap, hence to improve the integration of remote sensing data and to enhance the generation of knowledge-driven approaches in this domain. Following the suggestions made in Arvor et al. (2019), RESEO can be used for modelling elements, such as: Sentinel 2, NDVI, NDVI Processing, MSI, etc., which indeed can be used as linking concepts for the alignment with other related ontologies: OBOE, SSN, and SWEET. As a summary, Table 2 shows the main features characterizing the reported ontologies, with regards to the proposed approach.

Semantic approach
One of the main objectives of RESEO is to provide an ontological framework for the semantic consolidation of the data captured by Earth Observation satellites, in a way that it can be easily extended for adding new data sources, such as different satellites, UAVs or linked open data. To this end, the proposed ontology has been defined in OWL 2, following the Ontology Development 101 (Noy & McGuinness, 2001) seven-step methodology as detailed next: (i) Determine the domain and scope of the ontology. Although general enough to consider any kind of remote sensing product, for simplicity in this study, the scope of RESEO has been limited to the attributes of the Sentinel-2 and Landsat −8 meta-data. For example, the Sentinel 2 products include platform name, orbit number, orbit direction, format, filename, data take identifier, processing level, etc. (ii) Consider reusing existing ontologies. Several existing ontologies have been used to make the proposal easier to align with others. Firstly, the OWL-Time.owl (Cox & Little, 2017) is used to describe temporal instants, and GeoSPARQL.owl (Perry & Herring, 2012) to describe geographical positions, both ontologies are W3C standards. To integrate meteorological data, RESEO is aligned to the AEMET.owl ontology (Poveda Villalón, 2011), that defines meteorological data and how it is captured. In the context of EO and sensors, the OBOE and SSN ontologies have been partially reused to consider classes related to satellital sensors and indexes. Table 2 shows the domains of all the linked ontologies. RESEO's goal is not to replace any of this standard ontologies, but to integrate all their fields to enhance the generation of knowledge-driven approach in the field of remote sensing. (iii) Enumerate important terms in the ontology. The most important concepts for RESEO are the Product, its DataSource, the Snapshot (the analysis of a Product ) and the Scene of interest for an analysis. (iv) Define classes and the class hierarchy. The key concepts defined above have been modelled as the classes of the ontology. Fig. 2 shows the most important classes of RESEO. Some of these classes comprise a generalization of a set of more concrete classes. For example, DataSource is a general concept, which is specified in a hierarchy of subclasses that define more concrete data sources for a product, like Satellite, which later has other subclasses like Sentinel 2 or Landsat 8. If a use case needs any extra data source, it can be integrated in the ontology as a subclass of Data Source. For example, the range of the hasDataSource object property is a DataSource, and the domain of hasDataSource is Product ; the range and domain of the format data property are xsd:string and Sentinel2Product, respectively. Value restrictions are used in our ontology to specify, for example, that if Sentinel 2 is connected to Sensor through the hasSensor object property, at least one sensor has to be MSI (Multi-Spectral Imagery). (vii) Create instances. Individuals are specific data that belong to a class. These instances are created by mapping the data from the data sources (Copernicus Data Hub, 3 AEMET Open Data, 4 etc.) to RDF. The mappings are done by following the model defined by the ontology. Apart from the mapping of data, some static instances are created for the satellites in study, namely: Sentinel 2 and Landsat 8. These instances are created so that all the new ones generated from the data have a place to link to. For example, if a new Sentinel 2 product is included, its data source must be an instance of DataSource and all the Sentinel 2 products will link to the same instance.

Ontology model
In

Table 3
Product: Object properties.

Object properties
Description logic although a set of new generic ones have been designed to be inherited from. This allows the ontology to be easily expanded by adding new Scene, DataSource or Product types depending on the specific end user's case of study. For simplicity, a selection of the most important classes of RESEO are detailed as follows: -Product. This class defines the data products received from the remote sensing devices, i.e., EO satellites and UAVs. Each Product has associated a timestamp given from the owl-TIME ontology of type time:GeneralDateTimeDescription. 5 The Product class is modelled in this version with two subclasses, Sentinel2Product and Landsat8Product which correspond to the data sources worked at the moment in the semantic model, although it can be easily extended with more of them. A set of main object properties of the Product class are: hasDataSource, hasDate, hasProduct, and hasScene, which are defined in description logic in Table 3. Regarding data properties, they are mostly defined for the subclasses of Product, since they cover specific aspects of the contextual use cases where they are used. For example, in the case of Sentinel2Product, data properties are referred to this particular data structure and attributes as defined in Table 4.
-Data Source. It represents a data provider for a Product. In its current state, this class has the subclass Satellite, which in turn has other two subclasses: Sentinel2 and Landsat8. However, depending on the application, it could be extended with new hierarchies of subclasses to define different sources, such as: other EO satellites (MODIS, WorldView) and UAV products. DataSource includes two main object properties, hasDataSource and isDataSourceOf, which are defined in Table 5. It is worth noting that DataSource can also be used as linking element with other external ontologies, since it could cover sources of data in general, but focused on contextualizing with the remote sensing domain of knowledge. In this regard, the subclass Satellite has the object property processedBy with range NDVIProcessing, which is in turn linked with classes Procedure and Observation from the SSN ontology. NDVIProcessing has the data property formula to define how the NDVI is calculated for each satellite. In this regard, a subclass of satellite is Sentinel2, which is connected to MSI, being this last a subclass of Sensor (also taken from SSN). (See Table 4) -Scene. A Scene defines a region of interest with a specific location in the Earth. A scene is contained inside one or more products, which could be captured from different remote sensing devices (Sentinel 2, Landsat 8, UAVs, etc.), although referring to the same specific location and preferably to similar (or close) time instants. Scene is modelled as a subclass of the geosparql:Geometry class of the GeoSPARQL ontology. 6 Object properties defined for Scene are: hasNearestStation hasScene and isSceneOf, which descriptions are detailed in Table 6. Data properties of this class are defined in geosparql:Geometry. The property hasNear-estStation is used as linking element with the AEMET.owl ontology that incorporates meteorological data. In this way, the Scene allows to integrate different sensing data referring to a specific location, for a given time period and including the specific climatic conditions, e.g., the imagery products of Sentinel 2 and Landsat 8 capturing the area of the Strait of Gibraltar, and including maximum and minimum temperatures during the first week of August.
-Snapshot. A Snapshot represents the results of an analysis of a Product, over a concrete Scene of interest. Table 7 shows all the object properties of this class, namely: hasProduct, isProductOf and isSceneOf, while Table 8 contains a selection of representative data properties. Among these properties, it is worth mentioning those referring to vegetation indexes, such as EVI or NDVI, which are computed with different combination of spectral bands, depending on the remote sensing devices involved in the specific Scene (e.g., Sentinel 2, Landsat 8, etc.). The remaining indexes (SAVI, NSDI, etc.) are defined in this class by following similar schemes of data properties as done with EVI and NDVI. In this sense, NDVI is linked with the OBOE class Characteristic 6 More information about the GeoSPARQL standard: https://www.ogc.org/ standards/geosparql.  and with the SWEET ontology by means of the property hasCharacteristic. Another interesting property is the cloud cover percentage in sensed images, which is often used as a threshold to select an specific product, or discard it, for the analysis.   Table 7 Snapshot: Object properties.
Object properties Description logic

Data consolidation
At this point, the ontological framework of RESEO is then defined, including its linking mechanisms with other existing ontologies (OBOE, AEMET, GEOSParql, etc.). This constitutes the terminological component (TBox) of the proposed semantic approach. For model materialization through the associated ABox, a series of mapping functions have been defined to convert all processed data into RDF, according to the RESEO ontological scheme. All these RDF data are then stored and consolidated in a common RDF repository, which enables an SPARQL Endpoint for data querying.
An overall representation of the proposed semantic model is illustrated in Fig. 3. In the current version, two main data sources have been used for feeding the model consisting in: (1) Sentinel 2 data products collected from the Copernicus Data Hub and (2) meteorological data from the AEMET Open Data Portal. 7 This last source is obtained from the corresponding API in JSON format, so specific mappings have been also developed to transform them into RDF according to the AEMET.owl ontology. These data are collected for a given scene that is geo-localized in the area of interest (see Section 4), for a specific time period.
As commented before, this semantic model could be easily extended to consider other kinds of remote sensing data products (Landsat, MODIS, UAVs, etc.), as well as observation and meteorological data. In this regard, the linking properties defined in RESEO leads the RDF repository to be connected with other external knowledge graphs, so the use of federated SPARQL queries will allow semi-transparent data fusion for feeding advanced applications.
All the services and material generated for the semantic model are available in a CKAN organization devoted to the Green Senti Project (University of Malaga). 8 This Linked Open Data repository contains the OWL ontologies (RESEO and AEMET), use cases datasets, images, and SPARQL Endpoint. 9

Validation
To validate the proposed semantic model, four cases of study have been developed that represent featured functionalities to be provided by knowledge-based approaches (Arvor et al., 2019). These functionalities are mainly focused on: (1) data processing from multiple satellite products to generate time series; (2) for a given scene, querying to merge data from different data products; (3) querying to fuse different kind of data, e.g. vegetation indexes and meteorology; and (4) land-cover semantic classification of remote sensing imagery based on reasoning rules.
Most of these cases have been conducted on a common scene located in the Teatinos Campus of the University of Malaga, which is a semiurbanized area on the outskirts of the city of Malaga (Spain). Fig. 4 shows the selected area of this scene, which comprises 185 hectares in the west side of the metropolitan area of Malaga, containing: green zones, buddings, parkings, sports area, roads, and lakes.

Use case 1: Time series
One of the main tasks in remote sensing analysis is the generation of time series of a set of attributes, where vegetation indexes are often arranged with the observation dates, for monitoring the evolution of a certain factor. In these time series, additional information such as, climatic conditions or topological attributes, are usually incorporated, which are indeed useful for time series forecasting. J.F. Aldana-Martín et al. In this use case, the goal is to monitor the evolution of green zones in the university campus. To this end, a procedure has been developed to calculate the surface of green zones, as the number of pixels with > 0.5 for each product and observation date. In this way, the number of green hectares have been registered from 2016 to the date. It is worth noting that, as the first Sentinel 2 product that was distributed at a 2 A level was released in march 2018, the previous period was feed with products of the 1C level, including the application of the atmosphere correction.
All these data have been mapped into RDF (following RESEO) and stored in the repository, which indeed includes meteorological data (from AEMET) related to the same time period. The integrated data can be then queried by means of the SPARQL queries on Listings 1 and 2. The resulting time series can be plotted as shown in Fig. 5, where each point identifies the number of green hectares computed for each product and for each date of observation. The regression line fitting the time series reflects the seasonality induced by the vegetative stage of plants, as well as the amount of surface identifies as green zone. The increasing tendency can be explained not only by the growth of plants, but also for the planting of new ones.

Use case 2: Merging remote sensing data from different products
Besides Sentinel 2, there are many other EO satellites in orbit and aerial imagery vehicles, all of them supplying images that can provide information about the same region of interest (scene). Therefore, an interesting option is to use not only one remote sensing source, but several of them, getting the most amount of data for a given analysis. However due to the differences of the inboard sensors, trajectories, positions, etc., these remote sensing devices generate different data products of multi-spectral bands, which also involve differences in the calculation of indices. For example, there is a smooth different between the NDVI values calculated by Sentinel 2 and Landsat 8. In addition, there are different indices that could be calculated with a given remote sensor, but not with others. This entails the need of integrating information captured by multiple sensors, for a common scene.
For instance, in order to integrate data captured by Sentinel 2 and Landsat 8, it is required to obtain data products from both satellites captured at the same date. To perform this task manually, the human expert has to visit each satellite data portal to get a list of the products and check which dates have a product available in common. This manual step could be automated by including Landsat 8 in RESEO semantic model, so a SPARQL query would return integrated data from the two satellites in study.
In this use case, after querying all the available Sentinel 2 and Landsat 8 products for the region of and date of interest, a number of 4 pairs of products matched the requirements. This is mainly due by the difference in periods for both satellites. To reach a compromise between the amount of data and the difference in time between products, the time window for accepting products was set to 1 day. This means that if a Sentinel 2 product is released a given day, any Landsat 8 products released the day after or before, will be considered too. This increases the number of products available for this study to 16 pairs, from a starting pool of 140 Landsat 8 products and 96 Sentinel 2 products. From these selected products, an RDF subgraph including pixel per pixel values for all bands from both satellites and some indices (NDVI and EVI) was integrated. An example of SPARQL query to get the NDVI values calculated from Sentinel 2 and Landsat 8 products is shown below in Q3 (Listing 3).

Use case 3: Querying for merging heterogeneous data
Semantic integration and consolidation of heterogeneous data from multiple sources is a key functionality of ontological models, since it enables to enrich the initial knowledge bases with additional variables, hence allowing advanced analysis ( Barba-González, et al., 2019).
This use case is oriented to exploit such a functionality in the context of RESEO, by merging information about the vegetation stage of the region of interest (University Campus of Teatinos) with the weather conditions registered in this region and for a given time period (Aug. 2017to Feb. 2019. To this end, SPARQL query Q4 is formulated to obtain the NDVI calculated from Sentinel 2 products together with the registered temperature, while SPARQL query Q5 selects the Moisture index (also from Sentinel 2 products) with the temperature.
With the resulting data, it is now possible to calculate the correlation between these two couples of variables. In this way, Figs. 6 and 7 show the Pearson correlation of the NDVI with regards to the average temperature (after normalizing) and the same correlation of the Moisture with regards the temperature, respectively. In both cases, there is a negative correlation between these variables, since the NDVI and the Moisture indices decrease whereas the temperature is higher. This is a typical observation in southern Spain, where high temperatures are usually accompanied by a dry environment.

Use case 4: Remote sensing pixel classification with semantic reasoning
Land-cover classification of high resolution imagery is one of the main functionalities demanded in remote sensing, since it provides a framework for the identification, monitoring and traceability of important elements appearing in such images. It is used in important applications, such as: crop-land classification, urban monitoring and water reservoirs evolution. This problem has been successfully approached with two main strategies (Belgiu & Csillik, 2018;Weih & Riggan, 2010) by the remote sensing community, namely: object-based and pixel-based classification. Object-based classification has been recently tackled with semantic reasoning in some works (Andrés et al., 2017;Gu, et al., 2017) with success, although pixel-based semantic classification still remains an alternative to be checked (to the best of our knowledge).
In this regard, this use case is focused on performing pixel classification in Sentinel 2 products by means of a series of semantic reasoning tasks with SWRL rules under the knowledge-base of RESEO. Therefore, a set of rules have been constructed from a previous labelling process, where a series of thresholds were identified (on bands and NDVI) by training a decision tree on several products of the same scenario, spread across a year. These thresholds have been obtained with the ArcGIS 10 tool for discriminating the values of NDVI and Bands spectra, hence to separate the ground information of the study area into different land covers. In addition, colour mapping and class labelling were done to complete the classification process.
As a result, Table 9 contains the identified thresholds on pixel observation attributes (Bands and NDVI), together with the labelled classes, namely: Water, Bare Soil, Vegetation and Building. Column at right contains an identifier (from R1 to R11) to define the corresponding SWRL rule associated with each decision path in the classification procedure. Examples of three of these representative rules are R1, R5 and R11, which are defined in Listings 6, 7 and 8, respectively. These SWRL definitions show a common structure with a first block in the  antecedent of semantic element declarations, a second block (also in antecedent) of conditional definitions (with numeric thresholds defined in Table 9), and the labelled class in the consequent. For simplicity, the J.F. Aldana-Martín et al.  These SWRL rules are then incorporated to RESEO and used to classify a test set on the selected image products, by means of reasoning tasks with Pellet reasoner (Sirin, Parsia, Grau, Kalyanpur, & Katz, 2007). 12 The confusion matrix obtained from this semantic classification is shown in Fig. 8, resulting a global prediction accuracy of 92%. More in detail, Table 10 contains, for each class and in general, the precision, recall, f1-score and support registered for this classification. Taking into account the high number of classified pixels (close to 3.5e+6), the resulting precision and recall are in general successful, as only for class ''building'' there is a low recall, probably due to class imbalance in the whole training dataset for this label.

Discussions
The development of knowledge-driven approaches represents an active research line in the remote sensing community (Arvor et al., 2019), since it offers potentials enough to provide human experts with domain knowledge representation, support for data standardization and semantic integration of sources. These functionalities are highly valuable for the creation of advanced on-top applications.
In this sense, the capacity of ontologies to offer formal framework for symbolic representation entails a new component in the associated artificial intelligence processes, since they allow to make the expert knowledge explicit (in such processes), hence contributing to the interpretation of results. This is of especial importance in the field of remote 11 Available in URL https://opendata.khaos.uma.es/organization/greensenti. 12 Available at URLhttps://www.w3.org/2001/sw/wiki/Pellet. sensing imagery for Earth observation (Arvor, Durieux, Andrés, & Laporte, 2013), where research and industry communities are expending considerable efforts towards the knowledge extraction from sensed data and the interpretation of image products. RESEO aims at collaborating in this direction by offering an integration semantic model, as well as by incorporating mechanisms for the symbolic representation and standard access of data, in the specific field of remote sensing. In terms of integration, a key aspect in RESEO design is the possibility of alignment with other interdisciplinary domains, such as: ecology, agriculture, urbanism, geology, etc, which indeed will contribute to the generation of new extended versions of this ontology.
With regards to symbolic representation and standard access, the proposed semantic model allows the generation of well-defined APIs and services, with the capacity of federated querying and data enrichment with foreign variables, such as: meteorologic measures, plant phenotype features or soil geo-morphology tastings. These features incorporate added value in current analysis, as worked in the context of the research project Green-Senti 2019 PP Smart Campus of the University of Malaga. This project initiated the design and development of RE-SEO, including the use cases described in Section 4 for demonstration purposes.
Finally, concerning societal and economic implications, the European Commission identified data standardization and harmonization in Earth observation as one of the main challenges to be tackled in this field (BDVA, 2017), since they constitute central aspects in the data value chain, from data acquisition and storage, to data usage for supporting in business decision-making. In fact, the main players in this field are currently large companies with capacity to access major infrastructures, although new opportunities are appearing for SMEs in respond to the emerging demand of remote sensing services. Initiatives like the Green-Senti project and the RESEO ontology are focused on contributing in this direction.

Conclusions
In this work, the RESEO ontology is proposed for the semantic modelling of remote sensing data and meta-data, in the scope of Earth observation. This ontology is conceived to cover multiple kinds of data products of remote sensing imagery and their associated meta-data. RESEO is indeed linked with other existing ontologies in the field of Earth observation, as well as with ontologies devoted to meteorological open data, so an enriched knowledge framework is obtained as a result.
The proposed ontology is developed in OWL 2 and it has been linked with related external ontologies according to the same standard (OBOE, SSN, TIME-OWL, AEMET, and GeoSPARQL). A series of mapping functions have been developed for data consolidation in RDF, including automatic storage in common repository and Endpoint service. From this, a series of advanced SPARQL queries are set in form of API service to promote the use from the research community.
On top of this semantic model, a series of pilot applications have been generated in form of use cases on a selected area in the campus university of Malaga (Spain). These use cases consist of: time series analysis for environmental monitoring, multiple satellital data product consolidation (Sentinel 2 versus Landsat 8), data integration for analysis enrichment, and semantic reasoning for land-cover classification. RESEO has been shown to be useful to provide a knowledge framework for the data integration and enriched analysis, in the scope of remote sensing.
As future work, more data sources will be integrated in the ontology, including more Earth observation satellites like MODIS or Proba-V, and others types of data products involving hyper-spectral imagery from UAVs. In addition, the integration and linkage of other ontologies is a future task, since it will allow scaling the use cases to more complex scenarios, such as global climate change, bio-habitats conservation, forest monitoring, etc.