Ontology-based discovery of time-series data sources for landslide early warning system

Multi-hazard refers to a collection of multiple major hazards that a country faces [30]. There is a possibility that several hazardous events occur simultaneously and are interrelated. Tropical storms, for example, is one of the most common environmental hazards (in the tropics), which can trigger multiple hazards such as heavy rainfall that in turn can induce flash flooding. Furthermore, heavy rain and flooding increase the moisture content of soil in a mountainous area and this may induce landslide. To minimise the loss of life and property damage from these inter-related hazards a comprehensive strategy for hazard management is required. In general, a strategy for hazard management is comprised of four phases [32]: (i) mitigation the actions to minimise the cause and impact of hazards and prevent them from developing into full-blown disaster; (ii) preparedness the action plans and educational activities for communities to confront with unpreventable hazard events; (iii) response the actions for emergency situations to protect people life and properties during hazard or disaster events; and (iv) recovery the actions to restore damaged properties and communitys infrastructures and to cure people from their illnesses. These four phases demand supporting tools and technologies to enhance the effectiveness of hazard management.

Several modern multi-hazard early warning systems take advantage of the data explosion on the social media. Authors in [29] proposes using a twitter data analysis framework for identifying tweets that are relevant to a particular type of disaster (e.g. earthquake, flood, and wildfire). Several techniques, including matching-based and learning-based, to identify relevant tweets are also evaluated. The work in [14] studies the potential of using social media data to identify peatland fires and haze events in Sumatra Island, Indonesia. A data classification algorithm is used to analyse the tweets and the results are verified by using hotspot and air quality data from NASA satellite imagery. A data classification algorithm is also used in [26] to automatically classify tweets and text messages (from Ushahidi crowdsourcing application) generated during the Haiti earthquake in 2010. The goal of their work is to provide an information infrastructure for timely delivery of appropriately classified messages to the appropriate responsible departments. Work in [12] proposed a decision support system that integrates crowd sourcing information with Wireless Sensor Networks (WSN) to improve the coverage of monitoring area in flood risk management in Brazil. This research introduces the Open Geospatial Consortium (OGC) standards to facilitate the data integration among crowd sourcing information and WSN.

Earth Observation (EO) and urban data provided by multiple data sources are accessible by different methods ranging from direct download to various standard Web Services APIs (e.g. Web Map Services, Web Feature Services, Sensor Observation Services, RESTful API, SOAP-based API, etc.). In addition, there are heterogeneities among EO and urban data provided by different data sources [7] including: (i) syntactic heterogeneity the difference in data format or data model for presenting datasets (e.g. plain text, CSV, Excel, XML, JSON, O&M, SensorML, etc.); (ii) structural heterogeneity the difference in data schema for describing the same types of datasets (e.g. describing soil moisture using different XML Schemas); and (iii) semantic heterogeneity difference in meaning or context of the content in datasets. These heterogeneities reveal the challenging problems brought forth by the high variety data availability in multi-hazard applications. Semantic Web Technologies have thus played a significant role by providing languages and tools for modelling domains including describing the concept and relationship among the data and hazardous events. According to W3C definition [35, 36], the Semantic Web is a web of data that provides a common framework for data sharing and reuse across applications, enterprises, and communities.

Ontology, a key element of the Semantic Web, is a specification of a conceptual model for describing knowledge about a domain of interest. A basic concept in a form of ontology can be described by an Resource Description Framework (RDF) triple [33] which is comprised of a subject, a predicate and an object. Concepts described by RDF can be extended by Web Ontology Language (OWL) [34] to construct an ontology for representing rich and complex knowledge about things. In the case of multi-hazards application, an ontology can be used to: (i) represent domain knowledge through concepts, their attributes and relationships between data sources, data and hazards; and (ii) facilitate data integration across multiple data sources that represent varieties, velocity and volume characteristics of Big Data.

Ontologies are widely used in hazard management to model knowledge about hazards and use it to manage actual data derived from EO and urban sources. Hazard assessment and urbanisation analysis are two of the common application areas where ontologies are used. The Semantic Sensor Network Ontology (SSN) [22] and the Semantic Web for Earth and Environmental Terminology (SWEET) [25] are two significant ontologies that are commonly applied for hazard management. Authors in [22] reuse SWEET to conceptualize knowledge and expertise of several areas, such as buried assets (e.g. pipes and cables), soil, roads, the natural environment and human activities. Additionally, the Ontology of Soil Properties and Process (OSP) is proposed in their work to describe a concept of soil properties (e.g. soil strength) and process of soil (e.g. soil compaction). The OSP and other concepts are used to express how they affect each other in asset maintenance activities. Furthermore, [2] and [22] present the application of SSN for wind monitoring. The first work uses SSN with Ontology for Kinds and Units (QU) [18] to conceptualise wind properties (e.g. wind speed and direction) while the later uses SSN and SWEET to model the concepts of wind sensors and data streams of wind observations. The Landslides ontology [5] extends SSN to organized knowledge for the landslides domain such as the concepts of landslides, earthquake, geographical units, soil, precipitation and wind. Even though these ontologies provide comprehensive concepts for sensor data and hazard event, and provide a reusable, widely used semantic underpinning, they do not cover conceptual aspects on human sensors (e.g. social media data). Hence, currently additional processes are required when applying these ontologies to EWS for multi-hazard application.

The related literature in the context of multi-hazard management can be classified based on the following three perspectives, data sources, hazardous event analytics, and EO and urban time series data management. It can be seen that effective multi-hazard management demands high quality and rich data from vast amount of data sources that are related to the hazard of interest. Data sources utilized by multi-hazard management applications can be any sensors and/or data services that provide EO and urban data. Such data sources include physical sensor (e.g. remote sensing, in situ sensor, wireless sensor network) and human sensor (e.g. social media, blogs and crowd sourcing). Recent data analytics research for multi-hazard management focused on hazardous event analysis, which are conducted into three main directions, event identification, event verification, and event prediction. These research reveal the challenging problems in the EO and urban time series data management, especially the discovery of potential time series data sources over the complexity and high variety of such data sources in multi-hazard management applications. Ontology is a common method for not only modeling knowledge about hazard but also managing EO and urban data. Recent work around developing the ontology in this domain are classified as standardizing ontology and reusing ontology. They have shown that current standard ontologies for data sources discovery do not exist. In addition, existing applications of ontology in this domain mostly investigate specific problems, in other words these approaches are not generalized. They fail to model the relationship between data sources and the domain knowledge which is an important factor for efficient data integration and data sources discovery.

The Landslip scenario is co-created with domain experts who are members of Landslip. In addition, the Landslip is an NERC funded project involving the analysis of observation data and social media data provided by several institutions to contribute to the reduction of landslide impacts in the risk area of India. The scenario focuses on the preparedness phase of disaster management where warning signs of landslide from social media and observation data are detected before the occurrence of landslides. The landslide warning sign is an incident that indicates the potential of landslide hazard. It can be observed by human or an early warning system. The examples of landslide warning signs are the physical changes of utilities or infrastructure (e.g. blocked road, Leaning telephone poles, retaining walls or fences), a movement of soil from foundation, and a change of color in a river. In addition, a warning sign observed by people who are unknowledgeable about landslide hazard (e.g. social media users) requires additional process to verify the potential of landslide. Designing the scenario, historical event of landslide and domain experts experiences are considered. Figure 1 illustrates a situation before landslide that happens in a place located in an urban area. This area encompasses both natural environment (e.g. river, and mountain) and built environment (e.g. schools, hospitals, road, water supply and electricity). Locating in a high slope area, the place is prone to landslide and is monitoring by the National Disaster Management Authority (NDMA). Here, an expert from NDMA analyses satellite images to detect warning signs and informs decision makers for the potential landslide hazard. Meanwhile, person A who lives in the place is enjoying his leisure time by walking around his house. Living in the landslide prone area, it raises his awareness on possibility of landslide hazard. Then he is keen to contribute to his community by reporting any incident observed in his daily life via his social media account. While walking around his place, he has observed a leaning pole nearby his house. Thus, he takes a photo of the leaning pole and reports the incident to social network using his Twitter account. He also gives additional information such as observed time and place to the tweet. Besides, person B who lives in a nearby area has observed that the color of tab water in his house become brown. He thus uses hist Facebook account to report this incident. These messages from social media are collected by NDMA Early Warning System (EWS) to detect landslide warning signs. Here, an NDMA’s staff who is a member of decision making team receives a notification message from EWS with regard to the leaning telephone pole. The incident is considered as a warning sign for landslide hazard. Receiving such information from social media users, the decision maker needs to verify the information before making further decision. Base on this, the decision maker who is a domain expert in landslide hazard risk assessment performs data analysis using an appropriate landslide analytical model. This process requires adequate historical events of landslide, EO and urban data provided by various data sources to get more accuracy of the data analysis. Hence, the decision maker searches for potential data sources from the Data Sources Discovery Service (DS) and gathers EO and urban data from the discovered data sources. The gathered data is then used to verify the social media information. Furthermore, the decision maker utilizes the EWS to assess the risk and impact of the landslide event using EO and urban data and take timely actions against the event. An example of such actions is disseminating actionable warning information to people in the landslide prone area.

The above scenario reveals the essential role of data-driven early warning system for landslide hazard management which comprises of 5 main components.

The Data-driven early EWS analyses landslide-related data to enable dynamic and timely decision making against landslide hazard. Such data includes historical landslide events, historical and real-time observation data generated by physical sensors, and social media data. In addition, a number of sensor devices have been deployed in the landslide hazard prone area by organizations who are in charge of landslide hazard management. the organizations collect EO and urban data from their sensors to monitor landslide hazard events in real-time. Besides, the collected data is stored in their local repositories and is provided as data sources to their co-ordinated organizations for further analysis. Here, the data sources metadata is published to a Data Sources Discovery Services (DS) which is an application of Decision Support System. The DS enables data publishers to advertise their data sources by registering data sources metadata to a metadata registry service. Moreover, It allows data consumers to search for their potential data sources to be used in their applications.

The purpose of Landslip Common Ontology is to provide a conceptual knowledge model of landslide domain. In addition, the Common Ontology is a combination of theoretical knowledge and human experiences to identify warning sign before landslide. Basically, landslide is one of the most significant multihazards which can be found in many places around the globe [11]. Such hazard has interactions or can be triggered by another hazards [8]. Based on this, the Landslip Common Ontology conceptualizes knowledges of landslide and its interaction with other multi-hazards [8, 9] and knowledge of warning signs that can be observed by human and use such knowledge to indicate landslide event before the occurrence of landslide. The knowledge defined in the ontology can be used to facilitate landslide early warning based on warning signs observed and reported by people in social network. Figure 3a illustrates the concept of the Landslip Common Ontology which comprises of four main concepts as follow:

EO and urban data observed by sensors indicate events or changes of landslide phenomena [24]. Such data (e.g. rain, temperature, soil moisture. etc.) from a variety of data sources is collected by data provider and provides for stakeholders to be used in their landslide hazard applications [23]. Due to the high variety and geographically distributed nature of OE and urban data sources, effective data sources discovery [37] is thus required in order to provide sufficient amount and quality of data for landslide hazard risk assessment. Landslip Data Sources Ontology is developed to enable semantically discovery of data sources. In addition, the ontology describes concepts and relationships of EO and urban data, data sources, sensor devices, and data providers. With the combination of this ontology and Landslip Common Ontology, data sources discovery mechanism utilizes knowledge of landslide hazards to discover data sources which are related to the hazard of interest. Specifically, the knowledge of landslide warning sign can be used to identify appropriate observed properties and data sources for landslide precursor verifications. This capability enable EWS to provide dynamic and timely decision making against landslide hazards. Figure 4b illustrates the Landslip Data Sources Ontology which comprises of three main concepts and reuse SSN Ontology [22] and OGC standard [19‐21] for the concepts of observation and sensors.

Table 1 shows a summary of the ontological features of Landslip Ontology in terms of size (number of classes, properties, and individuals), expressivity, and complexity of the core knowledge captured by axioms.