Managing Heterogeneous Datasets for Dynamic Risk Analysis of Large-Scale Infrastructures

Abstract

Risk assessment and management are some of the major tasks of urban power-grid management. The growing amount of data from, e.g., prediction systems, sensors, and satellites has enabled access to numerous datasets originating from a diversity of heterogeneous data sources. While these advancements are of great importance for more accurate and trustable risk analyses, there is no guidance on selecting the best information available for power-grid risk analysis. This paper addresses this gap on the basis of existing standards in risk assessment. The key contributions of this research are twofold. First, it proposes a method for reinforcing data-related risk analysis steps. The use of this method ensures that risk analysts will methodically identify and assess the available data for informing the risk analysis key parameters. Second, it develops a method (named the three-phases method) based on metrology for selecting the best datasets according to their informative potential. The method, thus, formalizes, in a traceable and reproducible manner, the process for choosing one dataset to inform a parameter in detriment of another, which can lead to more accurate risk analyses. The method is applied to a case study of vegetation-related risk analysis in power grids, a common challenge faced by power-grid operators. The application demonstrates that a dataset originating from an initially less valued data source may be preferred to a dataset originating from a higher-ranked data source, the content of which is outdated or of too low quality. The results confirm that the method enables a dynamic optimization of dataset selection upfront of any risk analysis, supporting the application of dynamic risk analyses in real-case scenarios.

1. Introduction

Electric energy plays a crucial role in today’s society, and it is involved in almost all aspects of society’s daily routine [1]. The continuous development of the economy increases the need for energy, leading to larger-scale power systems and increasingly complex structures [2]. Furthermore, the scale and complexity of power grids are expected to increase with the growing use of renewable energy sources [3], as well as the development and implementation of smart grids [4]. As numerous businesses, public infrastructures, and private households rely on the provision of power for their daily tasks, companies in charge of the power supply need to provide energy management in a more reliable, effective, and secure way [1,5].

Power grids are exposed to a plurality of hazards such as hurricanes, earthquakes, ice storms, and floods, which can have severe consequences. The increasing frequency of extreme weather events increases the damage potential of these hazards, further highlighting the vulnerability of power systems [6]. Indeed, large-scale power outages have frequently occurred in recent years and have caused substantial economic losses [2].

Risk assessment and management have received significant attention as a tool to ensure the operational safety and reliability of power systems, becoming one of the major tasks for urban power-grid management [2]. Risk assessment of power grids generally makes use of traditional risk approaches such as reliability block diagram (RBD), fault tree analysis FTA [7], or failure mode and effect analysis [1]. Yet, the complexity of the power grids and the growing amount of data coming, e.g., from prediction systems favor the development and use of more advanced probabilistic risk-based approaches [6]. Applications of data-based approaches to power grids and energy systems range from energy management for smart buildings [8] to online fault diagnosis [9], among others (e.g., [10,11]). Risk analysis of power grids susceptible to vegetation-related hazards can adopt diverse data sources, ranging from satellite-based orthophotos to drone-based aerial images, including plane-based orthophotos or LiDAR 3D point clouds. Connected devices and access to more computing power provide additional opportunities for data-based, dynamically updated risk analyses. However, an updated and accurate risk analysis is highly dependent on the data used to inform the different parameters for calculating risk, e.g., the frequency of an event, the probability of failure, and the potential consequences of this event. Indeed, the use of different datasets for failure frequencies is an important source of uncertainty in risk analysis results [12].

In addition to better informing conventional risk analysis, multiple data sources present an opportunity for dynamic risk analysis (DRA). DRA is a concept that has mostly emerged over the last decade [13,14]. It aims to build on data availability to provide more frequent and performant risk pictures of infrastructures [15]. While DRA can benefit from the growing data source variability to diversify the possibilities of information acquisition relative to a particular parameter [4,16,17], numerous data sources may also increase data collection and processing complexity [4,18]. First, the data to process are intrinsically heterogeneous, requiring a large panel of competencies to manipulate and extract relevant information from the datasets. Second, a larger number of data sources requires selection rules for decision-making optimization, given the potential variability in the data quality. This variability can be due to, for instance, the type of considered datasets, the spatiotemporal resolution of the data, or the acquisition conditions of the datasets.

The International Standard Organization states that risk assessment should use the “best information available” and the implementation of “dynamic” approaches [19]. However, there is no guidance for applying those principles when multiple data sources are available. The present paper is a step toward closing this gap. We propose an approach for the dynamic optimization of dataset management to reduce uncertainties relative to data selection upfront of any risk calculation. The proposed method (called the three-phases method) is based on metrology concepts and metadata for characterizing the parameter-related information needed for a quantitative risk analysis (QRA).

The method focuses on three main features of the datasets impacting the quality and usability of the data for a QRA: the nature of a dataset, the discrepancies observed between the spatiotemporal attributes of the dataset and the spatiotemporal requirements for the risk analysis, and the agents and factors involved in the data management. The method integrates these three factors in a scoring system using meta-features, relying solely on metadata. The result is a ranking of the datasets, based on their informative potential relative to a baseline of “perfect information”. The method also predicts the informative potential of any new dataset originating from a list of preselected data sources using only the information available in the metadata, thus without factually analyzing the content of the datasets. Hence, the method’s application allows a continuous selection of the best candidate across all available datasets. While the implementation of the method is labor-intensive in the first iterations of the process, it can significantly increase data management efficiency in future assessments in the long term, avoiding suboptimal repetition of tasks.

The application of the method is demonstrated through a case study focusing on risk management in power grids. We focus on the role of vegetation along power lines, which represents a common source of outages in power grids, either via trees falling on the power lines or by growing under the infrastructure until grounding one phase [4,5,20,21,22].

The remainder of this paper is organized as follows: Section 2 describes the methods and concepts on which the three-phases method is founded: metadata and risk analysis/dynamic risk analysis. Section 3 presents the result, i.e., the method developed for dataset management on risk analysis. It describes the preliminary actions required for the application of the three-phases method, which is then fully detailed in the rest of the section as the main contribution of this work. Section 4 presents an application of the method to the case study. Section 5 discusses the case study results, as well as the benefits and limitations of the method, followed by conclusions in Section 6.

2. Materials and Methods—Metadata and Risk Analysis

Metadata can be defined as “data that provide information about other data” [23]. Metadata can provide structured information about a dataset without analyzing the dataset content. As highlighted by Wierling et al. [24], credible and traceable documentation of knowledge about the energy system is not possible without metadata. Despite its potential benefits for energy systems and data management optimization, the assessment of datasets through their metadata is not extensively explored in risk analysis. Indeed, there is no uniform definition of metadata to standardize the entire process of data production, processing, analysis, and use for prediction in the field of safety [25].

Data source management using its metadata in the context of risk analysis requires, at first, a clear definition of the level of analysis. In addition, it requires having a complete picture of all the datasets usable to inform the risk analysis parameters (i.e., an exhaustive description of the risk analysis parameters and a list of all the data sources usable to inform those parameters). While these two actions are common steps in risk analysis, they generally lack details that would enable an optimal dataset management. A reinforcement of those steps (“reinforcement actions”) is, thus, needed, as introduced in Section 2.2 and further detailed in Section 3.

This section presents an overview of concepts related to metadata and risk analysis. These do not constitute an exhaustive review and are limited to the description of the concepts applied in this paper.

2.1. Metadata Concepts

Metadata (i.e., “data that define and describe other data” [26]) report information concerning the structure and the content of a dataset or a service [27,28]. Metadata can be used for three main purposes: (1) content description (author, subject, etc.), (2) structural characterization (e.g., link between various parts of a resource), and (3) administrative management (access rights, file version, etc.) [29]. In addition to these features, metadata can be classified on the basis of a piece of information’s intrinsic vs. extrinsic property [30,31]. Although intrinsic properties may be assimilated to (1) content description and extrinsic properties cover, (2) structural characterization, and (3) administrative management, there is no broad consensus on the topic [32,33]. The classification and the metadata quality assessment depend, thus, on the task at hand [34], leading to new classifications if required.

Different metadata standards have been developed over the years, depending on the fields of application and the metadata’s purposes. The Metadata Standards Directory Working Group [35], a working group from the Research Data Alliance [36], has reported a community-maintained “open directory of metadata standards applicable to scientific data” [37]. An extract of this work is presented in Appendix A. This directory also reports the Dublin Core (DC), which is a generic standard developed on Semantic Web principles (or a “web of linked data”) [38,39] and managed by the Dublin Core™ Metadata Initiative, or DCMI. DCMI aims at developing and sharing best practices in the design and management of metadata. It is an open, collaborative, international, cross-disciplinary, technology-neutral, and business model-/purpose-neutral organization dedicated to supporting metadata management since the 1990s [40,41]. Dublin Core is a widely used standard, also published as an ISO standard and NISO standard [42,43,44]. It contains 15 core terms and several properties, classes, datatypes, and vocabulary encoding schemes maintained by DCMI (DC terms) [45].

The adoption of the DC standard data management for risk analysis presents several advantages, such as the following:

• many of the data sources not conventionally considered may be made available online,
• cross-disciplinary standards are critical to the comparison of heterogeneous data sources,
• the importance taken over the years and continuous increase in cloud-based technologies and web-based applications,
• the importance of facilitating the sharing of data and knowledge, the collaboration, the research and development, and the innovation adoption to third parties both in the risk community and across industries.

Furthermore, using the DC standard allows using DC-related crosswalks, facilitated by the international long-term recognition of the DC metadata standard. Crosswalks enable highlighting the nature of the overlap and gaps between different metadata standards through a table or a figure. In addition, they allow pinpointing the existence or the absence of relationships between terms existing in the respective standards [46]. Multiple examples of crosswalks linking recognized schemata can be found online, such as the one provided by the Getty Research Institute [47], the one provided by the Metadata Working Group of the Emory University [48], or the one provided by the UBC Faculty Research and Publications [49]. Non-standardized crosswalks (e.g., internal) may also be considered when discrepancies are observed between the format followed for metadata reporting in a selected file and the existing standards (e.g., due to explicit choices related to specific metadata needs, or due to a simple lack of competencies). Hence, the content from other schemas can always be linked to the Dublin Core schema.

2.2. Conventional Risk Analysis and Dynamic Risk Analysis

The concept of risk is generally related to three principal elements, as displayed in Equation (1) [50].

Risk = f(s, p, c),

(1)

where s corresponds to a specific scenario, p corresponds to the probability of occurrence of this specific scenario, and c corresponds to the resulting consequences.

Various standards adopt this definition for defining the steps of risk assessment [19,51,52]. Figure 1 presents the different steps to be followed within a risk assessment [51]. The figure also presents the placement of the proposed reinforcement actions to be described in Section 3.1, in blue and in red. Note that additional steps are identified in green in the figure: the management of datasets for informing the risk assessment. These steps consist of the method proposed in this paper, as described in Section 3.2 and Section 3.3.

Different sources of uncertainties may arise during a risk assessment, one of them being data processing [53]. The “level of knowledge” to represent some of these sources was added to Kaplan and Garrick’s definition (Equation (2)) by Aven and Krohn [54].

Risk = f(s, p, c, k),

(2)

where the variable k corresponds to the level of knowledge and is added to the variables s, p, and c corresponding to scenario, probability of occurrence, and consequence, respectively.

The assessment of the level of knowledge requires a proper characterization of the information pipeline, starting with the data acquisition [16,19]. The concept of “best level of information” selection, associated with the concept of “dynamicity”, can help in ensuring more efficient risk assessment and having a clear picture of the related uncertainties.

The notion of dynamicity was recently added to the principles of risk management presented within ISO 31000. Dynamic risk management approaches aim not only to update the data to consider, but also to adapt and reconsider, if necessary and on the basis of new risk evidence [55], the assumptions and models retained in previous cycles of the assessment [15,56,57,58]. As such, those techniques avoid lock-ins from initially considered conditions and process inertia by integrating, by design, the possibility to appropriately reshape the risk assessment process while minimizing the required efforts [59].

Despite the increasing number of publications and recognition of its relevance in ISO 31000:2018, DRA remains in an embryonic phase [15,60,61]. Limited research in the field hinders its implementation and the possibilities of improvements of DRA techniques. The lack of a systematic approach for identifying available data, as well as characterizing and managing data sources, also poses a challenge for the adoption of DRA, as it is a data-driven method. The method proposed in this paper intends to address this gap through the reinforcement actions detailed in Section 3.1 and the addition of two steps, presented in Section 3.2 and Section 3.3.

3. Results—Dataset Management Method for Dynamic Risk Analysis of Large-Scale Infrastructures

This section presents the resulting method developed for dataset management. It starts by describing the reinforcement steps required to apply the three-phases method. In Section 3.1, the main building blocks of the method are presented in Section 3.2, followed by the detailed description of the method elements in Section 3.3.

3.1. Risk Analysis Framework Reinforcement: Level of Analysis and Dataset Characterization

This subsection first details to which extent information should be characterized to enable a standard risk assessment. It then presents two reinforcement actions (RA1 and RA2) applied to existing steps of a standard risk assessment (Figure 1), namely, establishing the context (sub-steps 1₁, 1₂, 1₅) (RA1) and hazard identification (step 2), analysis of potential initiating event (sub-step 3₁), and analysis of potential consequences (sub-step 4₁) (RA2). The reinforcement of these steps is necessary for applying the proposed method for dataset selection (Section 3.2 and Section 3.3).

3.1.1. Information Characterization Requirements

Considering that the numerical values used within a QRA are all directly or indirectly based on measurements, best practices applied in metrology (i.e., the “science of measurement and its application” [62]) can be adopted as a reference. The measurement process in metrology is defined as “a set of operations to determine the value of a quantity” [63]. Its design represents a critical phase and consists, from a high-level point of view, in answering the following questions to execute a measurement adequately:

• Which quantity shall be measured?
• What are the required quality indicators (e.g., accuracy, precision, (see Figure 2))?
• Which measurement methods shall be used?
• Which equipment shall be used?
• Which software shall be used?
• Who is going to execute the measurement?
• What are the ambient conditions and influencing quantities affecting the measurement process?

Providing the described level of detail is critical for the validity of a measurement result, and to improve the traceability of a measurement. This is particularly relevant for risk assessment and recalls the paramount importance of a proper context characterization. Indeed, answering the question “Which quantity shall be measured?” requires first an adequate identification of the information that is sought out. This action should be executed within step (1) of the risk assessment (Figure 1) (“establishing the context”), as part of the global definition of the problem to address.

Three main points among those reported in the context establishment of the NORSOK Z-013 standard [51] need to be defined to adequately characterize the information one should look for:

• The objectives (defining the objective functions and indicating which type of information should be chosen),
• The scope (characterizing to which extent this information needs to be researched),
• The system boundaries (characterizing under which considerations and within which system delimitations the data need to be sought out).

3.1.2. Reinforcement Actions: Level of Analysis and Available Data Sources

The implementation of risk analyses is, in practice, strongly constrained by the availability of needed resources [64,65]. Hence, the adequate level of analysis is a tradeoff decision between stakeholder expectations and analytical possibilities [66,67]. Figure 3 illustrates the nature of the tradeoff to be found when defining the optimal level of analysis. The optimal analysis level can be considered as the level of convergence between a top-down and a bottom-up process. The top-down process consists of the progressive detailing of a global ambition associated with a resource budget allocation. The bottom-up process consists of progressively aggregating and restoring required information most efficiently while reducing information loss [68]. The dotted line in Figure 3 can be read as the level of convergence; it can be scrolled up or down depending on objectives and conditions. Note that no budget would enable a microscopic analysis of a large and complex system, and some level of abstraction is inevitable. On the other hand, no analysis can be limited to a high-level identification of risk-related objectives, and some level of details will always be required for meaningful decision making.

The definition of the optimal level of analysis is related to steps 1₁, 1₂, and 1₅ of the risk assessment method, namely, definition of objectives, definition of the scope, and definition of system boundaries and systems basis (Figure 1). Defining the level of convergence (reinforcement action 1) supports identifying the most relevant system, subsystem, assembly, subassembly, or component on which a risk assessment shall be executed. Following this identification, the next phase consists of building on the following steps commonly applied in risk assessment [69], augmented with reinforcement action 2, as presented in Figure 4:

• Task (I): applying a hazard identification (HAZID), i.e., identifying all relevant hazards and hazardous events,
• Task (II): describing the relevant accident scenarios,
• Task (III): reporting all dimensions to be considered for the hazardous events addressed in each scenario, from both a probability and a consequence perspective,
• Task (IV): identifying and characterizing all relevant parameters per reported dimension,
• Task (V): identifying all the data sources providing, to any extent, information to those parameters on the basis of experience, expertise, and further benchmarks.

Reinforcement action 2 consists, thus, of preselecting a list of data sources to inform different parameters which, in turn, inform different dimensions needed for quantifying the probability of occurrence and the consequences of a specific scenario. The list of preselected data sources should be completed by looking at all the accessible data sources and determining if those can provide (to any extent possible) knowledge about the needed parameters. For traceability, the preprocessing tasks enabling one to link a data source to a parameter shall also be reported.

The characterization of the parameters (Task IV) is a crucial step. It starts by reporting attributes relevant in any measurement process, i.e., the unit, the optimal resolution, and the range.

At this stage, considering that suboptimal resolution may often be faced, it is also strategical to define acceptable subcategories of information as second-best options to enable a semiquantitative evaluation when no other possibilities exist.

Risk analysis should ideally be site-specific [13,58] and performed in real time to avoid the possibility of building on outdated data and outdated considerations [70]. Therefore, risk analyses are time- and location-sensitive, and any spatiotemporal divergence between the site to be analyzed and the data that are considered will impact the results. Thus, the following questions should also be answered for characterizing the risk parameters:

• How location-sensitive is the parameter under review?
• What is the spatial extrapolation potential, i.e., the capacity, given data provided for a particular parameter in a delimited geographical area, to estimate values for that specific parameter in the surrounding of the initially considered area?
• How quickly does the parameter under review usually change over time?
• What is the relevant time changing rate?
• How long would it take before the dataset considered for the parameter under review to be outdated?

Depending on the scope of the risk assessment being performed, an applicable spatial scale may be the following (in square meters): “not applicable (NA), individual or <10⁰”, “<10¹”, “<10²”, “<10³”, “≥10³”. Similarly, a timescale could be reported as “hours”, “days”, “weeks”, “months”, “years”, “decades”, or “constant” (i.e., no change over the lifetime of the site).

In summary, a parameter pa can be characterized through the vector

$$pa=\left[\begin{array}{c}\begin{array}{c}R{s}_{pa}\\ SL{I}_{pa}\\ R{a}_{pa}\\ SE{P}_{pa}\\ T{S}_{pa}\end{array}\end{array}\right],$$

where, within a pre-defined scope, Rs_pa corresponds to the optimal resolution of the parameter pa based on the chosen unit, SLI_pa corresponds to the sublevel of information of the parameter pa acceptable for semiquantitative evaluations, Ra_pa corresponds to the range of values taken by the parameter pa, SEP_pa corresponds to the spatial extrapolation potential of the parameter pa, and TS_pa corresponds to the temporal sensitivity of the parameter pa.

Thus, the implementation of the actions reported up to step 3₁ in Figure 1, reinforced with the reinforcement actions 1 and 2, allows obtaining a preselection of all potentially relevant data sources. Additionally, it enables one to precisely list the attributes usable for a quality assessment of the information provided by a dataset in terms of risk quantification.

3.2. Dataset Management: Three-Phases Method Overview

Data quality assessment has a long research history [71] and is usually executed by comparing the value of specified data quality indicators to preliminary defined reference values. The quality of the information can be assessed using various dimensions, such as accuracy, precision, coverage, completeness, timeliness, reliability, trustworthiness, traceability, comparability, costs, and metadata [72,73,74,75,76,77,78,79]. Section 3.2.1 discusses the most relevant dimensions for risk analysis and shows how those can be characterized using the terms defined in the Dublin Core standard. This is then used as the foundation for the proposed data management method, described in Section 3.2.2.

3.2.1. Dataset Characterization for Risk Analysis

Efficient dataset management for risk analysis relies on the characterization of three main features, as described below: nature of the dataset, site/time specifications of the dataset, and agents and factors influencing data management.

(i)

Nature of the dataset

The technologies used to capture data determine which type of file will be generated. This directly impacts the obtainable performance in terms of resolution, range coverage percentage (how much of the predefined range can be covered), precision, and accuracy. For instance, the best spatial resolution available via commercial satellite images is much lower than that provided by LiDAR point clouds (30 cm vs. a few millimeters) [80,81,82]. Furthermore, satellite images are mainly used to provide 2D information, while LiDAR point clouds are usually used to obtain 3D insights.

(ii)

Spatiotemporal characterization of data

Figure 5 illustrates information provided for a unique and generic parameter, at three different resolutions, at three points in time (t − 2, t − 1, and t), for a specific area of interest (AoI). While the most recent dataset with the highest resolution would be ideal, datasets are most often incomplete. Therefore, one may face situations where the highest spatial resolution is only available within an older dataset (e.g., t − 2 here), making datasets with coarser spatial resolution the only up-to-date option [83]. Additionally, one may also face a total absence of information in some regions (represented by the black region).

The management of incomplete datasets is an important task to be performed for most of the parameters involved in a risk analysis. This highlights the importance of adequately addressing the spatiotemporal characterization of the information provided by a dataset, and including it as a comparison and evaluation criterion.

(iii)

Agents and factors influencing data management

The value of information available in a dataset strongly depends on the competencies of the actors involved in the various steps of the data management (i.e., data capture, data transmission, data storage, data pre-processing, information processing, results transmission) [16,71]. The trust to be given to the information provided by a dataset is, thus, strongly influenced by, e.g., the standards and protocols followed when managing the data, the authority, and legitimacy of the actors involved [39,84].

Identifying the “trust” level, the spatiotemporal features and the nature of the dataset are, thus, essential for the characterization of the datasets to be used for risk assessment. These three features are the foundation for the data management in the three-phases method. Note that the implementation of reinforcement actions 1 and 2 as previously described is required to apply the method (Figure 6).

3.2.2. Three-Phases Method—Logic Description

The Dublin Core standard presented in Section 2.1 is used as a foundation to exploit the metadata in the three-phases method. We start by only selecting the terms that are relevant for risk assessment purposes, i.e., those related to the three features defined in Section 3.2.1. We then regroup the terms into three classes by following a similar logic: (1) file (nature of the dataset), (2) scene (site-/time-specifications of the dataset), and (3) objectives/author/circumstances (agents and factors influencing data management).

Table A2. File-related DC terms.

N°	Term	Definition
1-1	Format	The file format, physical medium, or dimensions of the resource.
1-2	Type	The nature or genre of the resource.

,

Table A3. Scene-related DC terms.

N°	Term	Definition
2.1	Coverage	The spatial or temporal topic of the resource, spatial applicability of the resource, or jurisdiction under which the resource is relevant.
2.2-a	Spatial	Spatial characteristics of the resource (sub-property of coverage).
2.3-b	Temporal	Temporal characteristics of the resource (sub-property of coverage).
2.4-b	Date	A point or period of time associated with an event in the lifecycle of the resource.
2.5-b	Created	Date of creation of the resource (sub-property of date).
2.6-b	Issued	Date of formal issuance of the resource (sub-property of date).
2.7-b	AccrualPeriodicity	The frequency with which items are added to a collection.
2.8-b	Available	Date that the resource became or will become available (sub-property of date).
2.9-b	Modified	Date on which the resource was changed (sub-property of date).
2.10-b	Valid	Date (often a range) of validity of a resource (sub-property of date).
2.11-b	DateCopyrighted	Date of copyright of the resource (sub-property of date).
2.12-b	DateSubmitted	Date of submission of the resource (sub-property of date).
2.13-b	DateAccepted	Date of acceptance of the resource (sub-property of date).

and

Table A4. Objective/author/circumstance-related DC-terms.

N°	Term	Definition
3.1	Audience	A class of agents for whom the resource is intended or useful.
3.2	Abstract	A summary of the resource.
3.3	AccrualMethod	The method by which items are added to a collection.
3.4	BibliographicCitation	A bibliographic reference for the resource.
3.5	ConformsTo	An established standard to which the described resource conforms.
3.6	Contributor	An entity responsible for making contributions to the resource.
3.7	Creator	An entity responsible for making the resource.
3.8	Description	An account of the resource.
3.9	EducationLevel	A class of agents, defined in terms of progression through an educational or training context, for which the described resource is intended.
3.10	Extent	The size or duration of the resource.
3.11	HasVersion	A related resource that is a version, edition, or adaptation of the described resource.
3.12	IsReferencedBy	A related resource that references, cites, or otherwise points to the described resource.
3.13	IsReplacedBy	A related resource that supplants, displaces, or supersedes the described resource.
3.14	IsVersionOf	A related resource of which the described resource is a version, edition, or adaptation.
3.15	Provenance	A statement of any changes in ownership and custody of the resource since its creation that are significant for its authenticity, integrity, and interpretation.
3.16	Publisher	An entity responsible for making the resource available.
3.17	References	A related resource that is referenced, cited, or otherwise pointed to by the described resource.
3.18	Replaces	A related resource that is supplanted, displaced, or superseded by the described resource.
3.19	Source	A related resource from which the described resource is derived.
3.20	Subject	A topic of the resource.
3.21	Title	A name given to the resource.

in Appendix B detail this recategorization, together with the respective definition of each of the selected terms [45].

The terms categorized in the first class ((1) file) report the nature of the file. They are used to characterize the default maximum potential of knowledge (DMPK) that a specific data source can provide, based on the technological possibilities of the technique used to generate the dataset (e.g., satellite-based orthophoto, LiDAR-based point cloud).

The terms categorized in the second class ((2) scene) report the spatiotemporal properties of the file. This class can be divided into two subclasses: (2a) spatial and (2b) temporal. The use of information provided in class (2) scene enables one to calculate a first degradation factor (DF₁, composed of DF_1a and DF_1b, relative to spatial and temporal information, respectively) on the basis of the difference in nature between the spatiotemporal requirements of the site to be analyzed and the spatiotemporal properties of the considered dataset.

The terms categorized in the third class ((3) objectives/author/circumstances) report contextual information. They enable calculating a second degradation factor (DF₂), characterizing the level of trust one assigns to the analyzed dataset. In opposition to the first degradation factor, the second degradation factor calculation can be considered as a more dynamic and subjective task, as the trust level is strongly influenced by the stakeholders supervising the risk analysis [85]. For instance, understanding a problem and the knowledge of the mentioned actors/standards could be very different between two distinct teams [86], a standard may become outdated and withdrawn after some time, etc.

Figure 7 illustrates the sequencing of the phases required to calculate an assessed dataset’s informative potential (i.e., potential of knowledge).

The notions of “degradation factors” have been chosen because divergences observable via the analysis of properties relative to terms in class (2) scene and (3) objectives/author/circumstances can only neutrally or negatively impact the maximum performances of the knowledge acquired via the analysis of properties relative to terms in class (1) file.

The analytical order of the phases aims to optimize future data processing: the spatial overlap is assessed before analyzing the temporal properties to automatically discard non-overlapping datasets. Furthermore, one may still decide to valorize the analysis of properties relative to terms present in classes (1) file and (2) scene, despite a lack of qualifications leading to an absence of trust-related quality check. Including trust-related quality checks in the final calculation may, thus, have to be appreciated on a case-by-case study of the problems, justifying a final position for this task in the method.

Table 1. List of assumptions made for the development of the presented method.

N°	Assumptions
1	We place ourselves in a situation where we can apply all steps previously discussed (i.e., optimization of analysis level, HAZID on selected item, scenario identification, probability, and consequence characterization (i.e., identification of all involved dimensions, parameters and usable data sources), characterization of the required information for each parameter, and ability to report the metadata of the selected datasets following the DC standard).
2	A consensus is assumed among all the stakeholders involved in each method development step.
3	All datasets are initially considered external to the stakeholders involved in the risk analysis, thus needing to go through the method similarly.
4	All analyzed datasets are considered independent.
5	All datasets are considered to be analyzed independently and not leveraging on one another.
6	The quality of the datasets analyzed in earlier risk analyses is assumed to be optimized regarding acquisition conditions and state-of-the-art possibilities in the field (resolution, scale calibration, etc.), and the data are considered to be acquired by an expert.
7	The metadata of all datasets are convertible in DC terms.
8	No advanced natural language processing is used to extract information from text in this first version of the method.
9	A data source can uniquely be identified on the basis of the format and the type of a resource.
10	The number of most obvious invalid records can be indicated using dataset quality indicators. Although not originally reported in the DC standard, such information can easily be added to existing metadata.
11	The number of missing values can be indicated using dataset quality indicators. Although not originally reported in the DC standard, such information can easily be added to existing metadata.
12	Missing values can be characterized in time and/or space (when relevant).
13	Trust-related properties are dataset-specific and generalizable to all parameters informed.
14	All datapoints of the same dataset are acquired using a unique acquisition process.
15	The reported spatiotemporal information of datasets is assumed to be accurate (no mismatch).

presents the assumptions considered during the development of the method, followed by a detailed description of each of the phases in the next subsection.

3.3. Quantitative Elements of the Three-Phases Method

The three-phases method identifies and characterizes multiple data source/dataset properties through a number of classes and respective boundaries. The characterization of these classes is based on the authors’ experience with data management for risk analysis. Those choices are valid from a generic perspective to the best of our knowledge. However, the method offers the flexibility for this information to be adapted to the context in which the method will be applied. The values shall, thus, be seen as an indication instead of a static and rigid formalization. The implications of those choices are further discussed in Section 5.

3.3.1. Phase 1: Default Maximum Potential of Knowledge (DMPK)—Calculation

The evaluation approach of class (1) file consists of the calculation of the DMPK, which is assessed per data source sr and per parameter pa (i.e., DMPK_sr,pa). The DMPK_sr,pa is a function of four properties identified on the basis of common data quality assessment criteria [71] to estimate how well preselected data sources can inform a parameter. The DMPK can be calculated through a normalized weighted sum as in Equation (3).

$$DMP{K}_{sr,pa}=\frac{x_{LoI} \frac{Lo{I}_{sr,pa}}{3}+x_{RaC} \frac{Ra{C}_{sr,pa}}{3}+x_{Pr} \frac{P{r}_{sr,pa}}{2}+ x_{Ac} \frac{A{c}_{sr,pa}}{2}}{x_{LoI}+x_{RaC}+x_{Pr}+x_{Ac}},$$

(3)

where the variables are defined as follows:

• DMPK_sr,pa: default maximum potential of knowledge per data source sr and per parameter pa,
• LoI_sr,pa: the level of information for source sr and parameter pa,
• RaC_sr,pa: the range coverage for source sr and parameter pa,
• Pr_sr,pa: the precision for source sr and parameter pa,
• Ac_sr,pa: the accuracy for source sr and parameter pa,
• x_LoI, x_RaC, x_Pr, x_Ac: weights given by stakeholders to the level of information, the range coverage, the precision, and the accuracy of the data, respectively.

The weights give stakeholders the possibility to manage the importance given to meta-parameters as wished. For simplicity, a naïve approach setting those weights to 1 is applied for the rest of the present paper [87].

The use of the DMPK enables a first ranking of data sources based on their capacity to inform a specific parameter. Thereby, any new dataset ds originating from one of the reported data sources will automatically be given a DMPK score enabling an estimation of its a priori value for risk analysis.

Calculating the DMPK allows the stakeholders to identify the parameter characterization benefiting the most from data coming from a specific data source by assessing the DMPK scores for a unique source, and identify which dataset shall be used to inform a particular parameter depending on the origin of the different sets.

The four properties used for the DMPK calculation are described below.

Property 1.1.

Category of Obtainable Level of Information.

The obtainable level of information (LoI_sr,pa) required per parameter is based on the reachable resolution provided by the data source (

Table 2. LoI_sr,pa—obtainable level of information.

Question: “Which Level of Information Can Be Obtained?”
Classes	Score
None	0
Qualitative information	1
Acceptable sublevel of information	2
Precise measurement	3

), adapted from the classification of [62].

• Precise measurement, enabling to reach the expected resolution and, therefore, unlocking a potential full quantification,
• Acceptable sublevel of information, enabling a semiquantitative evaluation,
• Qualitative information (e.g., yes/no; +/−; shift of tendancy (e.g., mean)),
• None.

Property 1.2.

Range Coverage Potential.

The range coverage potential (RaC_sr,pa) concerns the completeness of a data source (i.e., the capacity for a data source to cover “all required parts of an entity’s description” [88]). It can be used for characterizing a candidate dataset by answering the question “How much of the predefined range can be covered?” (

Table 3. RaC_sr,pa—range coverage potential.

Question: “How much of the predefined range can be covered?”
Classes	Score
None	0
0% to 10%	1
10% to 90%	2
90% to 100%	3

).

Property 1.3.

Precision Estimation.

The precision meta-feature (Pr_sr,pa) indicates the precision of a data source, characterized through expert knowledge. The purpose is to evaluate, on the basis of experience, if the data source enables to systematically come to identical conclusions when assessing datasets acquired under repeatability conditions. This assessment is made by answering the question “Would an expert always come to the same conclusion when assessing datasets acquired under repeatability conditions?” (

Table 4. Pr_sr,pa—precision.

Question: “Would an expert always come to the same conclusion when assessing datasets acquired under repeatability conditions?”
Classes	Score
No	0
Probably to some extent	1
Yes, a priori	2

).

Property 1.4.

Accuracy Estimation.

The accuracy of a data source (Ac_sr,pa) is estimated through a classification built on expert knowledge. The purpose is to assess, on the basis of experience, the potential for the acquisition method to provide measurements centered around the true value. This assessment is, thus, made by answering the question “Does the method usually enable to provide conclusions centered around the true value?” (

Table 5. Ac_sr,pa—accuracy.

Question: Does the method usually enable to provide conclusions centered around the true value?”
Classes	Score
No	0
Probably to some extent	1
Yes, a priori	2

).

3.3.2. Phase 2: First Degradation Factor (DF₁)—Calculation and Application

The evaluation approach of class (2) scene is performed by calculating the first degradation factor DF₁. To calculate DF₁, we start by applying a subcategorization of the terms reported in Table A3, Appendix B. At this stage, one mainly looks for four types of information relative to the parameters pa evaluated in each considered dataset ds:

• Where the data were acquired (acquisition area (AA_ds,pa)),
• With which spatial resolution the data were acquired (spatial resolution (SRe_ds,pa)),
• When the recording of the data was initiated (Date_min,ds,pa) and, in case several recordings of them area are available, when the recording of the data was stopped (Date_max,ds,pa) (i.e., temporal range (TRa_ds,pa)),
• With which temporal resolution the data were acquired (temporal resolution (TRe_ds,pa)).

Therefore, assuming adequately registered metadata, one can decide to only focus on the terms 2.2a “spatial” and 2.2-b “temporal” in Table A3, Appendix B, for which detailing via subcategories (acquisition area, spatial resolution, temporal range, temporal resolution) can be used to report the required information. The rest of the terms in Table A3, Appendix B are considered redundant and potentially suboptimal for a spatiotemporal characterization relevant to risk analysis.

For simplicity, we further assume that no missing information is reported regarding the terms “spatial” and “temporal”. In addition, inspired by [79] and as further detailed where required in the following, we also suggest using additional dataset quality indicators. Although those are not initially reported in the DC standard, this information can automatically be added to existing metadata. In particular, we suggest inferring new spatiotemporal related terms using additional generic data quality measures, such as the number of missing values, non-expected records, or invalid records. This choice is further discussed in Section 5.

The global DF₁ can be calculated for any dataset ds and related parameter pa as

$$D{F}_{1,ds,pa}=\left(1-\Vert D{F}_{1a,ds,pa}\Vert \right)\times \left(1-\Vert D{F}_{1b,ds,pa}\Vert \right),$$

(4)

where the variables are defined as follows:

• DF_1,ds,pa: first degradation factor calculated per candidate dataset ds and per parameter pa,
• DF_1a,ds,pa: first degradation factor due to spatial properties, calculated per candidate dataset ds and per parameter pa,
• DF_1b,ds,pa: first degradation factor due to temporal properties, calculated per candidate dataset ds and per parameter pa.

The calculation detailing of DF_1a,ds,pa is presented in Section 3.3.2.1, and that of DF_1b,ds,pa is presented in Section 3.3.2.2.

An updated score can be given to the potential of knowledge (PK) for any dataset ds and related parameter pa as in Equation (5).

$$P{K}_{DF1,ds,pa}=DMP{K}_{sr,pa}\times D{F}_{1,ds,pa},$$

(5)

where the variables are defined as follows:

• PK_DF1,ds,pa: updated potential of knowledge of the dataset ds and related parameter pa after applying the first degradation factor,
• DMPK_sr,pa: default maximum potential of knowledge per data source sr and per parameter pa,
• DF_1,ds,pa: first degradation factor calculated per candidate dataset ds and per parameter pa.

The calculation of the updated potential of knowledge (PK_DF1,ds,pa) enables reconsideration and potentially reorganizing the ranking originally provided at the end of phase 1.

3.3.2.1. DF_1a—First Degradation Factor Due to Spatial Properties

DF_1a, the first element to be determined for setting up the global DF, is calculated per candidate dataset ds and per parameter pa (i.e., DF_1a,ds,pa). We consider five properties, which are further detailed below, to estimate the quality of a dataset with regard to its spatial characteristics. Those are used to determine the form of DF_1a through a normalized weighted sum as

$$D{F}_{1a,ds,pa}=\frac{x_{SC} \frac{S{C}_{ds,pa}}{3}+x_{SRe} \frac{SR{e}_{ds,pa}}{2}+x_{SDe} \frac{SD{e}_{ds,pa}}{3}+x_{SDi} SD{i}_{ds,pa}+x_{SN} \frac{S{N}_{ds,pa}}{3}}{x_{SC}+x_{SRe}+x_{SDe}+x_{SDi}+x_{SN}},$$

(6)

where the variables are defined as follows:

• DF_1a,ds,pa: first degradation factor due to spatial properties, calculated per candidate dataset ds and per parameter pa,
• SC_ds,pa: spatial coverage of candidate dataset ds per parameter pa,
• SRe_ds,pa: spatial resolution of candidate dataset ds per parameter pa,
• SDe_ds,pa: spatial density of candidate dataset ds per parameter pa,
• SDi_ds,pa: spatial distribution of candidate dataset ds per parameter pa,
• SN_ds,pa: spatial noise of candidate dataset ds per parameter pa,
• x_SC, x_SRe, x_SDe, x_SDi, x_SN: weights given by stakeholders to the spatial coverage, spatial resolution, spatial density, spatial distribution, and spatial noise of the data, respectively.

The weights give stakeholders the possibility to manage the importance given to meta-parameters as wished. For simplicity, a naïve approach setting those weights to 1 is applied for the rest of the present paper [87].

Given a list of predefined thresholds and the score obtained for DF_1a,ds,pa, one can automatically assess whether further processing a dataset under review is meaningful; further analysis of the dataset can be postponed and only reconsidered in the absence of other relevant datasets.

Property 2.1.

Spatial Coverage SC_ds,pa.

The spatial coverage indicates how much of the area of the selected item of interest (area of interest—AoI) is covered by the selected dataset (Acquisition area—AA). Mathematically, the percentage of spatial coverage sc_ds,pa of a dataset ds and for a parameter pa, with the AoI including the item under review in the risk analysis, can be expressed as in Equation (7).

$$s{c}_{ds,pa}=\frac{Part of AoI spatially covered by A{A}_{ds,pa} }{Total AoI},$$

(7)

where the variables are defined as follows:

• sc_ds,pa: spatial coverage of candidate dataset ds per parameter pa,
• AoI: area of interest,
• AA_ds,pa: acquisition area of candidate dataset ds and per parameter pa.

Table 6. SC_ds,pa—spatial coverage.

Classes	Score
Low (scds,pa < 50%)	−3
Medium (50% < scds,pa ≤ 80%)	−2
High (80% < scds,pa ≤ 95%)	−1
Very high (scds,pa > 95%)	0

presents the classes we propose to categorize sc_ds,pa for assessing the meta-feature SC_ds,pa.

Property 2.2.

Spatial Resolution SRe_ds,pa.

This property is used to assess if the dataset provides the minimum required level of information in terms of spatial resolution for a specific parameter. We suggest, for a dataset ds and a parameter pa, a relative classification based on the classes reported for the estimation of the spatial extrapolation potential SEP_pa in the parameter characterization (i.e., NA, individual, or <10⁰, <10¹, <10², <10³, ≥10³).

Table 7. SRe_ds,pa—spatial resolution.

Classes	Score
Distant (2 classes below or more)	−2
Close (1 class below)	−1
Sufficient (similar class or above)	0

presents the second meta-feature of learning phase 2a.

Property 2.3.

Spatial Density SDe_ds,pa.

This property is used to provide a statistical data quality check on the basis of the number of relevant missing values (spatially speaking). Mathematically, the spatial density sde_ds,pa of a dataset ds and for a parameter pa can be expressed as in Equation (8).

$$sd{e}_{ds,pa}=1-\frac{Number of missing value{s}_{ds,pa} }{Total number of expected record{s}_{ds,pa}},$$

(8)

where sde_ds,pa is the spatial density of candidate dataset ds per parameter pa.

The classes we propose to categorize sde_ds,pa, defining a third meta-feature of learning phase 2a, called SDe_ds,pa (spatial density for dataset ds and parameter pa), are reported in

Table 8. SDe_ds,pa—spatial density.

Classes	Score
Low (sdeds,pa < 50%)	−3
Medium (50% ≤ sdeds,pa < 80%)	−2
High (80% ≤ sdeds,pa < 95%)	−1
Very high (sdeds,pa ≥ 95%)	0

.

Property 2.4.

Spatial Distribution SDi_ds,pa.

This property is used to provide a statistical data quality check on the basis of the spatial distribution of missing values (spatially speaking). Mathematically, the spatial density sdi_ds,pa of a dataset ds and for a parameter pa can be expressed as in Equation (9).

$$sd{i}_{ds,pa}=\frac{Average distance bewteen missing value{s}_{ds,pa} }{Maximum distance between 2 expected record{s}_{ds,pa}},$$

(9)

where sdi_ds,pa is the spatial distribution of candidate dataset ds per parameter pa.

The classes we propose to categorize sdi_ds,pa, defining the fourth meta-feature of learning phase 2a, called SDi_ds,pa (spatial distribution for dataset ds and parameter pa), are presented in

Table 9. SDi_ds,pa—spatial distribution.

Classes	Score
Heterogeneous distribution (0 < sdids,pa < 30%)	−1
Homogeneous distribution (sdids,pa ≥ 30% or sdids,pa = 0)	0

.

Property 2.5.

Spatial Noise SN_ds,pa

This property is used to provide a statistical data quality check on the basis of the proportion of noise (spatially speaking). Mathematically, the spatial noise sn_ds,pa of a dataset ds and for a parameter pa can be expressed as in Equation (10).

$$s{n}_{ds,pa}=\frac{Number of invalid value{s}_{ds,pa} }{Total number of record{s}_{ds,pa}},$$

(10)

where sn_ds,pa is the spatial noise of candidate dataset ds per parameter pa.

A fifth meta-feature of learning phase 2a, called SN_ds,pa (spatial noise for dataset ds and parameter pa), can be used for characterizing a candidate dataset according to sn_ds,pa (

Table 10. SN_ds,pa—spatial noise.

Classes	Score
Low (snds,pa < 10%),	−3
Medium (10% ≤ snds,pa < 20%)	−2
High (20% ≤ snds,pa < 50%)	−1
Very high (snds,pa ≥ 50%)	0

).

3.3.2.2. DF_1b—First Degradation Factor Due to Temporal Properties

DF_1b, the second element to be determined for setting up the global DF₁, is calculated per candidate dataset ds and per parameter pa (i.e., DF_1b,ds,pa). We consider six properties-which are further detailed below-to estimate the quality of a dataset with regard to its temporal characteristics. Those are used to determine the form of DF_1b through a normalized weighted sum as:

$$D{F}_{1b, ds, pa}=\frac{x_{TP} \frac{T{P}_{ds, pa}}{2}+x_{TOU} TO{U}_{ds, pa}+x_{TRe} \frac{TR{e}_{ds,pa}}{2}+x_{TDe} \frac{TD{e}_{ds,pa}}{3}+x_{TDi} TD{i}_{ds,pa}+x_{TN} \frac{T{N}_{ds,pa}}{3}}{x_{TP}+x_{TOU}+x_{TRe}+x_{TDe}+x_{TDi}+x_{TN}}$$

(11)

where the variables are defined as follows:

• DF_1b,ds,pa: first degradation factor due to temporal properties, calculated per candidate dataset ds and per parameter pa,
• TP_ds,pa: temporal pertinence of candidate dataset ds per parameter pa,
• TOU_ds,pa: temporal overlap utility of candidate dataset ds per parameter pa,
• TRe_ds,pa: temporal resolution of candidate dataset ds per parameter pa,
• TDi_ds,pa: temporal distribution of candidate dataset ds per parameter pa,
• TN_ds,pa: temporal noise of candidate dataset ds per parameter pa,
• x_TP, x_TOU, x_TRe, x_TDe, x_TDi, x_TN: weights given by stakeholders to the temporal pertinence, temporal overlap utility, temporal resolution, temporal distribution, and temporal noise of the data, respectively.

The weights give stakeholders the possibility to manage the importance given to meta-parameters as wished. For simplicity, a naïve approach setting those weights to 1 is applied for the rest of the present paper [87].

Note that the calculations of the temporal resolution TRe_ds,pa, the temporal density TDe_ds,pa, the temporal distribution TDi_ds,pa, and the temporal noise TN_ds,pa are meaningless for datasets considered as punctual in the calculation of the temporal overlap utility TOU_ds,pa (see details below). Therefore, those terms are not considered in the calculation of DF_1b in such a situation.

Property 2.6.

Temporal Pertinence TP_ds,pa.

This property is used to assess how meaningful the exploitation of a dataset ds is for the analysis of a parameter pa in view of the age of the dataset at a given date d and the temporal sensitivity TS_pa reported in the parameter characterization (i.e., hours, days, weeks, months, years, decades, or “constant”).

Mathematically, the temporal pertinence tp_ds,pa of a dataset ds and for a parameter pa at a given date d can be expressed as in Equation (12).

$$t{p}_{ds,pa}=\frac{Date of analysis d-Dat{e}_{max,ds,pa}}{T{S}_{pa}},$$

(12)

where the variables are defined as follows:

• tp_ds,pa: temporal pertinence of candidate dataset ds per parameter pa,
• TS_pa: temporal sensitivity of parameter pa,
• Date_max,ds,pa: date when the recording of the data was stopped.

The classes we propose to categorize tp_ds,pa, defining a first meta-feature of learning phase 2b, called TP_ds,pa (temporal pertinence for dataset ds and parameter pa), are reported in

Table 11. TP_ds,pa—temporal pertinence.

Classes	Score
Distant (2 classes below or more)	−2
Close (1 class below)	−1
Sufficient (similar class or above)	0

.

Property 2.7.

Temporal Overlap Utility TOU_ds,pa.

This property enables one to qualify the utility of the temporal overlap of dataset ds for a parameter pa considering the temporal sensitivity TS_pa reported in the parameter characterization (i.e., hours, days, weeks, months, years, decades, or “constant”). Mathematically, the temporal overlap utility tou_ds,pa of dataset ds for a parameter pa can be expressed as in Equation (13).

$$to{u}_{ds,pa}=\frac{Dat{e}_{max,ds,pa}-Dat{e}_{min,ds,pa}}{T{S}_{pa}},$$

(13)

where the variables are defined as follows:

• tou_ds,pa: temporal overlap utility of candidate dataset ds per parameter pa,
• TS_pa: temporal sensitivity of parameter pa,
• Date_max,ds,pa: date when the recording of the data was stopped,
• Date_min,ds,pa: date when the recording of the data was initiated.

The classes we propose to categorize tou_ds,pa, defining a second meta-feature of learning phase 2b, called TOU_ds,pa (temporal overlap utility for dataset ds and parameter pa), are reported in

Table 12. TOU_ds,pa—temporal overlap utility.

Classes	Score
Timeseries (touds,pa > 1)	−1
Punctual (touds,pa ≤ 1)	0

.

Property 2.8.

Temporal Resolution TRe_ds,pa.

This property is used to assess if the dataset enables providing the minimum required level of information in terms of temporal resolution for a specific parameter. We suggest, for a dataset ds and a parameter pa, a relative classification based on the classes reported for the estimation of the temporal sensitivity TS_pa reported in the parameter characterization (i.e., hours, days, weeks, months, years, decades, or “constant”). Therefore, a third meta-feature of learning phase 2b, called TRe_ds,pa (temporal resolution for dataset ds and parameter pa), can be used for characterizing a candidate dataset (

Table 13. TRe_ds,pa—temporal resolution.

Classes	Score
Distant (2 classes below or more)	−2
Close (1 class below)	−1
Sufficient (similar class or above)	0

).

Property 2.9.

Temporal Density TDe_ds,pa.

This property is used to provide a statistical data quality check on the basis of the number of relevant missing values (temporally speaking). Mathematically, the temporal density tde_ds,pa of a dataset ds and for a parameter pa can be expressed as in Equation (14).

$$td{e}_{ds,pa}=1-\frac{Number of missing value{s}_{ds,pa} }{Total number of expected record{s}_{ds,pa}},$$

(14)

where tde_ds,pa is the temporal density of candidate dataset ds per parameter pa.

Table 14. TDe_ds,pa—temporal density.

Classes	Score
Low (tdeds,pa < 50%)	−3
Medium (50% ≤ tdeds,pa < 80%)	−2
High (80% ≤ tdeds,pa < 95%)	−1
Very high (tdeds,pa ≥ 95%)	0

presents the classes we propose to categorize tde_ds,pa, defining a fourth meta-feature of learning phase 2b, called TDe_ds,pa (temporal density for dataset ds and parameter pa).

Property 2.10.

Temporal Distribution TDi_ds,pa.

This property is used to provide a statistical data quality check on the basis of the temporal distribution of missing values (temporally speaking). Mathematically, the temporal distribution tdi_ds,pa of a dataset ds and for a parameter pa can be expressed as in Equation (15).

$$td{i}_{ds,pa}=\frac{Average time bewteen missing value{s}_{ds,pa} }{Maximum time between 2 expected record{s}_{ds,pa}},$$

(15)

where tdi_ds,pa is the temporal distribution of candidate dataset ds per parameter pa.

Table 15. TDi_ds,pa—temporal distribution.

Classes	Score
Heterogeneous distribution (0 < tdids,pa < 30%)	−1
Homogeneous distribution (tdids,pa ≥ 30% or tdids,pa = 0)	0

presents the classes we propose to categorize tdi_ds,pa, defining a fifth meta-feature of learning phase 2b, called TDi_ds,pa (temporal distribution for dataset ds and parameter pa).

Property 2.11.

Temporal Noise TN_ds,pa.

This property is used to provide a statistical data quality check on the basis of the proportion of noise (temporally speaking). Mathematically, the temporal noise tn_ds,pa of a dataset ds and for a parameter pa can be expressed as in Equation (16).

$$t{n}_{ds,pa}=\frac{Number of invalid value{s}_{ds,pa} }{Total number of record{s}_{ds,pa}},$$

(16)

where tn_ds,pa is the temporal noise of candidate dataset ds per parameter pa.

Table 16. TN_ds,pa—temporal noise.

Classes	Score
Very high (tnds,pa ≥ 50%)	−3
High (20% ≤ tnds,pa < 50%)	−2
Medium (10% ≤ tnds,pa < 20%)	−1
Low (tnds,pa < 10%)	0

presents the classes we propose for defining a sixth meta-feature of learning phase 2b, called TN_ds,pa (temporal noise for dataset ds and parameter pa).

3.3.3. Phase 3: Second Degradation Factor (DF₂)—Calculation and Application

The evaluation approach of class (3) objectives/author/circumstances consists of calculating the second degradation factor DF₂. To calculate DF₂, we also suggest a recategorization of the terms reported in Table A4 in Appendix B on the basis of two motivations:

• We do not apply advanced natural language processing techniques in this first version of the method,
• The terms 2.9-b “modified” and 2.10-b “valid” in Table A3, Appendix B may also be used for trust assessment of a dataset.

Table 17. Description and categorization of trust-related meta-features.

Term	Meta-Feature	Pre-Defined Classes and Respective Values
Audience	Ads	Regulatory authorities (0) Field specialists (−1) Targeted non-specialists (−2) Open access (−3) Not valued (−4)
BibliographicCitation	BCds	Official (regulations, standards, recognized journals, etc.) (0) Valued (−1) Not valued (−2)
ConformsTo	CTds	Actual (0) Depreciated (−1) Unrecognized (−2)
Contributor	Cods	Official/authorities (0) Valued (−1) Not valued (−2)
Creator	Crds	Official/authorities (0) Valued (−1) Not valued (−2)
EducationLevel	ELds	Senior (0) Junior (−1) Trainee (−2) Not related (−3)
HasVersion	HVds	Latest version (0) Not first/not last version (−1) First version (−2)
IsReferencedBy	IRefBds	Official (regulations, standards, recognized journals, etc.) (0) Valued (−1) Not valued (−2)
IsReplacedBy	IRepBds	Nothing (0) Something (−1)
IsVersionOf	IVOds	Latest version (0) Not first/not last version (−1) First version (−2)
Modified	Mds	Original file (0) Not original file (−1)
Provenance	Prds	Official (regulations, standards, recognized journals, etc.) (0) Valued (−1) Not valued (−2)
Publisher	Puds	Official (regulations, standards, recognized journals, etc.) (0) Valued (−1) Not valued (−2)
References	Refds	Official (regulations, standards, recognized journals, etc.) (0) Valued (−1) Not valued (−2)
Replaces	Repds	Something (0) Nothing (−1)
Source	Srds	Official (regulations, standards, recognized journals, etc.) (0) Valued (−1) Not valued (−2)
Valid	Vds	Valid (0) Not Valid (−1)

presents the retained terms and their associated meta-features. The management of trust-related properties consists of defining the value given to the meta-features on the basis of lists of actors, standards, references, etc. associated with predefined classes and identified over time [39]. Those meta-features are only dataset-specific and affect all parameters informed by the dataset identically.

As a result, we determine the form of the DF₂ for any dataset ds through a normalized weighted sum as in Equation (17).

$$\begin{array}{c}D{F}_{2,ds}=(x_A \frac{A_{ds}}{4}+x_{BC}\frac{B{C}_{ds}}{2}+x_{CT}\frac{C{T}_{ds}}{2}+x_{Co}\frac{C{o}_{ds}}{2}+x_{Cr}\frac{C{r}_{ds}}{2}+x_{EL}\frac{E{L}_{ds}}{3}+x_{HV}\frac{H{V}_{ds}}{2}+x_{IRefB}\frac{IRef{B}_{ds}}{2}+x_{IRepB} IRep{B}_{ds}\\ +x_{IVO}\frac{IV{O}_{ds}}{2}+x_M{M}_{ds}+x_{Pr}\frac{P{r}_{ds}}{2}+x_{Pu}\frac{P{u}_{ds}}{2}+x_{Ref}\frac{Re{f}_{ds}}{2}+x_{Rep}Re{p}_{ds}+x_{Sr}\frac{S{r}_{ds}}{2}+x_V{V}_{ds})/(x_A+x_{BC}\\ +x_{CT}+x_{Co}+x_{Cr}+x_{EL}+x_{HV}+x_{IRefB}+x_{IRepB}+x_{IVO}+x_M+x_{Pr}+x_{Pu}+x_{Ref}+x_{Rep}+x_{Sr}+x_V)\end{array}$$

(17)

where A_ds, BC_ds, CT_ds, Co_ds, Cr_ds, EL_ds, HV_ds, IRefB_ds, IRepB_ds, IVO_ds, M_ds, Pr_ds, Pu_ds, Ref_ds, Rep_ds, and Sr_ds are defined in Table 17, and x_A, x_BC, x_CT, x_Co, x_Cr, x_EL, x_HV, x_IRefB, x_IRepB, x_IVO, x_M, x_Pr, x_Pu, x_Ref, x_Rep, x_Sr, and x_V are weights given by stakeholders to the properties defined in Table 17.

The weights give stakeholders the possibility to manage the importance given to meta-parameters as wished. For simplicity, a naïve approach setting those weights to 1 is applied for the rest of the present paper [87].

Following the determination of DF₂, we can update the score given to the PK for any dataset ds and related parameter pa as in Equation (18).

$$P{K}_{DF1,DF2,ds,pa}=P{K}_{DF1,ds,pa}\times (1-\Vert D{F}_{2,ds}\Vert ),$$

(18)

where the variables are defined as follows:

• PK_{DF1,DF2,ds,pa}: updated potential of knowledge of the dataset ds and related parameter pa after applying the first and the second degradation factors,
• PK_DF1,ds,pa: updated potential of knowledge of the dataset ds and related parameter pa after applying the first degradation factor,
• DF_2,ds: second degradation factor calculated per candidate dataset ds.

The calculation of the updated potential of knowledge (PK_{DF1,DF2,ds,pa}) enables a final reconsideration and potential reorganization of the dataset ranking as an output of phase 2. The result is a ranking of data sources optimized for the potential of knowledge for each of the parameters that the available datasets can inform. The application of the presently described method ensures that the data used to estimate both probabilities and consequences required for the risk analysis correspond to the best level of information available to the stakeholders, as expected by ISO 31000 [19].

4. Case Study—Power-Grid Risk Analysis

This section illustrates the method described in Section 3 through a simplified application to vegetation management of power grids. It describes the context, the hazard identification, and the application of reinforcement actions 1 and 2 (Section 4.1 and Section 4.2). The three-phases method is applied in Section 4.3. The assessment is based on the evaluation of six experts specialized in risk analysis, data analytics, power-grid management, and vegetation analysis. The case study aims to illustrate the applicability and pertinence of the proposed method, rather than a full analysis covering all aspects required for executing a complete quantitative analysis. The scope is, thus, limited to large-scale power grids in Norway. Additionally, we consider only a sub-selection of parameters and a sub-selection of data sources/datasets relative to one specific dimension involved in the probability of outages due to tree fall on power lines, as detailed below.

Power grids are pillars for the good functioning of our modern and digitalized society. An important part of those networks consists of overhead power lines used for both transportation and distribution of power in regional, national, and international configurations [89]. Several hazards may compromise the integrity of those power lines. For instance, large-impact events can destroy overhead power lines, such as hurricanes, ice storms, and landslides [22]. They can also be damaged due to more local hazards, such as vegetation [5,83]. Indeed, vegetation represents a primary source of outages and has been identified as one of the root causes of some major blackouts in history [90].

Vegetation can lead to outages either via trees falling on the power lines (scenario 1) or by growing under the infrastructure until grounding one phase (scenario 2). Power-grid operators, thus, need to periodically inspect their entire network and trim vegetation in areas showing a higher probability of dangerous tree falls to avoid scenario 1. However, the way such operations are executed today (e.g., helicopter-based, foot patrols) is time-consuming, expensive, and challenging in remote and potentially hazardous areas. A risk-based approach can, thus, optimize the prioritization of actions to execute, and the decision making can be enhanced if supported by the maximum available existing data.

4.1. Reinforcement Action 1—Level of Analysis

The level of analysis for risk assessment of large-scale power grids can range from macroscopic perspectives (e.g., satellite-based inspections [20,22,91]) to a microscopic perspective (e.g., asset structural analysis [89]). Considering the nature of the infrastructure, the hazard, and resource constraints that power-grid operators usually face, we define the optimal level of analysis for risk assessment in power grids as the size of individual items (substations, power poles, etc.). We additionally break down power lines to obtain more localized items due to the extended nature of those assets. Figure 8 illustrates such a subdivision via an aerial perspective.

Figure 9 shows the item of interest chosen for the risk analysis. It furthermore illustrates the presence of trees along the power lines.

4.2. Reinforcement Action 2—Parameter Characterization and Data Source Identification

Three different dimensions can be reported when estimating the probability of outage due to a tree falling on a power line.:

• The physical configuration,
• The stability of the trees surrounding the power lines,
• External factors, such as strong winds.

The following parameters play a role in the definition of the physical configuration:

• Vegetation density/number of trees (*),
• Forest social configuration (i.e., distance characterization between trees),
• Height of tree (*),
• Structure of tree crown (depth),
• Structure of tree crown (width, diameter) (*),
• Terrain exposure to wind,
• X–Y direction from a tree to the power line,
• X–Y distance from a tree to the power line (*),
• Z-delta (intensity of altitude variation).

Table 18. Characterization of a subselection of parameters.

Parameters	Unit and Optimal Resolution	Acceptable Sub-Level of Information for Semi-Quantitative Evaluation	Minimum Range (Nominal Unit)	Spatial Extrapolation Potential (sqm): NA, Individual, or <100, <101, <102, <103, ≥103)	Relevant Time Changing Rate (Hours, Days, Weeks, Months, Years, Decades, “Constant”)
Density/number of trees	Number of trees/100 m2	High, medium, low (e.g., percentage of tree coverage/100 m2)	0–30	<102	Years
Height of tree	Meters (cm)	Large, medium, small	1–50	NA, individual, or <100	Months
Structure of tree crown (width, diameter)	Meters (cm)	Large, medium, small	0–30	NA, individual, or <100	Weeks
X–Y distance from a tree to the power line	Meters (cm)	High, medium, low, very low, e.g., high (x > 30), medium (10 < x < 30), low (1 < x < 10), very low (x < 1 m)	0–50	NA, individual, or <100	Weeks

reports the characterization we propose for the four parameters we selected in this case study designated by (*) in the preceding list.

Identification of data sources able to provide information for the four retained parameters is then executed. LiDAR point clouds, orthophotos based on aerial images, and satellite-based orthophotos correspond to some of the relevant data sources. The complete list of preselected sources is reported in Appendix C. Appendix C also reports suggestions of preprocessing methods usable to link each data source to the selected parameters.

4.3. Three-Phases Method Application

The geolocation of the AoI integrating the item of interest is given in the ETRS89/UTM32N coordinate system as follows:

• Minimum easting (X): 610,205,
• Minimum northing (Y): 6,561,098,
• Maximum easting (X): 610,253,
• Maximum northing (Y): 6,561,122.

The risk analysis is assumed to be made on 1 December 2021.

We consider a selection of three datasets to evaluate the probability of outage due to tree falls on power lines: a LiDAR point cloud, an orthophoto based on aerial images, and a satellite-based orthophoto.

The chosen files correspond to simulated realistic datasets generated for the present study. We assume having used crosswalks where required, and we report, for each dataset, the equivalent of original DC terms necessary for the analysis in

Table 19. DC-like terms for the three simulated datasets.

N°	Term		Dataset (a)	Dataset (b)	Dataset (c)
File-Related terms
1-1	Format		LASF	TIFF	JPEG2000
1-2	Type		LiDAR point cloud	Orthophoto based on aerial images	Satellite-based orthophoto
Spatiotemporal-related terms
2.2-a	Spatial
		Acquisition area	Min. easting (X): 609,600 Min. northing (Y): 6,561,000 Max. easting (X): 610,399 Max. northing (Y): 6,561,599	Min. easting (X): 609,731 Min. northing (Y): 6,560,621 Max. easting (X): 610,639 Max. northing (Y): 6,561,425	Min. easting (X): 599,395 Min. northing (Y): 6,514,003 Max. easting (X): 638,139 Max. northing (Y): 6,601,208
		Resolution	<1 m (5 points per square meter)	0.2 m	10 m
2.3-b	Temporal
		Datemin	25 April 2012	9 June 1989	14 June 2021
		Datemax	/	/	/
		Resolution	/	/	/
Objective/author/circumstance-related terms
3.1	Audience		Open access		Brief standard expertise
3.4	BibliographicCitation
3.5	ConformsTo
3.6	Contributor
3.7	Creator		Terratec AS	Norsk luftfoto og fjernmåling AS	ESA
3.9	EducationLevel		Not related
3.11	HasVersion		0.0.2
3.12	IsReferencedBy			Norge i bilder
3.13	IsReplacedBy		NDH Østfold 5pkt 2015
3.14	IsVersionOf
3.15	Provenance		Kartverket Oslo	Geovekst	ESA
3.16	Publisher			Rambøll Norge AS
3.17	References		‘LAStools (c) by rapidlasso GmbH’; ‘lasheight (141117) commercial’
3.18	Replaces		None
3.19	Source
2.9-b	Modified		10 July 2018
2.10-b	Valid

. Note that the nature of the considered files and the nature of the evaluated parameters lead the spatiotemporal characteristics (acquisition area, spatial resolution, etc.) considered in the present case study to similarly impact all addressed parameters. The outcome of each phase (i.e., the ranking of the datasets based on their estimated informative potential after the application of each phase) is summarized in a unique table (Table 25) at the end of Section 4.

4.3.1. Default Maximum Potential of Knowledge (DMPK)

The knowledge acquired during learning phase 1 enabling one to report the DMPK is detailed per data source and per parameter in

Table 20. Meta-features phase 1—DMPK calculation per data source per parameter.

Parameters per Source	Obtainable Level of Information LoIsr,pa	Range Coverage Potential RaCsr,pa	Precision Estimation Prsr,pa	Accuracy Potential Acsr,pa	Default Maximum Potential of Knowledge DMPKsr,pa
LiDAR point cloud
Density/number of trees	Precise measurement (3)	90% to 100% (3)	Yes, a priori (2)	Yes, a priori (2)	(3/3 + 3/3 + 2/2 + 2/2)/4 = 1
Height of tree	Precise measurement (3)	90% to 100% (3)	Yes, a priori (2)	Yes, a priori (2)	(3/3 + 3/3 + 2/2 + 2/2)/4 = 1
Structure of tree crown (width, diameter)	Precise measurement (3)	90% to 100% (3)	Yes, a priori (2)	Yes, a priori (2)	(3/3 + 3/3 + 2/2 + 2/2)/4 = 1
X–Y distance from a tree to the power line	Precise measurement (3)	90% to 100% (3)	Yes, a priori (2)	Yes, a priori (2)	(3/3 + 3/3 + 2/2 + 2/2)/4 = 1
Orthophoto based on aerial images
Density/number of trees	Precise measurement (3)	90% to 100% (3)	Yes, a priori (2)	Probably to some extent (1)	(3/3 + 3/3 + 2/2 + 1/2)/4 = 0.875
Height of tree	Acceptable sublevel of information (2)	10% to 90% (2)	Probably to some extent (1)	Probably to some extent (1)	(2/3 + 2/3 + 1/2 + 1/2)/4 = 0.5825
Structure of tree crown (width, diameter)	Precise measurement (3)	90% to 100% (3)	Yes, a priori (2)	Yes, a priori (2)	(3/3 + 3/3 + 2/2 + 2/2)/4 = 1
X–Y distance from a tree to the power line	Precise measurement (3)	90% to 100% (3)	Probably to some extent (1)	Yes, a priori (2)	(3/3 + 3/3 + 1/2 + 2/2)/4 = 0.875
Satellite-based orthophoto
Density/number of trees	Acceptable sublevel of information (2)	90% to 100% (3)	Probably to some extent (1)	Probably to some extent (1)	(2/3 + 3/3 + 1/2 + 1/2)/4 = 0.665
Height of tree	Acceptable sublevel of information (2)	10% to 90% (2)	No (0)	Probably to some extent (1)	(2/3 + 2/3 + 0/2 + 1/2)/4 = 0.4575
Structure of tree crown (width, diameter)	Acceptable sublevel of information (2)	10% to 90% (2)	Probably to some extent (1)	Probably to some extent (1)	(2/3 + 2/3 + 1/2 + 1/2)/4 = 0.5825
X–Y distance from a tree to the power line	Acceptable sublevel of information (2)	90% to 100% (3)	Probably to some extent (1)	Probably to some extent (1)	(2/3 + 3/3 + 1/2 + 1/2)/4 = 0.665

.

The scores obtained via the calculation of the DMPK for each data source and each parameter enable generating an initial ranking per parameter of the considered datasets, as described in Table 25.

4.3.2. First Degradation Factor (DF₁)

Table 21. Inferred scene-related DC terms.

Inferred Terms	Dataset (a)—LiDAR Point Cloud	Dataset (b)—Orthophoto Based on Aerial Images	Dataset (c)—Satellite-Based Orthophoto
Spatial inferred terms
Spatial coverage	100%	100%	100%
Spatial resolution	Comparison per parameter of the spatial resolution SReds,pa, with the spatial extrapolation potential SEPpa
Spatial density	0.738	1	1
Spatial distribution	15%	0	0
Spatial noise	18%	0	0
Temporal inferred terms
Temporal pertinence	Calculation per parameter of the ratio ((date of analysis d − Datemax,ds,pa)/temporal sensitivity TSpa)
Temporal overlap utility	0	0	0
Temporal resolution	Not applicable (NA)	Not applicable (NA)	Not applicable (NA)
Temporal density	Not applicable (NA)	Not applicable (NA)	Not applicable (NA)
Temporal distribution	Not applicable (NA)	Not applicable (NA)	Not applicable (NA)
Temporal noise	Not applicable (NA)	Not applicable (NA)	Not applicable (NA)

reports the results of calculations required for the quality assessment of (1) spatial inferred scene-related terms and (2) temporal inferred scene-related terms.

We characterize the contribution of the spatial information to the first degradation factor for each parameter informed by each of the retained dataset as reported in

Table 22. Meta-features phase 2—spatial contribution to DF₁ per data source per parameter.

Parameters per Data Source	Spatial Coverage SCds,pa	Spatial Resolution SReds,pa	Spatial Density SDeds,pa	Spatial Distribution SDids,pa	Spatial Noise SNds,pa	DF1a Spatial contribution to DF1
LiDAR point cloud
Density/number of trees	Very high (0)	Sufficient (0)	Medium (−2)	Heterogeneous distribution (−1)	Medium (−1)	(0/3 + 0/2 + (−2)/3 + (−1) + (−1)/3)/5 = −0.4
Height of tree	Very high (0)	Close (−1)	Medium (−2)	Heterogeneous distribution (−1)	Medium (−1)	(0/3 + (−1)/2 + (−2)/3 + (−1) + (−1)/3)/5 = −0.5
Structure of tree crown (width, diameter)	Very high (0)	Close (−1)	Medium (−2)	Heterogeneous distribution (−1)	Medium (−1)	(0/3 + (−1)/2 + (−2)/3 + (−1) + (−1)/3)/5 = −0.5
X–Y distance from a tree to the power line	Very high (0)	Close (−1)	Medium (−2)	Heterogeneous distribution (−1)	Medium (−1)	(0/3 + (−1)/2 + (−2)/3 + (−1) + (−1)/3)/5 = −0.5
Orthophoto based on aerial images
Density/number of trees	Very high (0)	Sufficient (0)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + 0/2 + 0/3 + 0 + 0/3)/5 = 0
Height of tree	Very high (0)	Sufficient (0)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + 0/2 + 0/3 + 0 + 0/3)/5 = 0
Structure of tree crown (width, diameter)	Very high (0)	Sufficient (0)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + 0/2 + 0/3 + 0 + 0/3)/5 = 0
X–Y distance from a tree to the power line	Very high (0)	Sufficient (0)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + 0/2 + 0/3 + 0 + 0/3)/5 = 0
Satellite-based orthophoto
Density/number of trees	Very high (0)	Sufficient (0)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + 0/2 + 0/3 + 0 + 0/3)/5 = 0
Height of tree	Very high (0)	Close (−1)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + (-1)/2 + 0/3 + 0 + 0/3)/5 = −0.1
Structure of tree crown (width, diameter)	Very high (0)	Close (−1)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + (-1)/2 + 0/3 + 0 + 0/3)/5 = −0.1
X–Y distance from a tree to the power line	Very high (0)	Close (−1)	Very high (0)	Homogeneous distribution (0)	Low (0)	(0/3 + (-1)/2 + 0/3 + 0 + 0/3)/5 = −0.1

.

The spatial overlap’s nature justifies further proceeding with a temporal analysis of the degradation factor DF₁.

We characterize the contribution of the temporal information to the first degradation factor for each parameter informed by each of the retained dataset as reported in

Table 23. Meta-features phase 2—temporal contribution to DF₁ per dataset per parameter.

Parameters per Dataset	Temporal Pertinence TPds,pa	Temporal Overlap Utility TOUds,pa	DF1b Temporal Contribution to DF1
LiDAR point cloud
Density/number of trees	Sufficient (0)	Punctual (−1)	(0/2 + (−1))/2 = −0.5
Height of tree	Close (−1)	Punctual (−1)	((−1)/2 + (−1))/2 = −0.75
Structure of tree crown (width, diameter)	Distant (−2)	Punctual (−1)	((−2)/2 + (−1))/2 = −1
X–Y distance from a tree to the power line	Distant (−2)	Punctual (−1)	((−2)/2 + (−1))/2 = −1
Orthophoto based on aerial images
Density/number of trees	Close (−1)	Punctual (−1)	((−1)/2 + (−1))/2 = −0.75
Height of tree	Distant (−2)	Punctual (−1)	((−2)/2 + (−1))/2 = −1
Structure of tree crown (width, diameter)	Distant (−2)	Punctual (−1)	((−2)/2 + (−1))/2 = −1
X–Y distance from a tree to the power line	Distant (−2)	Punctual (−1)	((−2)/2 + (−1))/2 = −1
Satellite-based orthophoto
Density/number of trees	Sufficient (0)	Punctual (−1)	(0/2 + (−1))/2 = −0.5
Height of tree	Sufficient (0)	Punctual (−1)	(0/2 + (−1))/2 = −0.5
Structure of tree crown (width, diameter)	Close (−1)	Punctual (−1)	((−1)/2 + (−1))/2 = −0.75
X–Y distance from a tree to the power line	Close (−1)	Punctual (−1)	((−1)/2 + (−1))/2 = −0.75

. Note that the detailing of the temporal resolution TRe_ds,pa, the temporal density TDe_ds,pa, the temporal distribution TDi_ds,pa, and the temporal noise TN_ds,pa is not reported here because they were identified as not applicable on the basis of the calculation of temporal overlap utility TOU_ds,pa.

We combine DF_1a and DF_1b to calculate the total degradation factor DF₁. This enables calculating the updated potential of knowledge (PK_DF1,ds,pa) per dataset and per parameter, as described in Table 25.

4.3.3. Second Degradation Factor (DF₂)

Metadata relative to the objectives, the authors, and the circumstances of the data acquisition enable calculating the value of the second degradation factor for each of the datasets, as detailed in

Table 24. Meta-features phase 3—DF₂ per dataset.

Terms	Meta-Feature	Dataset (a)—LiDAR Point Cloud	Dataset (b)—Orthophoto Based on Aerial Images	Dataset (c)—Satellite-Based Orthophoto
Audience	Ads	Open access (−3)	-	Field specialists (−1)
BibliographicCitation	BCds	-	-	-
ConformsTo	CTds	-	-	-
Contributor	Cods	-	-	-
Creator	Crds	Valued (−1)	Valued (−1)	Official/authorities (0)
EducationLevel	ELds	Not related (−3)	-	-
HasVersion	HVds	First version (−2)	-	-
IsReferencedBy	IRefBds	-	Official (regulations, standards, recognized journals) (0)	-
IsReplacedBy	IRepBds	Something (−1)	-	-
IsVersionOf	IVOds	-	-	-
Provenance	Prds	Official (regulations, standards, recognized journals) (0)	Official (regulations, standards, recognized journals) (0)	Official (regulations, standards, recognized journals) (0)
Publisher	Puds	-	Valued (−1)	-
References	Refds	Valued (−1)	-	-
Replaces	Repds	Nothing (−1)	-	-
Source	Srds	-	-	-
Modified	Mds	Not original file (−1)	-	-
Valid	Vds	-	-	-
$\begin{array}{c}D{F}_{2, ds}=(\frac{A_{ds}}{4}+\frac{B{C}_{ds}}{2}+\frac{C{T}_{ds}}{2}+\frac{C{o}_{ds}}{2}+\frac{C{r}_{ds}}{2}+\frac{E{L}_{ds}}{3}+\frac{H{V}_{ds}}{2}\\ +\frac{IRef{B}_{ds}}{2}+IRep{B}_{ds}+\frac{IV{O}_{ds}}{2}+\frac{P{r}_{ds}}{2}\\ +\frac{P{u}_{ds}}{2}+\frac{Re{f}_{ds}}{2}+Re{p}_{ds}+\frac{S{r}_{ds}}{2}+M_{ds}\\ +V_{ds})/17\end{array}$		((−3)/4 + (0)/2 + (0)/2 + (0)/2 + (−1)/2 + (−3)/3 + (−2)/2 + (0)/2 + (1) + (0)/2 + (0)/2 + (0)/2 + (−1)/2 + (−1) + (0)/2 + (−1) + (0))/17 = −0.397	((0)/4 + (0)/2 + (0)/2 + (0)/2 + (−1)/2 + (0)/3 + (0)/2 + (0)/2 + (0) + (0)/2 + (0)/2 + (−1)/2 + (0)/2 + (0) + (0)/2 + (0) + (−0))/17 = −0.059	((−1)/4 + (0)/2 + (0)/2 + (0)/2 + (0)/2 + (0)/3 + (0)/2 + (0)/2 + (0) + (0)/2 + (0)/2 + (0)/2 + (0)/2 + (0) + (0)/2 + (0) + (0))/17 = −0.015

Note: Note that an optimistic approach was applied in the absence of metadata, leading all the meta-features to be set to 0 in the absence of information. An alternative pessimistic approach penalizing the absence of information could also have been applied, putting all the scores to the most degraded level. This choice and the effects of this choice on the results are further discussed in Section 5.

.

By applying the degradation factor DF₂ and updating the calculation of the potential of knowledge (PK_{DF1,DF2,ds,pa}) per dataset and per parameter, we obtain a final ranking of the best candidate dataset for each of the considered parameter, as described in

Table 25. Spatiotemporal- and trust-influenced informative potential assessment per datasets per parameter.

	Phase 1	Phase 2		Phase 3
Parameters per Data Source	$DMP{K}_{sr,pa}$	$D{F}_{1,ds,pa}$	$P{K}_{DF1,ds,pa}$	$1-\Vert D{F}_{2,ds}\Vert$	$P{K}_{DF1,DF2,ds,pa}$
LiDAR point cloud
Density/number of trees	1	0.3	0.3	0.6029	0.1809
Height of tree	1	0.125	0.125	0.6029	0.0754
Structure of tree crown (width, diameter)	1	0	0	0.6029	0
X–Y distance from a tree to the power line	1	0	0	0.6029	0
Orthophoto based on aerial images
Density/number of trees	0.875	0.25	0.2188	0.941	0.2058
Height of tree	0.5825	0	0	0.941	0
Structure of tree crown (width, diameter)	1	0	0	0.941	0
X–Y distance from a tree to the power line	0.875	0	0	0.941	0
Satellite-based orthophoto
Density/number of trees	0.665	0.5	0.3325	0.985	0.3275
Height of tree	0.4575	0.45	0.2059	0.985	0.2028
Structure of tree crown (width, diameter)	0.5825	0.225	0.1311	0.985	0.1291
X–Y distance from a tree to the power line	0.665	0.225	0.1496	0.985	0.1474

.

Table 25 details the evolution of the ranking of the best candidate for each parameter based on the application of the 3 phases mentioned in our method.

The coloring (green, yellow, orange) of the scores within each phase as detailed in Table 25 reports the first, second, and third best candidate datasets for the characterization of each parameter. Table 25 indicates that LiDAR-based point clouds initially constitute the most interesting type of dataset for the present case study. Therefore, they should be preferred by default, in the absence of any other type of information relative to file properties.

However, the application of the degradation factors DF₁ and DF₂ leads to a new ranking of the best candidate datasets for each considered parameter. In particular, the satellite-based orthophoto obtains the first position in the ranking for all parameters after the application of DF₁ and keeps this position after application of DF₂. The LiDAR point cloud obtains the second position for informing the parameter “height of tree” after the application of DF₁ and keeps this position after application of DF₂. The orthophoto based on aerial images obtains the third position for the characterization of the parameter “density/number of trees” after the application of DF₁, but is reranked second after application of DF₂.

5. Discussion

5.1. Method Benefits and Contribution for Risk Assessment

The accessibility to many data sources for risk assessment is a recent phenomenon for many application areas. While this consists of a great opportunity for data-driven assessments and DRA, the process for choosing one dataset to inform a parameter in detriment of another was not yet formalized. The method proposed in this paper formalizes this process. One of the main advantages of the method is the use of metadata; the method provides the ranking of the best datasets according to their informative potential without the need for the analyst to open the file and assess its content. In addition to identifying the best dataset for each parameter, the proposed approach also identifies the best parameters per source. Such information can be used as an additional indicator for strategical decision making when deciding about investments related to future data acquisitions.

The results of the method application are, to the best of our knowledge, representative of the reality that power-grid operators can face. The impact of the degradation factors highlights that a dataset originating from a less valued data source may be preferred to a dataset originating from an initially higher-ranked data source, the content of which is outdated or of too low quality. The results also show that, while getting degraded and losing its leading position to inform a specific parameter, a dataset may still remain the best choice for another parameter. This indicates the importance of assessing the informative potential of a dataset on a parameter level rather than as one block.

Moreover, the method provides flexibility for the analyst to achieve the following:

• Tune the meta-features used to calculate DMPK in phase 1, if expertise/follow-up gain of knowledge shows that the initial estimation was not adequate, the initial estimation needed to be updated, or if the evolution in technologies/competencies of stakeholders enables improving the initially obtainable quality of information;
• Adequately maintain elements required for the calculation of the trust-related meta-features by adding, confirming, or removing entities in the lists used for the calculation of the second degradation factor (e.g., new standard or withdraw of a previous standard);
• Have the process iterated over time (even without new datasets) and readapt the ranking of the considered sources if required;
• Modify the weights given to any of the meta-features proposed in phases 1, 2, and 3 on the basis of what one decides to be important or if new risk evidence implies that changes are required;
• Assess the potential of new types of data sources not yet known and integrate the related datasets into the risk analysis by running them through the three phases.

This approach, thus, supports the development and implementation of DRAs by ensuring proper and dynamic dataset selection in an environment with ever-increasing access to more information. It should be noted that the main benefits of the method are to be gained in a long-term application, i.e., when used after the first time, the analyst needs to update it only in case new data sources are considered.

5.2. Limitations and Further Requirements

5.2.1. Reliance on Metadata Format

The proposed method is particularly sensitive to metadata existence and quality. Yet, experience shows that metadata can be in the wrong format (i.e., not respecting existing standards), inexistent, or incomplete, as illustrated by the absence of some objective/author/circumstance-related terms in the case study. Metadata may also be corrupted and contain inaccurate information, due to human actions or inadequate automatic processing. Metadata in the wrong format may lead to more challenging and time-consuming conversions into DC terms, as one would have to work with/generate nonconventional crosswalks. As a partial solution, and assuming that the number of metadata terms is expected to remain limited, one may create lists of “standard-like” terms to convert non-standardized metadata into information usable for the method. In line with this concept, one could also generate further quality indicators of the observed datasets by implementing and counting alarms used to highlight missing metadata. Additionally, it might be strategic to enable one to assign some “privileges” to specific datasets to avoid naïvely degrading the value attributed to datasets due to absence of trust information in the metadata (for instance, for data originating from reference entities such as governmental-based institutions). Such approaches may also be considered when acknowledging that some datasets initially lack metadata but are known to have been generated internally and can, thus, be considered as more “trustable”.

Lastly, two additional approaches—optimistic or pessimistic—may be chosen to face incomplete metadata reporting, as applied in the presented case study. For the former, stakeholders can apply the minimum degradation and generate alerts for manual verification of the final ranking when missing information has been detected. Alternatively, stakeholders may be more conservative and apply the maximum degradation to minimize the possibilities of building on uncertainties. An optimistic approach was chosen for the calculation of the second degradation factor in our work. The pessimistic approach would have led to the same ranking in the present case study because of the metadata we decided to report for those simulated datasets. However, this may usually not be the case. The choice of the strategy to follow is a more subjective task that, thus, needs to be addressed on a case-by-case analysis and is hardly generalizable in the generic description of our approach.

5.2.2. Three-Phases Method Elements

The characterization of the properties in phases 1, 2, and 3 of the method mainly consists of the definition of classes and class boundaries relative to each property, according to the authors’ experience. Yet, the method provides sufficient flexibility for the boundaries to be adapted if needed, especially in a context other than power-grid management. The choices made in the definition of the three-phases method may, thus, be seen as an illustration of a general guideline that can be adapted to the context in which it is applied rather than as a rigid formulation.

The weights associated with the meta-features of each phase are likely to be different from application to application or from organization to organization. Although the approach is already implementable as is for any project where the requirements related to the learning phases are fulfilled, another implementation would require first defining the value of the reported weights. Different approaches may be considered for this purpose, and one may, for instance, tackle this problem as a meta-learning task. Alternatively, one could use more straightforward solutions such as the application of Zip’s law, as applied in other work [92].

We chose a normalized weighted sum to summarize the results of all the properties used for the definition of the DMPK and DFs. This choice is considered robust, well-known, and straightforward [93,94], and it was adopted for a first formulation of the proposed methodology. Multicriteria decision-making methods could also be considered, for instance, if further dependencies between properties were to be considered in future applications of the method. Furthermore, the approach considers, right now, all datasets to be independent and analyzed independently and not leveraging one another. Future extensions should address the existence of links between datasets.

The implementation of the method is more labor-intensive in the first iterations of the process, especially when the knowledge of the involved stakeholders needs to be converted into information exploitable for the use of the presented method. Verifications required after detecting new terms in the lists used to calculate the second degradation factor may be particularly time-consuming. However, this workload and the general need for manual verifications are expected to diminish over time as the number of processed datasets increases, facilitating future automatic processing.

6. Conclusions

Despite being synonymous with considerable advantages, increasing data availability related to energy systems also implies numerous challenges from a data management perspective. In the present work, we suggest an approach enabling to tackle such challenges in the risk analysis field by analyzing features that may degrade the information potential of a dataset compared to a baseline (default maximum potential of knowledge (DMPK)). The key contributions of this research are twofold. First, it proposes a method for reinforcing data-related risk analysis steps. The use of this method ensures that risk analysts can methodically identify and assess the available data for informing the risk analysis key parameters. Second, it develops a method for selecting the best datasets according to their informative potential. The method, thus, formalizes, in a traceable and reproducible manner, the process for choosing one dataset to inform a parameter in detriment of another, which can lead to more accurate risk analyses.

The application of the method to vegetation-related risk assessment in power grids shows that the approach enables dynamically selecting the best information possible. It also demonstrates that a dataset originating from an initially less valued data source (the satellite-based orthophoto) may be preferred to a dataset originating from a higher-ranked data source, the content of which is outdated or of too low quality (LiDAR-based point clouds).

The proposed method reduces approximations and uncertainties in risk analysis by improving tracing of information and measurement characterization, thereby supporting the optimization of dataset preparation and integration for more efficient risk assessments. The generic aspect of the approach opens up further applications, such as management of safety barriers or risk management applied in other areas such as finance or insurance, which can also access large data lakes. The method is particularly valuable for the risk assessment of large-scale “open” systems, subject to environmental changes and presenting increasing opportunities for data capture. Such systems include not only power grids, but also transportation systems and gas pipelines. In addition to pressure, temperature, and flow sensors that might identify a failure after a leak is already established, external pipelines can benefit from data captured by autonomous systems (drones) or satellites for identifying a growing crack that can potentially result in a leak.

Lastly, although the exploitation of natural language processing techniques may in the future further support the application of the method (i.e., via advanced text-based quality assessment of metadata), there is an undiscussable need for proper and better metadata registration and maintenance. The community should encourage good practices in this direction, and our work can be used to illustrate the potential benefits of doing so.