A top-down approach for a multi-scale identification of risk areas in infrastructures: particularization in a case study on road safety

Transport infrastructures are important components in modern societies and in developed countries to ensure communication and mobility of people and goods, thus favoring social and economic growth and development.

Focusing only on rail and road transport infrastructure, Europe has 79,142 km of motorways (year 2018) with Spain the country with the greatest length with 15,585 km. In terms of railway, Europe has 231,284 km (year 2019) with Spain the fourth European country with 21,988 km of railways [1].

Both rail and road infrastructures play a very important role, with both road and rail being the most used means of transport by travelers, well ahead of maritime and air transport [2]. For this reason, the development of reliable and resilient infrastructures is essential, being one of the Sustainable Development Goals (SDG) for UN’s 2030 Agenda [3]. Within the current societal context, including SDGs in a challenging environmental, economic, and political situation, a new concept for resilience-oriented maintenance that efficiently maximizes transport infrastructure performance is mandatory.

The assessment of risks, considered as the combination of the probability of occurrence of an event and the magnitude of its consequences, is crucial to improving infrastructure resilience [4]. Risk assessment is a general process for establishing the context of the infrastructure, the identification of hazards, and risk appreciation [4], which are addressed more in detail in the following paragraphs.

Regarding the contextualization of infrastructures, this article will focus on road and rail infrastructures in Galicia (located in northwest Spain). These infrastructures will be considered from a global-to-specific point of view in a multiscale approach. For this purpose, the terms network, system, object, component, and element are defined as follows:

Hazards are defined as a potential source of undesirable consequences [4]. In terms of safety, in the last decade (2010–2019), 18,419 people have died in road traffic accidents in Spain, 1503 of them in Galicia [7]. As for railways, the number of fatalities in Spain was 334 for the same period [1]. Several natural hazards are considered in Europe, such as among others: heat waves, heavy precipitation, river floods, windstorms, landslides, droughts, forest fires, and avalanches. However, most of these hazards are influenced by other hazards. Forest fires are influenced by heat waves and droughts, and landslides are influenced by heat waves, heavy precipitation, river floods, windstorms, and droughts [8]. Both fires and landslides are hazards that severely affect safety in road infrastructures and railways [9‐12]. Accordingly, we focus on fire, landslides, and safety.

Finally, risk appreciation aims to provide evidence-based data and information to make decisions on how to deal with risk. Risk appreciation is divided into three phases: identification, analysis, and evaluation of the risk. Risk identification includes how risks are discovered, recognized, and recorded. Risk analysis is the foundation for understanding the risk, providing inputs for risk appreciation and decision-making. Lastly, risk evaluation consists of the comparison of risk levels under the defined criteria [13].

The purpose of this work is to determine which areas of road and rail infrastructures are subject to the aforementioned hazards. After these risk areas are identified, their prioritization is carried out to select the riskiest one to be studied in detail. For that purpose, we define and validate a multiscale methodology developed in three steps, accounting for the considered scales with a top-down approach. The core of the work consists of identifying risky areas in road infrastructures based on proven and standardized procedures using only information in the public domain. The main novelty of the proposed methodology consists of the multiscale procedure that ranges from a global risk identification at network-level to a site-specific object-level risk assessment, with a particular focus in this case study on road safety, in which a risk map has been defined by quantifying the safety factors related to the infrastructure. These factors have been weighted through expert opinion, obtaining a map that allows updating the risk zones according to changes in the reality observed through access to public databases, which is relevant in situations of climate change. This methodology allows an individualized analysis of the risk factors to estimate which are more relevant and to be able to act in a specific and preventive way. Obtaining the risk map can be done with real-time information, being the basis of a simulation tool for a digital twin focused on the analysis of road behavior in terms of road safety.

In this section, we describe the literature related to the objectives of our work. This review is carried out in a top-down fashion starting with that works focused on multiscale approaches, describing their limitations, and remarking on the differences with our proposal. Then, we progress with the literature related to the risk assessment in our scenario, regarding road safety issues, multi-criteria methods for decision-making and, to conclude, the illustration of the results as risk maps.

The top-down multiscale approach is one of the novelties of the methodology proposed in this paper. Some works related to this multiscale approach are mentioned as follows. Berres et al. [14] presented techniques for the exploration of interconnected traffic dynamics at intersections and highways. These techniques are based on sensors at different scales, which were named microscale, mesoscale, and macroscale. Considering the multiscale approach and also in a road safety context is the work of Thorisson and Lambert [15], in which they integrated road safety metrics of road segments in re-scalable straight-line diagrams. They identified the road segment under stress by searching for one or more metrics that are outliers concerning the contextual data. In the work of Achillopoulou et al. [16] provided the link between the components of multiple hazard resilience assessment in transport infrastructures based on a variety of Structural Health Monitoring, considering three scales: components, assets, and networks. The work carried out by Dragićević et al. [17] develops a multiscale analysis and fuzzy sets combined with GIS-based multicriteria evaluation to determine landslide susceptibility for regional, municipal and local scales at resolutions of 50, 10 and 1 m respectively. Similar to the previous work, Bernardo et al. [18] also treated the problem of landslides. They created a map of susceptibility with which identified the areas of the road most exposed to landslide and they focused on local monitoring of those parts identified.

Many of the previous works mentioned [14, 15, 17, 18] only consider one hazard (such as road safety or landslide). The only one that considers several risks [16] does not have the same approach as this work because it does not treat the hazards in a geolocated way to identify risk areas. However, it is a good example of the different scale that they consider in an infrastructure. In addition of consider several hazards, in the present work, such hazards are studied from a general scale to a site-specific scale, identifying and prioritizing both risk areas (general scale) and factors affecting road safety (site scale). None of the previous works has studied the hazards of this method. Another aspect that differentiates this work from previous ones is the identification and prioritization that has been done in a Geographic Information System (GIS), which allows the creation of risk maps in the last step.

In our specific case study, the risk analyzed in detail is that of road safety. Multi-Criteria Decision-Making (MCDM) models are used to weigh the factors affecting this risk. MCDM are models that analyze many conflicting criteria in decision-making, thus aiding and improving the decision-making process. These models were used in numerous works for the selection and weighting of the most important road safety factors, and the most used method is the Analytical Hierarchy Process (AHP). The most outstanding works that use AHP in road safety aspects are the following.

Seven causes of accidents were selected by Nanda & Singh [19] and created a table indicating the number of accidents associated with each factor. Then, using AHP, they obtained a weighting for each factor. With the weighting and the number of accidents, they generated a ranking of the states with the highest accident rate. In the case of the work carried out by Keymanesh et al. [20], they identified nine factors that contribute to accidents on an entire road. They then divided a road into 8 sections and selected potential black spots in each one. For each section, the most important factors and the most dangerous potential black spots were weighted by five experts using AHP. Both the identification and the prioritization of black spots were compared with the data collected from accidents by the police.

The following articles should also be highlighted, in which a very similar methodology was followed in all of them. Cheng et al. [21] classified road factors by taking into account the factors of driver, vehicle, road, and environment. They also defined their subfactors and obtained the weighting with the AHP. Farooq et al. [22] selected 20 driver behavior factors that have a critical impact on road safety and weights them in a three-level hierarchical structure using AHP. Sordyl [23] performed a hierarchy of the road traffic safety factors in a general way under the levels: driver, vehicle, and environment. It used the AHP to obtain the weighting of these factors.

The AHP not only stands out for being the most used individually but also when combined with other methods, such as an integrated multi-criteria decision-making model combining AHP and Best Worst Method (BWM) to weight driver behavior factors according to Moslem et al. [24].

In all the above-mentioned works, many different road safety factors have been considered and prioritized through different MCDM [19, 21‐24]. However, these works neither validate the results with real data nor generate risk maps.

One of the previous works that compare its MCDM with real data is the work carried out by Keymanesh et al. [20]. However, they first identified the hot spots and then prioritized them considering the factors that affect by AHP. The last step was comparing their results with the data from the police, but they do not generate a risk map. In contrast to this work, we select the factors that influence road accidents, weighting them and then we obtain the segments with higher risk.

This section shows the results obtained by applying the proposed methodology. As the methodology has been divided and explained in three steps, the results are also shown in this way.

This section analyses and discusses the results obtained. First, the risk map obtained is compared with historical accident data [53] and validated by performing an exact Binomial test. In this way, the results obtained in steps 2 and 3, which refer to the system and object level respectively, are discussed. The results obtained in step 1 at the network level and the proposed methodology, in general, are discussed at the end of this section concerning previous works.

The risk map obtained has been validated with real accident data providing the location of the accident with an accuracy of 100 m, between the years 2016 and 2019. It should be noted that these real accident data come from another source and have a different scale than those used in network level 1 and shown in Table 2. This comparison supports the validation of the risk map derived through the multi-criteria decision methodology based on AHP and BWM. The hazard index (Haz) [45] is calculated using Eq. 2.

Though Haz is usually calculated for longer segments of roads, in our case it is derived for 100-m segments to be directly compared to the risk map. With the hazard indices obtained for both the increasing and decreasing directions, 5 equal intervals are defined for visualization, where the higher the index, the greater the hazard. The three highest levels of the hazard index are shown in Figs. 8 and 9 with the risk map previously drawn up for both directions. A zoom-in of the most critical points is also shown.

There were not any sections with values 1 (corresponding to “no risk”) or 5 (corresponding to “extreme risk”). As the values obtained were classified into these levels with linear interpolation, there are segments with intermediate colors, and so their risk was between these two values.

As can be seen in Figs. 8 and 9, maps show a good correlation, as points with a higher danger index coincided with high-risk areas. This good correlation can be seen in zoom 3 in the increasing direction (Fig. 8), in one point of zoom 1, and in one point of zoom 4 in the decreasing direction (Fig. 9). In these three examples, the hazard index points had a moderate level and corresponded to medium risk on the map. The same happens in zoom 3 of the decreasing direction (Fig. 9), but in this case, the hazard index point was very high, and the risk was extreme.

Cases with a lower hazard index (or simply no accidents) than the risk index cannot be considered incorrect, as this indicates that these segments present some potential risk factors. Although no accidents have occurred yet, they should be considered because, in the case of future changes of different nature (traffic flow increase, new types of vehicles, climate change, aging of the infrastructure, etc.), these sections may reach high Hazard Indexes. Therefore, predictive maintenance policies and early decision-making should be followed, with special attention to these segments, before accidents occur. An example of this was two points in zoom 1 and one point in zoom 2 in the increasing direction (Fig. 8), and all points in zoom 2 in the decreasing direction (Fig. 9). However, in the case of a high hazard index and a low-risk index, this is considered an incorrect result, as all points with a high hazard index should be perfectly identified within the risk map. This occurred only for 3 points: one point in zoom 1 in the increasing direction (Fig. 8), one point of zoom 1, and one point of zoom 4 in the decreasing direction (Fig. 9). In the example of the increasing direction, the calculated risk was high, and the hazard index was very high. In both examples of decreasing direction, the calculated risk was moderate and hazard indexes were high.

Figure 10 shows the risk map for the increasing and decreasing direction with all the points of the hazard index with which the validation is done.

Criteria followed for the validation were the following to be on the side of security:

The road was divided into segments of 100 m, which resulted in 170 segments for the decreasing and the same for the increasing direction. Therefore, the risk was calculated for 340 segments. However, real data was only provided for 153 of these segments, because there were not any accidents in the rest of the segments, and the hazard index was very low. With this, the analysis of the results was derived considering only the segments with real data.

To know whether the segments were correctly or incorrectly classified, the assessment takes into account that these results follow a binomial distribution where the outcomes are Boolean-valued: success or failure. According to the results, we tested if this methodology had a probability of correctly classifying the segments with a probability greater than 95%, with a hypothesis contrast where the statements are (Eq. 3):

The hypothesis established in Eq. 3 was calculated with an exact Binomial test in the software R with a 5% significance level. The results are shown in Table 11.

As can be seen in Table 10, the number of segments classified correctly was 150 out of 153, so with these data, the probability of correct classification was higher than 98%. However, with the established hypothesis, we tested if the probability of correct classification was higher than 95%. The results show that the p-value was less than 0.05 (5% significance level), resulting in the rejection of the null hypothesis, and the acceptance of the alternative hypothesis. This concluded that the methodology correctness was greater than 95%.

The advantages of the proposed methodology over alternative methods reviewed in the related work are explained in detail below.

If we focus on discussing our results compared to results obtained in previous works, we can say the top-down multi-scale methodology shows results at the three scales: location and prioritization of the hot spots at the network level, weighing of the safety factor at the system level and risk map creation at the object level. None of the previous works [14‐18] related to this multiscale approach, showed results for the different scales and none considered together the hazards that are taken into account in the present work.

Regarding the MCDM and validation of the result that is developed in parts 2 and 3 of our methodology, the works revised [19, 21‐24, 54] used some MCDM but they do not validate their results with real data. In the work carried out by Keymanesh et al. [20] despite using the AHP method and comparing it with real data, they do not generate a risk map. The work carried out by Fuller et al. [55] does generate a risk map and compares it with real data, but they only take into account four road safety factors and use a multi-criteria evaluation method that does not take into account the weighting of either experts or road users. The work of Driss et al. [56] validates the generated map, but they did not obtain it using MCDM. In the work carried out by Hu et al. [57], in addition to not using any MCDM, they only compared the generated map with field-based interviews and not with real accident data.

In this work, a multiscale methodology with 3 steps has been proposed and carried out. The first step consists of identifying the hot spots at the network level, the second step consists of identifying and prioritizing the factors that affect the road safety of the system, and the last step is the creation and validation of a risk map.

For the identification of risky areas, landslide, fire, and safety hazards have been considered. Once obtained the hot spots, they were prioritized to obtain the area with the highest risk, where the second step was applied. The risk factor in this area was safety in road infrastructure, so the factors affecting road safety were analyzed. Only factors related to traffic were considered, leaving other factors which could affect road safety, such as human or vehicle factors, unanalyzed. These factors were classified into different levels and weighted using a Multi-Criteria Decision Method. In this case, the Analytical Hierarchy Process (AHP) and the Best Worst Method (BWM) were applied, and the opinions of 67 participants were obtained through surveys.

Once the results of the weighting of factors were obtained, in the third step two risk maps were created with these factors: one for the increasing and one for the decreasing direction. The resulting risk maps were compared with the hazard index of real accidents on the road to validate the risk map and the methodology applied. This validation has concluded that the methodology has a probability of success in classification greater than 95%. In this way, we obtain a risk map based on the safety factors related to the infrastructure. Thanks to the weighting of factors obtained with the methodology, we have quantified each 100-m segment according to the risk presented by these factors. Validation with real data shows that these infrastructure-related safety factors do affect the occurrence of accidents. However, if we only obtained the risk map with these real data, we would not have quantified the influencing factors. Therefore, preventive action can be taken on those factors that have the greatest influence on the risk.

In general conclusion, this methodology shows a good overall result as the area identified as the riskiest has a probability of successful classification higher than 95%. All this methodology is developed with a multiscale top-down approach and in as standardized a way as possible since numerous road and risk assessment standards are followed. The fact that the information is in the public domain means that the methodology developed here is applicable to any environment (at least in Spain), and its application is easily automated so that it is quick to apply as the public data used hardly needs any processing. The advantage over previous works is that this methodology presents results for the three scales that have been considered in this multi-scale approach and it is suitable for analyzing a transport infrastructure, from a network completeness point of view to the quantification of safety factors in 100-m sections to obtain a risk map. This methodology allows an individualized analysis of the risk factors to estimate which are most relevant, and to act in a specific and preventive manner. Obtaining the risk map from these factors could be done in real-time. For this reason, the proposed methodology can be the basis for a simulation tool of a Digital Twin focused on the analysis of road behavior in terms of road safety. A weakness of this methodology is given by its application in the case study of road safety on the A-55 motorway. In this case, the hierarchy of factors can be applicated to other segments of the motorway, but not the weightings obtained since these weights are particular to this road due to its special characteristics.