Data and reports on Nat. Hazards show an increasing frequency of such disasters worldwide [1
]. In addition, the number of people affected and the economic losses resulting from natural disasters have increased worldwide except in some developed countries, which have experienced a decreasing trend since 1900 [2
]. Globally, communities in general are facing a wide range of natural hazards. These hazards cannot be eliminated totally, but their effects can be minimized by applying new methods and providing data support for the decision-making process [4
Mountainous areas are characterized as social-ecological systems and, because of their geotechnical and hydrological characteristics, are associated with elements of danger. For example, mountainous regions are relatively active hydrologically and geophysically and there is considerable topography-driven variation in vegetation, moisture, and energy [6
]. Mountainous environments commonly experience a wide range of natural disasters, such as avalanches [8
], landslides [11
], floods [5
], debris flows [19
], soil erosion [22
], and rockfalls [25
]. Therefore, planners need hazard susceptibility maps to cope with natural disasters in mountainous areas, particularly flash floods, avalanches, and rockfalls [8
]. These three most common hazards can damage transportation systems, property, and danger human life in risk areas [28
]. However, even in areas with a wide range of natural risks, previous studies have mainly focused on one type of hazard [8
]. As a result, integrated multi-hazard analysis has been under-emphasized in natural disaster management and planning. More importantly, most previous studies have focused solely on forecasting and controlling Nat. Hazards [33
], neglecting the human effects and the exposure component [4
]. In fact, the exposure of mountain social–ecological systems to geohazards has increased because of a combination of the dynamic bio-geophysical environment and the challenges of land use planning [34
]. Separate spatial studies of different risk types in mountainous regions could in fact result in greater underestimation of risks in total. A multi-risk assessment technique could thus be very useful for considering the interactive reactions of different hazards on surrounding areas [36
Multi-hazard exposure mapping is an important key tool in enhancing resilience in the context of Nat. Hazards in mountain areas and can play an important role in ensuring the long-term sustainability of local ecosystems and communities [38
]. Resilient social–ecological systems have the ability to adapt and utilize relevant knowledge in order to mitigate natural disasters by self-organizing and developing safe infrastructures based on nature-based solution [41
]. Fortunately, there is growing recognition of the importance of multi-hazard exposure analysis, and hazard-based attitudes and policies are slowly changing to risk-based management in different parts of the world [1
]. In an early study, Bell and Glader [4
] developed a general method for analyzing multi-hazards in mountainous areas for multiple processes, focusing on avalanches, rockfalls, and debris flow hazards. In a later study, Gruber and Mergili [36
] performed a regional-scale analysis of multi-hazards and risk indicators in high mountainous regions of Tajikistan. Schmidt et al. [45
] introduced a generic framework for multi-risk modeling developed in New Zealand and developed a prototype software that is capable of “plugging in” various natural hazards. Recently, Emmer [24
] provided a review on natural disasters worldwide and found that research on hydro-meteorological disasters prevails over the research on geological and geomorphic disasters. However, their framework does not provide any guidance on producing a single natural hazard map.
Avalanches, rockfalls, and floods frequently occur in mountainous regions of Iran and cause severe damage to buildings, people, and natural environments [11
]. This study investigates multi-hazard exposure using different machine learning approaches due to the following reasons: (1) Machine learning (ML) is a subfield of artificial intelligence where models can learn and improve themselves based on historical events; (2) ML models can easily identify trends and patterns in a large volume of data and involve continuous improvement during operation, which lets them make better decisions [48
]; (3) they are capable of handling data that are multi-dimensional and multi-variety [30
]; and (4) they can directly extract knowledge of natural disaster processes based on previous disaster occurrences and geo-environmental factors without human intervention, thus, they do not need experts’ experiences and judgements to determine the importance of predictive variables [49
]. To evaluate this framework in an area with multiple hazards, the Asara watershed (Iran) was selected because it has become considerably disaster-prone in the past decade and has experienced a disproportionately high number of catastrophic natural disasters. In work to develop a global methodology for multi-hazard exposure mapping in mountainous regions based on historical disaster events, the novelty of this study lies in constructing an integrated framework based on different state-of-the-art machine learning algorithms and using several morphometric, geo-environmental, and topo-hydrological factors for a multi-hazard exposure analysis.
Specific objectives of the study were to: (1) Produce an exposure map based on proximity to residential areas, main roads, rural roads, and power transmission lines; (2) apply three advanced machine learning models including BRT (Boosted Regression Trees), GAM (Generalized Additive Model), and SVM (Support Vector Machine) to produce a multi-hazard map (avalanche, flood, and rockfall hazards); (3) evaluate the performance of each model using threshold-dependent and threshold-independent methods; and (4) produce an integrated multi-hazard exposure map. The intention was for the practical framework of multi-hazard exposure mapping developed in the study to be transferable to other mountainous areas around the world.
2. Study Area
The study area, the Asara watershed, is in Alborz province, which is located in the Alborz mountains of Iran (Figure 1
). The watershed occupies an area of 1094.9 km2
, with a range of elevation from 1332 m.a.s.l at the watershed outlet to 4323 m.a.s.l at the headwaters. Mean annual precipitation in the study area is 265 mm, of which 40% falls as snow. Chalous Road, which is one of the most important transfer lines to the north of Iran, passes through the Asara watershed with 79.54 km. Traffic on the road is generally very heavy, especially on the weekends and on state holidays.
In a land use perspective, the watershed is almost entirely covered by rangeland (95%), but it has 46 villages with around 5000 residents and many restaurants (near the main road). However, only 0.33% of the watershed is residential area. In this mountainous watershed, there are markedly high frequencies and magnitudes of natural hazards, but the level of exposure and risk are also increasing due to socio-economic changes such as the increase of population [50
]. Some rivers and watercourses in this semiarid region generate flashy stormflows that rapidly peak following brief and intense storms which cause flashy floods in the area. The geological structure and geomorphological characteristics of the study area is considerably complicated which often prone to rockfall and snow avalanche events occurring [52
The results of this study indicated that the predictive ability of models in Nat. Hazards is highly affected by the type of the hazard. In this study, we found that SVM was the most accurate model for analyzing avalanche and rockfall hazards in the Asara watershed, while BRT had the highest accuracy for spatially predicting flood hazard. Therefore, a machine learning model that has a high predictive performance for a natural disaster (e.g., snow avalanche) may showed a lower performance for another natural disaster (e.g., flood). The SVM-based classifiers automatically determine classification areas by statistical analysis of observed training datasets. In addition, the SVM model does not offer direct access to classification borders and thus it is not easy to check the classification, so users should carefully select positive and negative events. The results obtained in the present study confirm previous findings on the ability of SVM to predict avalanche hazards [73
] and rockfall hazards [25
]. The results of this study confirm previous findings that BRT is one of the most accurate models for mapping flood-prone areas [57
]. Tree-based models automatically fit non-linear relationships to flood area input factors [75
]. Therefore, since Nat. Hazards present a high level of complexity, especially in a complex system such as a mountainous watershed, multi-hazard analysis should be undertaken using different advanced machine learning techniques [17
]. ML models are critical in helping landscape managers to understand the complex factors controlling Nat. Hazards and more efficiently plan the mitigation measures [71
Production of an exposure map based on AHP was one of the important steps in our multi-hazard exposure mapping approach. Analytical hierarchy process offers a flexible, step-by-step, transparent way to combine factors that play an important role for exposure in different dimensions [77
The present study identified the possible high-risk areas in one of the most important mountainous watersheds (in terms of travel and traffic) in Iran. In general, the results should not be viewed as definite risk maps, but should be used as basis for planning risk management and implementing ecological engineering measures to the high-risk areas. The multi-hazard exposure map (Figure 10
) showed that areas associated with a high and very high levels of risk represented a very small proportion of the Asara watershed. However, the areas with high and very high risk levels were mostly located in the valleys which are also the concentration points of infrastructure and population.
These are the areas where the preventive actions should be prioritized. In case of floods, it is necessary to protect lower lands that are exposed to flood hazards by constructing impounding dams upstream of the larger villages [79
]. The dams will not only prevent flooding but also provide water during the dry season. For villages that are located headwaters and are less exposed to flood risk, a construction of flood protection walls should reduce the risk considerably. To reduce the rock-fall risk, structural protection measures in the areas with high risk, existing infrastructure, and residential areas need to be implemented. These measures can include the removal of unstable rock elements, slope reshaping or stabilization of unstable rock masses by the use of vegetation or steel wire mesh. New buildings should be carefully planned in all high risk multi-hazard areas. They should be sited and designed according to the character of the local landscape and very high risk areas should be avoided [80
]. Moreover, all aforementioned engineering measures should be accompanied with raising the people’s awareness in the high risk areas.