The role of susceptibility, exposure and vulnerability as drivers of flood disaster risk at the parish level

Fluvial flooding continues to be a process that has a major impact on society, the environment and the economy. Although its natural triggering factors, the spatial configuration of exposure and vulnerability is expected to play a relevant role in explaining the damage records. The starting point of this research is the use of existing flood susceptibility, exposure and social vulnerability mapping, produced at the parish level, as input data in a Classification and Regression Trees’ (CART) model. Two models were ran, autonomously, that use two databases of flood damage as dependent variables: one including the human damages (fatalities, missing, injured, displaced and evacuated people) from flood events—the DISASTER database; another one that sums the DISASTER cases and the lower impact damages (damages to roads, railroads and buildings). The results show a quite distinct classification of parishes, whether one database is used or the other. The DISASTER database reveals susceptibility as the most relevant flood risk driver in explaining the damage patterns, while the database with all the flood cases identifies exposure as the more relevant driver. In the end, the degree of damages as documented in databases is conditioned by the geographical distribution and overlay configuration of the three flood risk drivers. Finally, the CART classification groups are analyzed at the light of the European Union’s Floods Directive areas of significant potential flood risk. This analysis showed that the Directive’s parishes are interpreted differently—in terms of their positioning in face of the risk drivers—which is explained by the use of distinct impacting-criteria in the construction of the flood damage databases.


Introduction
The recent report on human losses from disasters caused by natural hazards, based on the EM-DAT data from 2000 to 2019 (CRED & UNDRR 2020) includes flooding within the hydrological hazards-type, along with landslides and wave action. Floods alone were responsible for 44% of all occurrences, 41% of affected people, 9% of casualties (representing 104,614 persons) and 22% of the economic losses (US$ 651 billion) recorded in the EM-DAT (CRED and UNDRR 2020).
Natural and human environments, as in many other risks, interact significantly to generate flood risk (Zischg et al. 2018). The spatial relation of these environments is usually expressed by hazard, exposure and vulnerability with emphasis on social vulnerability (Koks et al. 2015;Santos et al. 2020). Several approaches are based on the traditional and current understanding of the risk concept (UNDRR 2019) and define risk as a product of hazard, exposure and This article is part of a Topical Collection in Environmental Earth Sciences on Earth Surface Processes and Environment in a Changing World: Sustainability, Climate Change and Society, guest edited by Alberto Gomes, Horácio García, Aejandro Gomez, Helder I. Chaminé. vulnerability. Other approaches perform a statistical analysis of the patterns of flood damages along with the characteristics of the hazard and exposure (Mazzoleni et al. 2020). Independently of the approach, databases of past floods and respective damages are essential in the calibration and validation process of flood hazard and risk models (e.g., Khosravi et al. 2019;Li et al. 2012;Termeh et al. 2018;Santos et al. 2019).
An increasingly number of global flood risk indexes has been produced by academic institutions and the business sector. The calculation of such indexes is taking advantage of the high availability of Earth Observation products and cloud computing capacity, producing static risk assessments (Assteerawatt et al. 2016;Phongsapan et al. 2019;Ward et al. 2013Ward et al. , 2015Wing et al. 2018) or near-real time updated risk assessments (Dottori et al. 2017;Todini 1999). Resolution and diversity of input data representing exposure and vulnerability varies significantly among the available risk models according to the scale of the analysis.
Flood risk indexes have been designed worldwide to support risk characterization, analysis and management. In the Portuguese context, a static flood risk index was recently proposed at the municipality level, in support of strategic levels of decision in flood risk management (Santos et al. 2020). Subsequently, the research team involved in that work identified the need to explore deeper the relation between risk drivers-susceptibility, exposure and vulnerabilityand historical records using smaller units of analysis like the parish. Such approach would be implemented using Geographic Information Systems (GIS) and machine learning.
The mix of GIS and other techniques like multivariate statistics, multicriteria analysis, physically based and machine learning models is recognized to be appropriate to flood analysis and modeling (Arabameri et al. 2020;. In this context, several methods have been used. Regarding statistically-based methods, one can highlight the use of bivariate methods as the weights of evidence (Tehrany et al. 2014) and the frequency ratio (Samanta et al. 2018;Siahkamari et al. 2018). Other data-driven methods include the entropy index (Hong et al. 2018), the statistical index (Khosravi et al. 2016), the evidential belief function (Tien ), logistic regression (Ali et al. 2020) and the k-nearest neighbor (Costache et al. 2020a, b, c). However, these procedures greatly rely on the relation between dependent and independent variables and are heavily influenced by the size of datasets (McLay et al. 2001).
As for the multicriteria analysis to evaluate flood hazard and risk, several methods have also been used (Souissi et al. 2020), like the analytical network process (Cao et al. 2016), the analytical hierarchy process (Ghosh and Kar 2018;Luu et al. 2018;Tang et al. 2018), the simple additive weighting ) and the technique for order preference by similarity to ideal solution ). These approaches rely on expert knowledge that can be twisted by confuse rulings and ambiguity (Miles and Snow 1984).
As far as we know, within the diverse decision tree models, like Random Forest (RF), Quick Unbiased and Efficient Statistic Tree, Classification and Regression Trees (CART) and Chi-squared Automatic Interaction Detection (CHAID), only RF (Lee et al. 2017) and CHAID (Tehrany et al. 2013) have been applied until now in flood analysis and modeling.
In this work, the CART algorithm is used because this method has proved to work well on procedures with nonlinear behavior and substantial inner heterogeneity (Ji et al. 2013). The CART has a number of proficiencies, like the insensitivity to outliers and data spatial distribution, the capability of integrating both categorical and continuous variables in the model and the ability of using several trees to characterize the modeling processes .
The main objective of this paper is to understand the drivers of flood disaster risk at the parish level in the Northern region of Portugal. This study has the following specific objectives: a) To identify the role of susceptibility, exposure and vulnerability, as the main drivers of flood risk, in justifying human losses and damages caused by floods in the XXI century (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015), suggesting a classification of parishes based on those drivers' role; b) To discuss the selection of the European Union's Floods Directive risk areas in the study area, according to the dominant disaster flood driving forcers of each parish.

Study area
The study area corresponds to the Portuguese NUTS II Northern region (Nomenclature of Territorial Units for Statistics, an European Union administrative and statistical division). The Northern region has an area of 21 287 km 2 and 3 689 682 inhabitants. It includes 86 municipalities and 2028 parishes at the 2011 Census date. After the administrative reorganization performed in 2012, the number of parishes dropped to 1426, which represents a loss in geographic detail. A parish is defined as the smallest administrative union with elected authorities and its delimitation is used for censitarian purposes as well. Considering the 2028 parishes, the average size is 10.5 km 2 . A municipality is composed of several parishes (ranging from 1 to 89 parishes, depending on the municipalities' history, size and urban development, but with an average of 23 parishes by municipality). The map in Fig. 1 also shows the six areas of significant potential flood risk (ASPFR) following the 1st cycle of the Floods Directive (EU Directive 2007/60/EC). The study area is dominated by mountains in the West (with elevation above 1000 m), plateaus in the East, a narrow coastal platform, tectonic depressions and deeply incised valleys with steep slopes (Fig. 2). Low permeability rocks (granites and metamorphic rocks) dominate, often highly fractured and covered by weathered materials resulting from chemical weathering.
Regarding the basins of the main rivers that drain into the Northern region, three of them have their headwaters in Spain-Minho (9091 km 2 , of which 8276 km 2 in Portugal), Lima (2522 km 2 , of which 1199 km 2 in Portugal) and Douro (97 478 km 2 , of which 18 588 km 2 in Portugal). The other main drainage basins are fully located in Portugal-Cávado (1699 km 2 ), Ave (1391 km 2 ) and Leça (185 km 2 ) (Fig. 2). The river regime is controlled by the rainfall seasonal patterns.
According to the Köppen-Geiger-Pohl climate zones' classification, two subtypes of temperate climate can be found in the study area: temperate with dry or temperate summer (Csb) in the western and central areas; and temperate with dry or hot summer (Csa) in the inner eastern area. The climate in the western area is largely influenced by the proximity to the Atlantic Ocean, while in the east it features a more pronounced continental facies. The mean annual precipitation (MAP) is higher (2000-3500 mm) in the western mountains, mainly oriented parallel to the coastline, thus blocking the moist westerly winds blowing from the North Atlantic, while the eastern Douro river valley is one of the driest regions in the country (MAP = 300-500 mm). Precipitation concentrates during the autumn and winter seasons and the summer drought typically lasts for three months (June, July and August). Up to 50% of the total rainfall on winter (D-J-F months) is due to cyclonic and directional W and SW atmospheric rivers (Ramos et al. 2015). During the summer months rainfall is usually associated with N and E flows, although local factors such as relief and deep convective depressions are relevant in conditioning the spatial variability of intense rainfall events (Ramos et al. 2014). Spring and autumn months are transitional months and total rainfall can vary significantly (Gallego et al. 2011;Miranda et al. 2002;Trigo and DaCamara 2000).
River floods in the study area have been occurring in large rivers with large drainage basins, mainly triggered by precipitation periods lasting for several days (Fig. 2, Pereira et al. 2017). Damaging river floods have been more frequent from November to February, although they also occur on spring and autumn. On the other hand, flash floods are the result of short duration and intense precipitation events that affected small river basins with low concentration times, most frequently occurring during the autumn and winter, but may also occur in any month along the hydrological year (Zêzere et al. 2014).

Flood damage databases
Data of damaging floods occurred in the Northern region of Portugal, regardless of the number of affected people or the economic value of the resulting damages, was obtained from national and regional newspapers.
Human damages (fatalities, missing, injured, displaced and evacuated people) caused by floods in mainland Portugal are gathered in the DISASTER database (Zêzere et al. 2014) for the period 1865-2010, further updated until 2020. The DISASTER database includes floods and landslides that caused human damages, independently of the number of people affected. The methodology used to collect flood DISASTER cases and human damages is summarized in Zêzere et al. (2014).
Minor consequences caused by floods (road and railroad circulation disruption, flooded buildings and corresponding damages) were collected by Santos et al. (2015 for the period 1865-2016 for the Northern region of Portugal, also using newspapers as data sources. Additional information on the methodology used to collect flood cases that caused material damages is summarized in Santos et al. (2014Santos et al. ( , 2015. On both databases, a significant amount of work was developed to check and validate the flood cases, crosschecking different sources, from national to regional and local newspapers. In this work, we explore the damaging floods (river floods and flash floods) that caused both DISASTER-type losses and minor (uniquely material) losses in the Northern region of Portugal for the period 2000-2015, for which the exact location is known (Fig. 2). Along this period, the DISAS-TER database counts 40 flood cases that caused human damages (3 fatalities, 1 injured, 267 evacuated and 197 displaced people). Regarding minor damages, there are 213 flood cases (Table 1), of which 114 caused road disruptions (53.8%) and 94 caused damages in the built environment (44.3%).
Human damages caused by floods were recorded in 9 years (2000, 2001, 2002, 2004, 2006, 2010, 2011, 2013 and 2014), while minor damages were recorded in every year. The hydrological year of 2000/2001 recorded the highest number of damaging floods in the Northern region. River   floods, in opposition to flash floods, are the fluvial process that causes most of flood damages (whether with human consequences or with only minor damages) in the study area (89%). The Douro river basin concentrates 48% of the damaging floods, followed by the Ave river basin with 26% (Fig. 2). The majority of the minor damages are located near the coastal urban areas.

Susceptibility
Flood susceptibility for the study area was calculated at the parish level from the results obtained in a national-scale assessment . The model evaluates the propensity of each stream to flooding using four types of raw input data: a hydrologically corrected DEM, land use, dominant parent material and the fine fraction of the topsoil. From the first input data, flow accumulation (F acc ) and average slope (S avg ) of each cell (with 3 arc-second resolution) was derived. The remaining data was used to derive the third susceptibility conditioning factor, relative permeability (P rel ). Weights assigned to the three factors are validated using historical evidences of past major floods (Zêzere et al. 2014) in four validation areas: mainland Portugal, the Northern region, and three sub-basins where a historical record of minor flood cases was also available. The cell-defined susceptibility score is based on the sum of those weighted values, transformed through the min-max method to a range from 0 to 5. The advantage of this method is the evaluation of flood susceptibility in any given cell considering the contribution of the entire upstream drainage area. Additional information for this method can be found in Santos et al. (2019). The flood susceptibility value calculated per each parish is the simple arithmetic average of the susceptibility scores (Fig. 3).

Exposure
Two variables were used in the exposure assessment to floods at the parish level: population density and the percentage of artificial surfaces (Fig. 4). Population density was computed for each parish using the resident population in the 2011 population Census, divided by the parish area in km 2 . Artificial surfaces include the urban areas, road and railroad networks, construction areas, quarries and urban green spaces. This data was extracted from the Land Cover map of 2010 produced by the Portuguese Directorate-General of the Territory, with a minimum cartographic unit of 1 ha. The legend codes at level I were used, and the percentage of the artificial surface in the total area of each parish was computed. The final exposure value corresponds to the arithmetic average between population density and the percentage of artificial surfaces. Exposure in the Northern region of Portugal shows a very polarized spatial pattern, with a concentrated presence of population and urbanized areas in the metropolitan area of Porto, around the mouth of the Douro river, and from here northwards to the major cities of Braga and Viana do Castelo (Fig. 4).

Vulnerability
Vulnerability is represented in this research by social vulnerability (SV), which is defined as the propensity of individuals, communities and systems to be harmed by hazardous processes, based on their social and demographic characteristics and territorial context (Chen et al. 2013;Cutter et al. 2003;Mendes 2009;Mendes et al. 2019;Ogie and Pradhan 2019;Tavares et al. 2018;Yoon 2012). The statistical procedure to assess SV consists in the iterative application of principal component analysis until a robust and interpretable relationship of principal components is obtained from the explicative variables. In this research, an initial set of 24 variables collected at the parish level from the population Census of 2011 was used. After the iterative process, the final SV model ran with 15 variables, that were grouped around four principal components (PC) ( Table 2).
The first principal component (PC1) represents the contexts of demographic aging, low qualifications and low levels of women empowerment. PC2 represents the demographic and economic dynamism: positive loadings in variables related with qualifications, socially valued employments,  and internal and external attractiveness, whose PC scores needed to be inverted to be concordant with SV interpretation (cf. the negative cardinality in PC2 on Table 2). PC3 represents the economic condition of residents: positive loadings in such variables as overcrowded households and households without at least one basic infrastructure, opposing to the negative loading in the variable car usage in daily commuting. Finally, PC4 represents the mobility, as an SV driver expressing the need of long commuting times, concordant with high proportion of residents who work or study outside the municipality of residence. A score of SV is obtained (Fig. 5) by applying weights to the components' scores based on the percentage of explained variance. The parishes located on rural areas, with aged population and households without basic infrastructures, are characterized by high social vulnerability. Some parishes in the old city centers and outskirts of the Porto city are also evidenced as highly vulnerable.

Geographical and statistical analysis
The flood damages, susceptibility, exposure and vulnerability scores were statistically analyzed using their transformed values to the range [0, 1], using the min-max method.
The reported damaging floods were split in two databases, to which the geographical and statistical analysis was conducted: -Database of DISASTER-damaging flood cases, which only assembles the DISASTER-type cases; -Database of All-damaging flood cases, which assembles damaging flood cases with minor consequences and DISASTER-type cases.
The purpose of this procedure is to test if the risk components-susceptibility, exposure and vulnerability-play the same role independently of the type of damaging flood that is reported.
Classification And Regression Trees (CART) was developed by Breiman et al. (2017). It is a recurrent distribution non-parametric algorithm used to analyze and predict data relations . It is also applied to predict both categorical, i.e., classification, and continuous variables, i.e., regression (Felicísimo et al. 2013;Youssef et al. 2016). The dissimilarity among classification and regression approaches is one of the methods used to split and assemble data (Prasad et al. 2006).
In this study, CART was applied on two groups of cases (a group with All-damaging flood cases and another with DISASTER-damaging flood cases) to divide the flood cases into a sequence of classes upon their internal homogeneity, and to build a model for each group. The objective of this procedure is to create a tree and evaluate the set of logical if-then-else split conditions used. In this case, starting from the root, the method splits the data subsets, generating two child nodes. If the separated data belongs to the same class, then it is merged and forms a leaf. If not, the division process continues. The sample is divided and cataloged according to the squared residuals minimization algorithm (Timofeev 2004). This method creates data subsets the most likely  (Mahjoobi and Etemad-Shahidi 2008). More information concerning the CART algorithm can be found in .

Validation and sensitivity
To have an enhanced understanding of our analysis accuracy a k-fold cross validation procedure was applied. One randomly split the training data into 20 subsamples, i.e., folds, and trained each one of them with 80% of the data. The remaining 20% (that are mutually exclusive) were used once for validation and 19 times as a portion of the learning subsample (20-fold validation). This procedure returns a good estimate of the accuracy of the optimized tree, as it tests each of the 20 trees on a different subsample. The evaluated cross-validation is a measure of the average probability of misclassification among the 20 test subsamples.

Classification and regression trees
Using the All-damaging flood cases database and using the DISASTER-damaging flood cases database (i.e., that with only the most severe human consequences), results in a quite distinct classification of parishes according to the regression trees algorithm.
Both trees are simple structures with three nodes but that is the only similarity you get using the two databases. The first two leaves of the tree ran with All-damaging flood cases (Fig. 6) come from splitting exposure values using somewhat low thresholds i.e., low exposure (0.299 and 0.143 in a 0 to 1 scale), due to the high asymmetric distribution of urban areas in the study area (cf. Fig. 4). These two nodes classify 20% of the parishes giving us groups '2' and '4'. Finally, the last splitting is supported by vulnerability and separate the parishes with low values (= < 0.5, 25%) from those with high values (> 0.5, 55%). This gives us groups '5' and '6', respectively.
For the database that uses only the DISASTER-damaging flood cases (Fig. 7) the situation is diverse. The three splitting nodes use, sequentially, susceptibility, exposure and vulnerability. The first retrieves a leaf (4.9%) of high susceptibility values (> 0.6) and corresponds to group '2'. The second, splits between the 0.297 exposure threshold, defining group '4' (4.9%) as the one with values above it. Finally, the third one, separates two leaves, groups '5' and '6', the first with vulnerability values below or equal to 0.479, and the second with values above it.
In summary, using the database with All-damaging flood cases, the 4 groups are interpreted as: It is worth mentioning that in this regression tree, flood susceptibility does not play a key role in the classification of parishes using historical flood cases.
The groups resulting from using only the DISASTERdamaging flood cases can be interpreted as: -Group '2': parishes with high susceptibility: -Group '4': parishes with low susceptibility and high exposure: -Group '5': parishes with low susceptibility, low exposure and low vulnerability: -Group '6': parishes with low susceptibility, low exposure and high vulnerability.
Spatially, the groups based on the analysis with the Alldamaging flood cases (Fig. 8) tend to isolate important urban (highly populated) parishes and their expansion crowns, whereas in the DISASTER-damaging flood cases classification (Fig. 9) this spatial sorting is more fuzzy, and the proximity to two important rivers, i.e., Minho and Douro, is highlighted (Fig. 10). Table 3 shows that more confidence is put on the model created upon the DISASTER-damaging flood cases than in the model with the All-damaging flood cases. The former has a much lower-and acceptable-misclassification estimation and a residual standard error.

Validation and sensitivity
One used a measure of how much the model predicts value changes according to the dependent variable, in response to changes on the independent variables, to analyze the variables' importance (Table 4). The normalized value is just the importance of the variable regarding the highest importance value, displayed as percentage. Analyzing such results (Table 4), it can be drawn that in both databases exposure has a high importance when it comes to justify the grouping of parishes: normalized importance of 100% using the All-damaging and 98.5% using the DISASTER-damaging flood cases. The less important variable is vulnerability, which is however, four times more important when using the DISASTER-damaging flood cases in comparison with the All-damaging cases database: 25.1% comparing to 6.5% of the normalized importance, respectively (Table 4). The most opposite position in both databases, is that manifested by susceptibility, that has no importance in discriminating the All-damaging flood cases but arises as the most important when analyzing the DISASTER-damaging cases (Table 4).

Discussion
From the 2028 parishes, 57 parishes are included in the six Floods Directive areas of significant potential flood risk (ASPFR). Considering the three drivers of flood damagessusceptibility, exposure and vulnerability-the ASPFR parishes present low exposure, with the exception of the Porto ASPFR (Fig. 10). Porto (#5, comprising 13 parishes) and Régua (#6, comprising 10 parishes) ASPFR are the ones located along the major river, the Douro river, giving credit to the high susceptibility index values found in those parishes. Comparing the parishes covered by the Directive with the parishes outside the ASPFR (Fig. 10), it is verified that exposure and susceptibility are, on average, higher in the first group, while vulnerability is higher in the parishes outside the ASPFR. This confirms the findings presented in Table 4, which assign less importance to vulnerability as a splitting criterion in the CART classification method. Looking at the number of flood cases per parish, the Directive's parishes are capturing 23% of the DISASTER cases (9 in 40 cases) and 31% of the All-damaging flood cases database (79 in 253 cases). This means that around ¾ of the DISASTERdamaging flood cases have occurred in streams not covered by the Floods Directive. Although not available yet, it is expected that several of these locations will be covered in the 2nd cycle, as the number of ASPFR in the study area has been expanded from 6 to 17.
A more detailed analysis of the groups-resulting from the CART-and the Flood Directive's areas demonstrates that the Directive's parishes differ significantly within each ASPFR as well, and that using criteria-differentiated flood damage databases provides quite distinctive interpretations for a same parish (Fig. 11). In fact. using the database with All-damaging flood cases, 18 parishes are individualized from the set of 57 simply considering the exposure driver (group 2), with relevance in the Porto ASPFR. The DIS-ASTER-damaging flood cases complement that analysis by introducing the role of susceptibility: i.e., 8 of the 10 parishes of Régua ASPFR are classified in the All-damaging flood cases database as having low exposure and high vulnerability, while considering the DISASTER cases allows to characterize the hazard dimension, represented by high susceptibility. Chaves ASPFR is another interesting example of improved interpretation capacity using the two CART analysis: in the All-damaging flood cases CART, 2 parishes show high exposure and 8 show low exposure; in the DISASTER-damaging flood cases CART, the same 2 and 8 parishes maintain the exposure classifications, but it is also verified that they present low susceptibility. Although the DISASTER-damaging flood cases database produces a classification with less standard error (Table 3), both databases render an understanding about flood risk drivers, linked to small and frequent flood episodes, that is simultaneously relevant to flood risk decision-makers. Regarding the use of historical flood damage databases in risk characterization, this study concludes that using databases differentiated, or filtered, by the type of impact provides an equally differentiated classification of the role of risk drivers. This means that the role of susceptibility, exposure and vulnerability varies according to the damage characteristics of each record. When filtered by the degree of damage, the two distinct databases related differently with the flood risk drivers (Fig. 12).
Narrow inclusion criteria of flood damage databasesrequiring the existence of 10 or more fatalities, 100 or more affected, declaration of a state of emergency and a call for international assistance, as in the case of the EM-DAT-is optimal for a global comparison of contexts by reporting only the most important events. In local studies, like the one here presented for the Portuguese Northern region, the narrower set of data (the DISASTER-damaging flood cases), eventually and surprisingly, provided a more holistic representation of flood drivers when compared to the Alldamaging flood cases database. However, it cannot be disregarded that minor and small disasters are more frequent and are explained by societal and human-related drivers-not so dependent on the hazard characteristics-than the most impacting flood events. In resume, and considering three types of impact databases-global ones like the EM-DAT, national-level like the DISASTER and the All-cases type

Conclusions
Two types of flood damage databases were used as dependent variables to train a machine learning classification method, based on regression trees (CART), that uses flood susceptibility, exposure and vulnerability as independent variables. All the data is expressed at the parish level, a local administrative unit, summing 2028 territorial units of analysis in the Northern region of Portugal.
The results show that the risk drivers assume distinct roles in relation to the dependent variable, according to the inclusion criteria adopted in the definition of the two databases. Ultimately, the degree of damages (serious and minor), as documented in databases, is conditioned by the geographical distribution of the flood risk drivers. The inclusion of all types of damages highlights the role of exposure first and vulnerability later, but when the cases are filtered to include only human consequences-casualties, injuries, missing people, evacuated or displaced people-flood susceptibility emerges as the most relevant driver, followed by exposure and vulnerability.
Next steps of research in the field should explore in greater detail the combined effect of (i) the scale at which damage data is aggregated, (ii) and the distinct inclusion criteria on databases, over the explicative and predictive capacity of the three flood risk drivers.
Funding This research is funded by the project RISKCOAST-Development of tools to prevent and manage geological risks on the coast linked to climate change (SOE3/P4/EO868, Interreg Sudoe) and MIT-RSC-Multi-risk interactions toward resilient and sustainable cities (MIT-EXPL/CS/0018/2019, Portuguese Foundation for Science and Technology I.P. (FCT)). Sérgio Cruz Oliveira participation is funded by FCT through the project BeSafeSlide-Landslide Early Warning soft technology prototype to improve community resilience and adaptation to environmental change (PTDC/GES-AMB/30052/2017). Pedro Pinto Santos is financed by FCT under the contract CEECIND/00268/2017.

Conflict of interest
The authors declare that they have no known competing interests or personal relationships that could have appeared to influence the work reported in this paper.