Landslides in the Styrian Basin, Austria, have received much attention, especially recently after the 2009 and 2014 compound events . Our study area (3831 km2) is characterised by flat lowlands in the east and a hilly topography in the west, ranging from 196 to 1167 m above the Adriatic Sea (m AA) with a relative relief of 3–550 m km-2. The Styrian Prealps border the study area from the southwest (Possruck Mountains, Koralpe) to northwest (Stubalpe Mountains, Graz Mountains). According to and , the Styrian Basin has a high predisposition to landslides due to the underlying geology of Neogene sediments (thick Miocene and minor Pliocene sediments) mainly consisting of a heterogeneous mixture and interbedded strata of sands, silts, clays, marl, and gravels (more geological details are given in ).

In June 2009 and September 2014, heavy-rainfall events occurred in southeastern Styria. From 22 to 25 June 2009, a series of heavy thunder- and rainstorms brought over 100 mm of precipitation within 24 h in some places, which corresponds to a 50-year return period (Fig. c left; meteorological details in ). In the region, around 1700 landslide-related private damage claims were submitted to the state government, surpassing EUR 13.4 million in cost for reconstruction and emergency response . The northern part of the district of South East Styria was particularly severely affected (Fig. b), so for several municipalities a state of emergency had to be declared . The rainfall event in 2014 was similar, although less severe. In September 2014, several heavy thunderstorms occurred within a 3-week period . From 12 to 15 September, between 30 and 100 mm of rain (Fig. c right) was brought to the region by an upper-level low . Again, landslides occurred across the entire Styrian Basin but were clustered with more than 500 landslides south of Leibnitz and north of Gleisdorf (Fig. b).

The database of the analysis is consistent with the data described in and . While detailed information on the landslide inventory, the sampling design, and the landslide susceptibility model can be found in , model extensions and applications that account for soil moisture and environmental changes, respectively, were presented in . Similarly to the above-mentioned studies, we used a target resolution of 10 m × 10 m for our analysis. Here we will present the database, only briefly giving the details relevant for our model construction, and we refer the reader to the cited publications for further detail.

Our analysis built on the landslide susceptibility model of , which was a generalised additive model (GAM). A GAM is a semi-parametric extension of a generalised linear model (GLM), since it has the ability to model non-linear relationships by automatically fitting transformations, or so-called component smooth functions . The additive structure of a GAM allows common model diagnostics (e.g. predictor–response relationship, variable importance, odds ratios), and thus GAMs have become popular in landslide susceptibility studies in recent years .

For the landslide susceptibility analysis, we linked predisposing and time-varying preparatory and triggering factors to landslide occurrences by following recent recommendations for non-stationary landslide susceptibility models . Therefore, we are analysing conditional landslide susceptibility in the sense of a susceptibility that applies under given (time-varying) environmental conditions. Predictor variables were land surface variables (convergence index of 100 and 500 m, plan and profile curvature, logarithmic D-Infinity flow accumulation, normalised height, slope angle, slope angle of catchment area, north and west exposedness, topographic position index (TPI), and topographic wetness index (SWI)), hydrometeorological variables (soil moisture, 5 d rainfall, and maximum 3 h rainfall intensity), geology, and forest type (representing LULC). For more information on the delineation of the predictor variables and their maps, refer to and Fig. S6, respectively.

We developed landslide susceptibility models with different model settings, some of which have been published before (Table  in Appendix ). As a new model, we decided to explicitly account for residual spatial autocorrelation using a low-rank Gaussian process (GP) smoother as an additional predictor (GAM-Spatial). This feature is often missing in landslide susceptibility modelling, which may result in residual patterns that are unaccounted for by the model and may introduce errors into predictions . Furthermore, we kept the influence of the 5 d rainfall variable constant beyond 80 mm (“top-coded”) to counteract a physically implausible predictor–response relationship beyond this threshold (GAM-SM+TC in ). In other model settings, the 5 d rainfall variable was not modified (i.e. GAM-SM; ; cf. Fig. , Appendix ), and soil moisture was not included as model predictor (i.e. GAM-Co; ).

We used a GAM-Spatial implementation that incorporates residual spatial autocorrelation through the GP representation . The hyperparameters necessary for the GP implementation are a range (Φ) and correlation function and were estimated based on the residuals of GAM-SM+TC. We followed in iteratively searching for a suitable range (from 10 to 1000 m) and correlation function (spherical, exponential, Gaussian, and Matérn, the latter with κ values of 1.5, 2.5, and 3.5). The models' restricted maximum likelihood (REML) score was assessed to identify optimal hyperparameters (i.e. lowest REML score), which were then applied to fit the GAM-Spatial model with a GP smoother. Additionally, the effective range, at which the residual autocorrelation drops below 0.05, was calculated using StempCens in R (EffectiveRange with cor =0.05; ).

For model assessment, we used well-established diagnostic tools. The model performance was assessed using a 5-fold spatial cross-validation (SpCV) with five repetitions and measured using the area under the receiver operating characteristic curve (AUROC) . We ensured that the training and validation data were identical to those in to achieve a fair comparison of the landslide susceptibility models. The AUROC values were interpreted following the interpretation guide of . To assess variable importance, we calculated the model's mean decrease in deviance explained (mDD, %) after removing the respective variable from the model . Larger mDD values indicate greater explanatory power. Predictor–response relationships were analysed visually using transformation functions and quantitatively using odds ratios (ORs). An OR represents the chance that an outcome happened given a specific exposure, compared to odds of the outcome under a reference exposure . An OR greater than 1 means an exposure with higher odds of landslide occurrence, while an OR lower than 1 is associated with lower odds of landslide occurrence while accounting for the other variables in the model (OR =1 means no association with the exposure). In contrast to , the variable importance and predictor–response relationships were assessed based on GAMs using the complete dataset (GAM-Spatial, GAM-SM+TC, GAM-SM), i.e. not on subsets of the SpCV models (GAM-Co).

We focused on highly susceptible areas for analysing the uncertainty cascade in storylines of the landslide susceptibility. The area of high landslide susceptibility was defined using the thresholding approach of , according to which 70 % of the observed landslides fall into that class; low and medium susceptibility account for 5 % and 25 % of observed landslides.

We used the free and open-source computing environment R (R version 4.1.0; ) with the GAM implementation in the mgcv package .

Landslide susceptibility is commonly considered stationary (or invariant in time; ). However, this assumption is often violated in space (e.g. anthropogenic land use changes; ) and time (e.g. “follow-up” landslides; ). In this study, we assessed NO-CC and future landslide susceptibility by considering the LULC and hydrometeorological variables as time-varying predictors, and therefore we modified their values according to the storylines. While the hydrometeorological predictors were multiplied with the corresponding delta change factor, the LULC data were replaced with the developed scenario only for the future. The other predictors (land surface variables and geology) were considered invariant in time.

To estimate changes in landslide susceptibility between a storyline and the present day, we applied two approaches: first, we calculated the relative change in the areal extent of a susceptibility class (in % and percentage points, pp) to quantitatively derive the potential change in the area at risk. And second, we expressed the effect size of change using the average change in the odds of landslide occurrence of a susceptibility class as OR class for those locations which fall into identical susceptibility classes in both predictions: 1ORclass=exp⁡1n∑i=1nlogitstory(i)-logitref(i), where logitstory(i) and logitref(i) denote the logits of landslide occurrence probabilities at a specific location i and susceptibility class (low, medium, or high) of the storyline and reference scenario (present day), respectively, and n is all locations of the same landslide susceptibility class in both scenarios.

In integrated modelling, there are various sources of uncertainty affecting the predicted outcomes. We quantified the uncertainty cascade in storylines of landslide susceptibility by accounting for climate model uncertainty and landslide model parametric uncertainty (or sampling variability).

For the climate model uncertainty we assessed within-event internal model variability and scenario uncertainty for each hydrometeorological variable. The applied single event storyline approach includes a particular realisation of internal climate variability; thus the established storylines are conditioned on it . By using estimates from a climate model ensemble with different realisations of internal variability, the uncertainty related to conditional internal climate model variability can essentially be removed . In this study, we explicitly analyse within-event internal climate model variability and therefore address the question of how the event could locally unfold in the present climate as well as cooler and warmer climates. Regarding the climate signal or forced change, the influence of within-event internal climate model variability can be accounted for by averaging across the ensemble members of a climate models' storyline .

For the within-event internal climate model variability, we estimated the 2.5th and 97.5th percentiles from the delta change factor distribution (i.e. 100 pairs per storyline) describing the lower and upper bound of local climate model variability; the mean delta change factor represents a models’ climate signal (or forced change). The estimated delta change factors were subsequently offset against the corresponding values of the hydrometeorological predictor. For the climate scenario uncertainty, the width of the predicted outcomes using the climate signals of all climate models within a scenario (NO-CC, PARIS, 3 and 4 K warming) was calculated.

The landslide models’ parametric uncertainty is associated with measurement errors, sampling errors and variability, misclassification of data, and surrogate data weaknesses and, from a frequentist perspective, is commonly expressed using confidence intervals. A common approach to quantify parametric uncertainty is bootstrapping e.g.. However, in the framework of the mgcv::gam, this approach has limitations when penalties are present and some observations are sampled twice. This may cause undersmoothing and is therefore not recommended . However, the manner in which the mgcv GAM is implemented can be viewed as “empirical Bayes” . This allows the calculation of pointwise Bayesian credible intervals for the estimated model parameters using the standard error derived from the Bayesian posterior covariance matrix . The pointwise Bayesian credible intervals have good frequentist coverage probabilities when averaged across the function domain but with over- and under-coverage in some parts . In contrast, a simultaneous Bayesian credible interval that contains the entire true function for a given credible level (1-α) can be achieved by posterior simulation from the posterior distribution of the model coefficients ; this is the approach we adopt in this study.

To simulate GAM coefficients, a Gaussian approximation to the posterior for the coefficients is a computationally very efficient approach but not recommended for spatial models with a logit link in the mgcv framework . Therefore, we used a simple Metropolis–Hastings (MH) sampler with random-walk proposals as implemented in the mgcv::gam.mh function, but we also calculated a Gaussian approximation for comparison purposes (mgcv::rmvn function). The MH function reports two types of acceptance rate: a fixed acceptance rate and a random-walk acceptance rate (refer to , for more details). The simulations of the MH sampler are controlled by hyperparameters that need to be tuned to achieve an optimal proposal distribution. A proposal distribution is optimal for a random-walk acceptance rate of 23.4 %, while high values are to be achieved for the fixed acceptance rate . Thus, the optimum of the hyperparameter tuning was selected by minimising the sum of the absolute difference of both acceptance rates to their optimal rates. For the tuning, we set the initial burn-in period to 5000 simulations, the number of (actual) “simulated GAMs” to 10 000, and optimised rw.scale (random-walk scale factor for a posterior covariance matrix) and t.df (degrees of freedom of the t distribution) using a grid search. We defined a regular grid ranging from 0.005 to 0.02 for rw.scale (step size of 0.0005, i.e. 31 total steps) and from 25 to 10 000 (step size of 25, i.e. 400 total steps) for t.df (in total 12 400 executions).

Furthermore, the simulation from the posterior distribution allows us to derive the so-called critical value for each non-parametric transformation function . The critical value is a factor which modifies the pointwise Bayesian credible interval to achieve the simultaneous Bayesian credible interval for a given credible level. A critical value greater than the coverage factor for a given credible level means an underestimation of the true function for the pointwise interval compared to the simultaneous interval (e.g. critical value of 2.5 vs. coverage factor of 1.96 for 95 % credible level, 28 % underestimation; ), and large critical values may indicate particularly uncertain predictors.

For the uncertainty cascade, landslide susceptibility prediction uncertainty intervals were estimated as described in the following. First, we classified the predictions of each of the simulated GAM-Spatial models into (low, medium, and) high landslide susceptibilities. And second, we derived the 2.5th and 97.5th percentiles as lower and upper uncertainty intervals of the corresponding storyline from the resulting distribution of the high-landslide-susceptibility area (i.e. 3×10000 values of high landslide susceptibility for mean and percentiles of delta change factor distribution). Finally, to compare the different sources of uncertainty in the uncertainty cascade, we calculated the following ratios: 2RIV;Lsl=width of CIIVwidth of CILsl, 3RCS;Lsl=width of CICSwidth of CILsl, where RIV;Lsl and RCS;Lsl denote the ratio of the uncertainty width of within-event internal climate model variability (CIIV) and scenario uncertainty (i.e. width of climate signals, CICS), respectively, to the parametric uncertainty width of the landslide model (CILsl). The joint uncertainty distribution of the landslide models' parametric uncertainty and within-event internal climate model variability is here referred to as the storyline uncertainty (CIStory).

We identified different patterns of delta change factor distributions depending on the considered hydrometeorological variable and the storyline (Fig. ).

The mean delta change factors (i.e. climate signals) for 5 d rainfall and maximum 3 h rainfall intensity were generally lower in NO-CC and higher in the future relative to the present-day climate (Table  in Appendix ). In the future climate, with the exception of the GFDL model, the mean increase ranged from 3 % (IPSL, PARIS) to 34 % (MIROC, 4 K warming) for 5 d rainfall and from 4 % (HadGEM, PARIS) to 61 % (MIROC, 4 K warming) for maximum 3 h rainfall intensity. In NO-CC, with the exception of the GFDL model, the changes were negative on average (i.e. factors <1). Specifically, mean decreases were between 3 % (IPSL) and 6 % (HadGEM) for 5 d rainfall and ranged from 10 % (HadGEM) to 13 % (MIROC) for maximum 3 h rainfall intensity. The averaged delta change factor of GFDL indicated projected changes of ±1 % in both meteorological variables in NO-CC as well as future climates. For soil moisture, the models projected on average drier soils in the future and wetter soil in NO-CC relative to the present-day climate, with the exception of HadGEM. IPSL in 3 K warming projected the driest soils (-16 %, on average), and it also projected the wettest soil conditions overall (+2 %) in NO-CC, on average. Following , the hydrometeorological storylines for the future can thus be summarised as (much) heavier rain (HadGEM, MIROC, IPSL) and (much) drier soil (IPSL, GFDL, MIROC), although there were also some neutral projections (rain: GFDL; soil: HadGEM; Table  in Appendix ).

Regarding the 2.5th and 97.5th percentiles of the storylines' delta change factor distributions (i.e. within-event internal climate model variability), we discovered opposed delta change factors in the GFDL storylines (Table  in Appendix ). Similarly, in the NO-CC and PARIS scenarios for 5 d rainfall, all other climate models had contrasting delta change factors, while for the maximum 3 h rainfall intensity this was only the case for HadGEM and IPSL in PARIS. Regarding the percentiles of the soil moisture storylines, only coherent delta change factors were found, yet some climate signals were weak (e.g. nearly 1 for HadGEM).

We established a spatial GAM using an optimised GP smoother. The spatial structure of GAM-SM+TC residuals was best described by a Gaussian correlation function with an effective range of 524 m (Φ of 303 m). Nevertheless, the other correlation models achieved nearly identical REML scores with varying effective ranges (582 to 710 m, Table in Appendix  and Fig. S2). The following subsections describe the model assessment (Sect. ) and the posterior simulation of the coefficients (Sect. ) for GAM-Spatial. Detailed results for the other model settings (GAM-Co and GAM-SM+TC) are included in the Supplement (Fig. S3).

The landslide susceptibility predictions of the simulated GAM-Spatial models were classified into low, medium, and high susceptibility . The thresholds to discriminate low from medium and medium from high landslide susceptibility were 0.16 and 0.57 on the probability scale, respectively. In the following, we focus on the storylines of highly susceptible areas and their uncertainties (Fig. ).

Regarding the present-day landslide susceptibility, around 4.9 % of the study area was highly susceptible to landslides (Fig. a, Table  in Appendix ). In the NO-CC scenario, high susceptibility was generally less common (4.4 %–4.8 %), with the exception of GFDL, which showed a slightly higher share (5.1 %).

The projected future signal, in contrast, was less coherent. In PARIS, the mean signal was similar to present-day susceptibility (-0.2 pp (percentage points difference in area) for GFDL and +0.2 pp for HadGEM), while the mean changes in 3 and 4 K warming were substantial. In 3 and 4 K warming, GFDL and IPSL projected a decrease in the highly susceptible area by up to -1.3 and -2.2 pp (-27 % and -45 %), respectively, while HadGEM and MIROC showed the opposite signal with increases of up to +1.7 and +0.8 pp (+35 % and +16 %), respectively. The mean signals of climate models in a 3 K warmer world were lower than in 4 K. Additionally, the considered LULC scenario resulted in a generally smaller extent of highly susceptible areas (up to 0.5 pp).

Regarding the OR of the storylines relative to present-day landslide susceptibility within the high-susceptibility class (see Eq. ), the pattern is similar to the change in susceptible area but more coherent (Fig. b, Table  in Appendix ). The chances of landslide occurrence in the NO-CC scenario were 0.91 (HadGEM) to 0.98 times (IPSL) as high as in the present-day settings (with the exception of GFDL with an OR of 1.05), based on the modelled climate signals. For the future, while in PARIS the ORs showed small effect sizes (ORs between 0.97 for GFDL and 1.04 for HadGEM), in 3 and 4 K warming the effect sizes were medium to large. In the worst-case storyline, the chances of landslide occurrence in highly susceptible areas was 1.37 times as high as in the present day (i.e. for HadGEM in 4 K warming), while it was 0.60 times as high in the best-case storyline (i.e. for IPSL in 3 K warming).

Regarding the storyline uncertainty (CIStory, Fig. a, Table  in Appendix ), for present-day landslide susceptibility, the uncertainty width was from +1.4 pp (+29 % area change) to -1.1 pp (-22 %). While present-day predictions were not affected by within-event internal climate model variability and thus reflected only landslide model uncertainty, the other storyline uncertainties accounted for both. The largest uncertainty interval for NO-CC covered 3.0 pp (GFDL). For the future, we identified larger uncertainty intervals with increased warming levels (largest intervals: PARIS, 3.2 pp for HadGEM; 3 K warming, 4.4 pp for MIROC; 4 K warming, 5.4 pp for MIROC).

Regarding the contributions of the uncertainty cascade in the storylines, the uncertainty introduced by within-event internal climate model variability is around 0.13 (IPSL in 3 K warming) to 0.35 (HadGEM LULCC in 4 K warming) times as large as parametric landslide model uncertainty (RIV;Lsl, Fig. c, Table  in Appendix ; see Eq. ). Instead, the ratio of climate scenario uncertainty to parametric landslide model uncertainty (RCS;Lsl; see Eq. ) showed a distinct pattern depending on the scenario: for NO-CC and PARIS, the uncertainty introduced by the climate models was lower relative to parametric landslide model uncertainties (ratios between 0.28 and 0.34). Instead, for 3 K warming, the climate scenario uncertainty was generally larger than the parametric landslide model uncertainty with ratios between 1.05 (IPSL) and 1.46 (GFDL LULCC) (excluding MIROC with a ratio of 0.97). For 4 K warming, while the climate scenario uncertainty was larger than the landslide model uncertainties for GFDL and GFDL LULCC (ratios of 1.23 and 1.38, respectively), for the other models the ratios were equal to (i.e. HadGEM LULCC) or lower than 1 (e.g. ratio of 0.77 for MIROC).

We identified different trends of projected landslide susceptibility for future storylines relative to the present day, while for the NO-CC storylines (i.e. pre-industrial, counterfactual climate), landslide susceptibility was generally estimated to be lower (exception: GFDL). For projected future changes in hydrometeorological conditions (i.e. climate signals), the storyline of much heavier rain showed the strongest increase in affected highly susceptible area (+35 %) with additionally higher chances of landslide occurrence relative to present-day susceptibility (OR of 1.37, HadGEM in 4 K). However, the effect of much drier soil may compensate for heavier rain and thus reduce the affected area (-45 %) and the chance of landslide occurrence (OR of 0.6, IPSL in 3 K). The projected changes in highly susceptible areas and their magnitudes were in general agreement with the findings of , who used a small part of our study area (Feldbach region, 95th percentile for high-susceptibility thresholding). For the worst- and best-case storylines, found changes of +45 % (4 K warming) and -37 % (3 K warming) in highly susceptible area and ORs of 1.66 (4 K warming) and 0.8 (3 K warming) in landslide susceptibility in those areas.

For other regions, some authors have projected an increased landslide occurrence or affected area that is attributable to climate change , while other authors have found more complex patterns that depend on the seasonal period and location under consideration . Projected increases in slope failure probability amounted to approximately 25 % for mountainous areas in Norway (for 2050; ) or 40 % for the Umbria region in central Italy (for 1990–2013 to 2070–2099 under RCP8.5; ), and an increase in susceptible area by around 32 % was projected for southwestern Taiwan (for 2090; ). With a focus on seasonal differences, projected increases in landslide occurrence mainly in winter, while in the warm and wet season, very low soil moisture and increased evaporation (; for 2071–2100 under RCP8.5) might even improve slope stability. Analysing spatial patterns in Calabria, southern Italy, identified for around 27 % of the municipalities a significant increase in rainfall-events with landslides (western part of the region and along the main mountain chains), while for 38 % of the municipalities a reduction was estimated (for 1981–2010 to 2036–2065 under RCP8.5). Therefore, we agree with that projecting the interplay between a natural hazard and climatic changes is still a challenging task, especially if multiple triggers and locally driven ground responses are present.

Regarding changes in LULC, our findings suggest that active LULC management and afforestation may have a beneficial effect on landslide occurrences (-0.5 pp in highly susceptible areas). Such effects have also been reported by other authors .

The analysis of uncertainties in landslide risk management under environmental change is an important, yet often neglected, task . In particular, the analysis of uncertainty cascades in integrated modelling is still an open research gap with great potential to understand and subsequently reduce uncertainty sources .

In our analysis, we could successfully quantify uncertainty in landslide susceptibility predictions by accounting for parametric uncertainty in the landslide model and climate change uncertainty (within-event internal model variability and scenario uncertainty). However, we found the estimated uncertainties in predictions to be generally high, which is not unusual for an uncertainty cascade . Furthermore, we discovered a tendency towards higher uncertainty with increased warming level for both parametric uncertainty and within-event internal climate model variability. Uncertainty from within-event internal climate model variability was much lower than parametric uncertainty (RIV;Lsl of around 0.25). Additionally, parametric uncertainty even exceeded scenario uncertainty for specific climate models (e.g. RCS;Lsl of 0.77 for MIROC in 4 K warming). Other authors have reported similar uncertainty components in the context of hydrological modelling under climate change .

In the following subsection the uncertainties in climate change modelling (Sect. ) and landslide susceptibility modelling (Sect. ) are further discussed.

In this study, we successfully modified a GAM for landslide susceptibility analysis to explicitly account for spatial autocorrelation effects using a GP component (GAM-Spatial) – an often missed feature in landslide susceptibility modelling . The spatial GP component explained a meaningful fraction of the deviance (6 % mDD), indicating that a spatial pattern remained unaccounted for by the other predictors in the model. Furthermore, the modified GAM showed similar predictor–response relationships and variable importance as well as an outstanding discrimination skill (mAUROC of 0.94) comparable to GAMs used in previous studies , confirming the general reliability of the landslide susceptibility model. Additionally, the time-varying modelling perspective on (conditional) landslide susceptibility as recommended by various authors allowed us to analyse the effects of LULC and climate change dynamics.

For the assessment of the uncertainties related to GAM-Spatial, we identified and applied optimal hyperparameters for a simple Metropolis–Hastings sampler (mgcv::gam.mh function in R) to stochastically simulate GAMs by approximating the posterior distribution of the coefficients, as this approach is recommended for logit link functions . However, the identified very large optimal hyperparameter value for t.df (i.e. 6050 degrees of freedom of the t distribution) indicated a nearly Gaussian approximation for the posterior distribution. Furthermore, regarding the similarity in the critical values for simultaneous intervals between the Metropolis–Hastings sampler and a simple Gaussian approximation (mgcv::rmvn function in R, Table in Appendix ), the effort put into tuning and applying the Metropolis–Hastings sampler is worth reconsidering in future studies with similar settings.

The presented methodological approaches are generally applicable to other regions. However, the established landslide susceptibility model was based on only two landslide-event inventories at the regional scale for a single study area. Thus, the projected uncertainties and effects of climate and LULC change on rainfall-induced landslide occurrences may be specific to the local geological and geomorphological setting. With a greater awareness of the need for a reliable landslide-event inventory and with additional landslides surveyed in the future, a more generalisable and less uncertain landslide model may be established.

As a practical application, the prediction of landslide susceptibility, its projections of change in area and magnitude with climate change, and its uncertainties may provide important tools for planners and decision-makers. With the chosen approach it is possible not only to consider susceptibility maps statically and retrospectively but also to project them into the future, which is an essential prerequisite for climate change adaptation in the context of slope stability. In the spatial context, recommendations for the construction of new settlement infrastructure may be derived based on a map, and zones of high landslide susceptibility along with uncertainty graphs (Fig. S7) can be communicated to local planners and environmental managers cf.. In the engineering context, grey (e.g. engineering structure) or green (e.g. forest management) approaches may be considered a scenario, and their effectiveness for potential slope stabilisation on a regional scale may be assessed cf.. Such development scenarios also enable the analysis of the exposure of elements at risk, potential mitigation measurements, and sustainable adaptations plans.