Based on a literature review, an initial set of landslides and flash flood causative factors was identified (Alarifi et al., 2022; Bui et al., 2019; Chowdhury, 2024; Elghouat et al., 2024; He et al., 2025; Khodaei et al., 2025; Luu et al., 2023; Pham et al., 2021; Rayamajhi et al., 2025; Shahabi et al., 2021; Wahba et al., 2024). These factors are listed in Table 1 and briefly described afterwards. Figures 4 and 5 provide a visual overview of the spatial distribution of these input layers, as used in the machine learning algorithm.

Unlike other studies that include rainfall as a dynamic input variable in landslide susceptibility mapping (Ahmed et al., 2023; Lee et al., 2022), we chose not to incorporate rainfall data in our model. Although rainfall is a well-known trigger for both landslides and flash floods, its use as a predictive requires rainfall datasets with sufficient spatial and temporal resolution. This is particularly important in mountainous areas such as Liguria, where rainfall patterns are highly variable, and short-duration, high-intensity convective storms frequently occur. These are often not captured by coarse-resolution (coarser than 4 km) datasets, such as those provided by satellite-based or global reanalysis products e.g. ERA5 (Hersbach et al., 2020), CHIRPS (Funk et al., 2015), PERSIANN (Sorooshian et al., 2014), IMERG (Huffman et al., 2020). Moreover, the other input data used in this study – such as topography and land cover – are available at much finer resolutions (∼ 30–250 m). Such a mismatch between input resolutions can lead to inaccuracies in the model and reduce confidence in the resulting susceptibility maps. Future studies may incorporate dynamic rainfall information, provided that high-resolution precipitation datasets become available and are spatially consistent with the other model inputs.

This study was conducted following a stepwise framework. First, a multicollinearity assessment and a frequency ratio analysis were carried out to evaluate the relevance and consistency of the conditioning factors for both flash floods and landslides. Second, three machine learning models were calibrated and evaluated using identical conditioning factors and different inventories to produce flash flood and landslide susceptibility maps.

Figure 7 shows the calculated Variance Inflation Factor (VIF) for all selected causative factors. Only the Stream Power Index (SPI) exhibited moderate multicollinearity (VIF =5.7), while all other factors showed low level of multicollinearity. Therefore, all factors were included in the ML-models.

Figures 8 and 9 present the frequency ratio (FR) values for landslides and flash floods across numerical variables. Overall, the FR patterns were similar for both hazards.

Figure 8 shows that lower elevations and gentler slopes corresponded to higher likelihood of both flash floods and landslides occurrences. Although steep slopes tend to rapidly collect surface runoff, the water is accumulated in valleys. The differences in FR for landslides between slope classes is relatively small (0.5< FR <1.5) meaning that slope exerted a limited influence on landslide susceptibility in the study area. Regarding aspect, both flash floods and landslides are most likely to occur in areas facing South, and highly unlikely on flat surfaces. Convex surfaces such as hilltops and ridges are very unlikely (FR <0.5) for flash floods to occur and unlikely (FR <1) for landslides to occur on. On the contrary, flat areas were associated with the highest FR for flash floods. For areas close to roads (0–21.6 m) the FR is high (>1.5) for both flash floods and landslides and drastically drops for the larger distances.

In Fig. 9, we observe that for the Topographic Wetness Index (TWI), the highest FR values – areas that are most susceptible to saturation – correspond to the highest likelihood for both hazards. The FR decreases with decreasing TWI, a trend pronounced more for flash floods than for landslides. On the contrary, areas with low Stream Power Index (SPI) – lower erosive power of streams – exhibit the highest FR (∼ 1.6) for both hazards, with FR strongly decreasing as SPI increases. Proximity to rivers showed a clear spatial relationship: the closer to rivers, the higher the likelihood of both flood and landslides. This likelihood decreases sharply with distance. We obtained high FR values for areas with low NDVI (0.2–0.5), and with increasing NDVI values, we observed decreasing FR values.

The influence of surface lithology and land cover on hazard occurrence is shown in Fig. 10. For lithology, the highest FR values for both hazards corresponded to marlstones and limestones which have low permeability. However, the FR for landslides and flash floods remained close to 1 (0.8–1.2 and 0.7–1.3 respectively) across all lithological classes, indicating that lithology exerted only a limited influence on landslide and flash flood occurrence in the study area. Regarding land cover, the highest FR values are found for artificial soils. Flash floods are unlikely to occur in the other land cover classes (FR <1) whereas landslide likelihood occurrence is more evenly distributed across land cover classes except for closed forest in which they are unlikely to occur (FR <1).

Figures 11 and 12 show the susceptibility maps for flash floods and landslides, respectively, for each of the ML models: Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM). The susceptibility maps obtained are continuous values ranging from 0 to 1 and have been reclassified into five discrete classes to improve interpretability of the outcomes using Jenks natural break (see Table A1 in the Appendix for the exact classes). The best parameters as given by the highest AUC together with the parameters tested are shown in Table B1 of Appendix B.

Overall, the maps follow a similar pattern for all algorithms and both hazard types. Along the coastline, areas of high susceptibility are widespread. Numerous villages and cities are located along this coast, including the historic city of Genoa, which has been repeatedly affected by severe flash floods in its urban valleys and by rainfall-induced landslides on the surrounding hills. Susceptibility is also high for both hazards along the Polcevera river valley, north of Genoa, (refer to Northern Liguria zoomed-in area) where extensive urbanization has led to the confinement or disappearance of natural waterways, ephemeral streams, and artificial channels, thereby increasing the likelihood and severity of catastrophic events (Faccini et al., 2015; Lanza, 2003).

Although there are many areas susceptible to both hazards, pronounced differences between flash floods and landslides occur in the Western part of the Liguria region (see Western Liguria zoomed-in area). Here the area with high landslide susceptibility is larger than the area with high flash flood susceptibility. This sector is characterized by relatively steep slopes (Fig. 4) and extensive crop land (Fig. 5), including vineyards. In case of non-permanent crops and frequent tillage, soil stability tends to be reduced (Giarola et al., 2024) leading to high occurrence of landslide events. In addition, the western Liguria region is driest part of the study area, with less than 600 mm yr⁻¹, therefore leading to fewer flash floods recorded in the inventory.

Across all methods and ML algorithms, the total area identified as highly susceptible is larger for landslides than for flash floods. This difference largely reflects the characteristics of the two inventories (see Fig. 1). The landslide events are more widespread, particularly in the northern and western sector of Liguria, whereas flash flood events are concentrated in the lower valleys. The landslide inventory used in the study consisted of 2412 events, whereas we only have 526 historic flash flood events.

As discussed in Sect. 3.2.2 Hazard Susceptibility Modelling, the ML models were trained using area under the curve (AUC), tested and evaluated using the receiver operating characteristic (ROC) curve, the AUC, the confusion matrix, the accuracy score, and the F1 score. The ROC and the AUC for flash floods and landslides for each of the three ML models are shown in Fig. 13. For both hazards, we observe that the Random Forest (RF) achieved the highest results with an AUC of 0.95 and 0.93 for flash floods and landslides respectively whereas LR showed the poorest results. The confusion matrix which summarizes how many pixels (as defined in the reference dataset) were correctly predicted is shown in Table B2 of the Appendix B for the training dataset and in Table 2 for the testing dataset. The RF, exhibited the best results (flash floods, landslides) with the highest number of TP (593, 3318) and TN (591, 2967), and the lowest number of FP (96, 690) and FN (80, 381), followed by the SVM. The highest accuracy is achieved by the RF with values 0.87 and 0.86 for flash floods and landslides respectively. Similarly, the highest F1 score is attained by the RF with the exact values.

As indicated in Fig. 8 lower elevations and gentler slopes, e.g. valleys, had the highest likelihood of flash floods and landslides occurrences which agrees with previous studies on flash floods (Chowdhury, 2024; Elghouat et al., 2024; Pham et al., 2021; Rayamajhi et al., 2025). For landslides, however, the apparent preference, though limited, for gentle slopes is unexpected. This pattern may be partly influenced by the characteristics of the ITALICA landslide database. As noted by the database authors, landslides that took place in remote or uninhabited regions, or for which no institutional or media reports exist, are rarely included. Consequently, the inventory likely overrepresents events on gentler slopes and in lower-elevation areas near infrastructure, while underrepresenting those in steep, remote terrain. As to aspect, the orographic precipitation resulting from the Ligurian Sea partly explains the heavier rainfalls on South facing slopes contributing to an increase in both hazards. Figure 8 showed that flat areas were associated with the highest FR for flash floods, which was also observed in Chowdhury (2024). Similarly to our findings, Mancini et al. (2010) obtained the highest FR for landslides near the roads in Daunia, Italy located in the Apennines while highlighting the inventory bias.

Previous studies on flash floods (Chowdhury, 2024; Elghouat et al., 2024; Rayamajhi et al., 2025) have indicated as well that flood-prone areas exhibited high TWI values (Fig. 9). In line with our findings, Chowdhury (2024) obtained for flash floods the highest FR values for SPI values close to 0. The patterns we observed regarding proximity to rivers have been reported for other flash floods or floods studies too (Elghouat et al., 2024; Pham et al., 2021; Rayamajhi et al., 2025). We expected to obtain high FR values for areas with low NDVI values which correspond to bare soil or areas with limited vegetation and usually experience more frequent landslide and flash floods occurrences.

As pointed out in Fig. 10, the highest FR values are found for artificial soils. For flash floods this can be explained by the limited infiltration capacity whereas for landslides, we expect the human-made changes to topography and loss of vegetation in urban areas to play a bigger role. Indeed, urbanization increases landslide hazard (Johnston et al., 2021). Landslides were unlikely to occur in closed forest confirming that strong root cohesion and vegetation cover enhance slope stability.

The unified framework applied in this study ensured methodological consistency between the analyses of flash floods and landslides, allowing the same set of conditioning factors and modelling procedures to be used for both hazards. All tested models showed good predictive capability, with Random Forest achieving the best overall performance (AUC =0.95, accuracy =87%, F1 score =87% for flash floods, and AUC =0.93, accuracy =86%, and F1 score =86% for landslides). Furthermore, as summarized in Table 2, the confusion matrix which allows a disaggregated view of the model performance, showed that all models – particularly the Random Forest – had promising results, i.e. high TP and TN while keeping the FP and the FN low. These findings are consistent with other studies that reported Random Forest as the most effective model for predicting the occurrence of flash floods and floods (e.g., Elghouat et al., 2024; Khodaei et al., 2025) and landslides (e.g., Youssef and Pourghasemi, 2021). The similarity in performance between the accuracy and the F1 scores likely arises from the use of the AUC metric for the model calibration. The susceptibility maps Figs. 11 and 12 revealed that the RF method is relatively conservative, identifying a smaller area as highly susceptible (red zone) compared to Logistic Regression (LR) and Support Vector Machine (SVM) models. This suggests that while RF offers superior classification performance, it tends to minimize false positives by restricting the extent of predicted high-susceptibility zones. The high performance achieved for both hazards confirm that the adopted set of conditioning factors can effectively describe both processes. This suggests that the same modelling structure can be successfully applied to different types of hazards, allowing a homogeneous multi-hazard evaluation across the regional scale.

We overlaid the flash flood and landslide RF susceptibility maps into one as shown in Fig. 14 to identify areas susceptible to both. The classification used is found in Table of the Appendix A. There is a large overlap in terms of high susceptibility for both hazards mainly along the coast but also further inland along valleys (see DEM map in Fig. 4). The areas where the susceptibility is high for landslides but not for flash floods (dark green) is more spread out across Liguria as opposed to vice versa where the blue areas are mostly clustered in three areas (from left to right): (i) in the city of Albenga along the Centa River, (ii) close to Savona inland between two valleys, and (iii) along the Magra and Vara rivers including the surrounding urban areas (e.g. La Spezia). The higher precipitation received on the East (Sect. 2 Study Area) and the higher relative historical flash flood occurrences on the East (Fig. 1) can help explain the larger spatial extent of the rightmost blue cluster.

Nevertheless, the spatial susceptibility patterns for both hazards are mostly similar which is crucial to highlight as one hazard may amplify the other (Borga et al., 2014), e.g. a landslide may block a stream elevating the flood hazard. Gill and Malamud (2014) found that floods and landslides both trigger and increase each other's probability, hence, the importance of studying them together.

The susceptibility maps in Figs. 11 and 12 show a marked spatial correspondence between flash floods and landslides, mainly along the coastal areas and the main valleys of Liguria. These zones are characterized by steep slopes, low-permeability lithologies and strong anthropogenic pressure, which favour the simultaneous occurrence of both processes. Flash floods tend to be concentrated in low-elevation and urbanized valleys, whereas landslides extend toward the upper slopes and inland sectors. The two processes, although controlled by partly different conditions, share several predisposing factors. This confirms that they should be analysed within an integrated and unified framework. Our methodology, however, does not consider joined occurrence of flash floods and landslides as in e.g. Claassen et al. (2023) for which multi-hazard events are defined as events overlapping spatially with a time lag between their dates of occurrence. Due to the limited spatial scale of flash floods and landslides we expect a limited overlap but a more realistic event footprint could be studied such as slope units for landslides (Woodard et al., 2024) and sub catchments for flash floods (Yin et al., 2023).

The generated Random Forest landslide susceptibility map was compared against the European Landslide Susceptibility Map version 2 (ELSUS v2) (Joint Research Centre (JRC), European Commission, 2018; Wilde et al., 2018) at a 200 m resolution. The ELSUS map was developed using elevation, climatic conditions, slope, sub-surface lithology, and land cover datasets. Compared to European regions, the geohydrological vulnerability of Liguria is particularly high (Faccini et al., 2015). This is evident from the ELSUS map, in which most of the region is classified as having high or very high susceptibility. The Western section nearly completely has a high susceptibility. For the coastline, landslide susceptibility is lower than in the here generated maps. Possibly, for the ELSUS map more emphasis was paid to the slope and elevation that are both lower in these areas, leading to a lower landslide susceptibility.

Nevertheless, because the ELSUS dataset provides continental-scale classes, the spatial granularity for small regions like Liguria is limited. Although our study also uses a globally applicable framework, by focusing on a specific region, a more detailed and locally representative susceptibility map could be generated.

The ML framework is globally applicable; however, the accuracy of the generated susceptibility maps remains highly dependent on the completeness of hazard inventories. In the Liguria region, the frequency of occurrence of events is relatively high, and records are generally well documented. Moreover, the inventory prepared for this study relied on multiple sources to increase the coverage in space and time partially minimizing the spatial bias (Bornaetxea et al., 2023). Nevertheless, biases remain in the spatial distribution of recorded events with more observations near roads and in lower-elevation areas, reflecting the greater accessibility and population density in these zones. As a consequence, the susceptibilities are potentially less reliable for remote areas. The influence of the spatial bias could be explored by removing part of the inventory contributing to it and assessing the difference in susceptibility outcomes (Steger et al., 2017).

In addition, the framework has been applied and trained on the full period of data availability at once (only excluding scattered events from before 1940), while especially land use changes such as urbanization may have affected the landslide susceptibility in the region over time. To verify that the long inventory does not introduce temporal inconsistencies that may affect our conclusion, we derived new susceptibility maps for both hazards using only the more recent events (2000–2024) (refer to Fig. 3). We found no significant changes in the susceptibility maps with more limited changes for landslides than for flash floods (figures not shown).

Previous studies (Free et al., 2022; Modrick and Georgakakos, 2015; Uwihirwe et al., 2022) have shown that in many other, especially remote, regions of the world the number of recorded events is substantially lower. Global inventories can fill some gaps but are often coarse, incomplete or inconsistent in spatial and temporal coverage. This will affect the training of the ML algorithm and its accuracy, limiting the transfer of the framework to other data rich regions, or the transfer of the trained framework to regions with similar terrain, meteorological conditions, land cover, soil types, and land use including degree of urbanization. To overcome the latter, transfer learning for landslides susceptibility modelling in dissimilar areas was applied by Wang et al. (2022) by using Domain Adaptation (DA) in which a latent feature space is defined where the source and target areas have the same distribution. In particular, we expect physics-informed predictors like TWI, SPI, slope, curvature to be more easily transferable (after aligning their distributions) to other regions. On the other hand, elevation, proximity to roads, proximity to rivers, NDVI, land cover, and lithology are region dependent and could be transferred to similar regions using e.g. Case-Based Reasoning (CBR) (Wang et al., 2022). Alternatively, they may be harmonized to represent properties rather than classes e.g. grain size instead of lithology classes. Aspect on its own is not an informative variable but could be rederived into Windward-Leeward Index (WLI) that incorporates the influence of orographic precipitation. Still, there may be other aspects like wetness, seasonality and variability of precipitation that are not similar between regions affecting the accuracy of the transferred framework.

For the current study, the added value of extending the national inventories (AVI and ITALICA) with a more local data (ARPAL), supported by the use of Large Language Models to enable automatic information extraction is highlighted. This approach worked for Liguria, but relied on the available institutional event reports. In more data-sparse regions, collecting event data would be more difficult, but LLMs could potentially be applied for mining newspapers, social media and global disaster report archives.

The exact definition of flash floods influences the number and specifications of the events considered in the final inventory and hence the susceptibility maps. We note that our definition of flash floods for the AVI database is more restrictive than that of Vennari et al. (2016) which focused on Campania (to the South but also extending from the Sea to the Apennine) where they assumed all floods in the AVI inventory to be flash floods except the ones in alluvial plains. In our case, we explicitly used the description and the precipitation duration (if available) to select flash flood events, consequently 1149 entries with dates and without duplicates were excluded.

In this study all conditioning factors were static. Including dynamic parameters would enhance the capacity of the models to capture transient conditions and could potentially even allow the framework to update susceptibility in near real time. Moreover the framework holds the potential to be applied in early warning applications, as we previously explored in Uwihirwe et al. (2022), using high-resolution forecasted, now-casted precipitation (short-term). The same holds for future climate change assessments using high resolution precipitation projections from convection-permitting climate models (Zander et al., 2022). The framework would anyhow benefit from higher resolution precipitation products as these capture better the effect of the high spatial variability in heterogeneous terrain (Lee et al., 2022).

By integrating dynamic rainfall information, the framework could further support early warning applications through the definition of rainfall-triggering thresholds for both landslides and flash floods. Such thresholds, defined in a deterministic or probabilistic manner, could be coupled with the susceptibility component to provide time-dependent hazard estimates during forecasted precipitation events.

Finally, the hazard maps could be adapted to account for projected changes in rainfall regimes. By modifying the frequency or magnitude of triggering precipitation events in line with climate change scenarios, the framework would allow assessment of potential changes in landslide and flash flood hazard under evolving climatic conditions. The detailed implementation of these dynamic extensions was considered beyond the scope of the current study but represents a logical direction for future research.

While the present study focuses on relative susceptibility, the proposed modelling structure can be extended towards hazard assessment by incorporating information on the probability of occurrence of landslides and flash floods. In this way, the spatial susceptibility patterns derived here could be translated into estimates of event probability over a given time horizon (Wu and Yeh, 2020).

Although the proposed approach produced promising results, several limitations should be acknowledged. The current framework relies on static conditioning factors and does not yet incorporate dynamic variables such as rainfall, soil moisture, or land-use change. Applying the framework for longer-term climate projections will be more complex as, next to the climate change induced rainfall changes, the landscape and vegetation will also be influenced by anthropogenic forces which are even more uncertain. This could potentially make soils more erodible, increase the impermeable surface or even further reduce the network of natural waterways (Stalhandske et al., 2024) and thus require changes in the static maps underlying the framework.

Here the framework was applied in a multi-hazard setting, assessing both landslide and flash flood susceptibility, unfortunately, the inventories for both hazards were not equally sized (526 for flash floods and 2412 events for landslides). As the susceptibility mapping was done for both hazards independently this imbalance between hazards may not have affected the accuracy that much. Yet, the imbalance of the events in the inventory over time, with an increased number of events for later periods (Fig. 3), suggests that there may have been many non-included events in the past. After training this imbalance may reduce the number of false-positive events, as the model's ability to distinguish non-events becomes stronger. This also requires an analysis of historic changes in the landscape. In the past a more natural system with possibly more vegetation and less infrastructure could have been less susceptible to landslides and flash floods, thus the imbalance may also have a natural cause.

The fewer occurrences of events – whether landslides or flash floods – with respect to non-events lead to a class imbalance which can cause bias in the ML model by predicting the event (majority class) more often. One approach to handle this imbalance is to under sample the majority class and/or oversample the minority class, the latter is the more common in hazard mapping. A popular method is the synthetic minority oversampling technique (SMOTE) which for each event finds the k nearest neighbours and generates a random event at the chosen neighbour creating new events along the segment between the original event and the chosen neighbour. Han and Semnani (2025), who confirm the challenges of addressing class imbalance in landslides, found the best overall results for gridded hyperspace even sampling with a variant of SMOTE. However, this method can potentially create physically unrealistic events. Another approach is to apply class weights or cost sensitive weighting (Chen et al., 2004) in which the model is penalized for misclassifying the minority class. This means that by balancing the classes, a trade-off occurs: an increase of TP to the detriment of creating more FP. As previously indicated, our confusion matrix results are quite positive, limiting but not removing the need of applying a trade-off which could be studied in future work.

Despite these limitations, the results demonstrate the potential of the proposed framework for regional multi-hazard susceptibility assessment.