This manuscript investigates flash flood and landslide susceptibility using a unified machine-learning framework applied to the Liguria region in Italy. The study addresses an important topic in the context of multi-hazard risk assessment and benefits from the use of multiple event inventories, a consistent modelling framework across hazard types, and a careful discussion of several limitations. Overall, the paper is technically sound and clearly written, and it has the potential to make a valuable contribution to regional-scale multi-hazard susceptibility analysis. The comments below are intended to help strengthen the methodological clarity, interpretability, and positioning of the contribution.

1. The flash flood inventory spans a very long time period (approximately 1950 to present), whereas the susceptibility modelling relies on present-day static conditioning factors (e.g. land cover, river network, road networks, …). Over this period, substantial changes have occurred in Liguria, particularly related to urbanisation, river modification, and land-use change, which may violate the assumption of stationarity between past events and current susceptibility conditions. While the long inventory improves statistical robustness, it may also introduce temporal inconsistencies that affect the interpretation of the resulting susceptibility maps, especially in urbanised valleys where many early events predate current conditions. This issue is particularly relevant for flash floods, which are highly sensitive to anthropogenic hydrological modifications. The authors are encouraged to (i) acknowledge this limitation more explicitly, and/or (ii) demonstrate that it does not substantially affect the results. Possible approaches could include comparing susceptibility maps derived from a more recent subset of events (e.g. post-1990) with the full inventory-based map, or applying temporal weighting to events based on their distance from present-day conditions.
2. The Introduction states that this study transfers the framework of Tehrani et al. (2021) to flash flood susceptibility and applies it within a unified multi-hazard context. However, the precise methodological and conceptual advances beyond the original framework are not sufficiently articulated. The authors are encouraged to clarify more explicitly what aspects of the framework are novel (e.g. unified hazard treatment, inventory construction, comparative analysis), and how this work differs from or extends Tehrani et al. (2021) beyond a change in hazard type and study area. Strengthening this framing would help readers better appreciate the innovative potential of this study.
3. The title emphasises a Mediterranean-scale application, while the analysis is conducted at a local/regional scale focused on Liguria. While Liguria is representative of Mediterranean hydro-geomorphic settings, the current title may overstate the spatial scope of the application. The authors may wish to consider revising the title to better reflect the local-scale case study, while still acknowledging Mediterranean relevance.
4. In the Methodology section, the manuscript states that hyperparameters were calibrated using grid search and cross-validation, but the specific parameter ranges and final selections for each algorithm (LR, RF, SVM) are not described in detail. For transparency and reproducibility, the authors are encouraged to provide: the tested hyperparameter ranges, and the final selected values for each model, possibly in an appendix or supplementary material. This would strengthen the methodological clarity of the paper.
5. Class imbalance: the flash flood and landslide inventories differ substantially in size, and the binary classification problem is inherently unbalanced (event vs non-event pixels). While model performance metrics (AUC, accuracy, confusion matrices) are reported, the role of class imbalance in influencing false positives and false negatives is not explicitly discussed. The authors are encouraged to either: discuss how class imbalance may affect the reported performance metrics, or justify the decision not to apply balancing techniques (e.g. class weighting, SMOTE). Explicitly addressing this point would improve confidence in the interpretation of model performance.
6. To account for location uncertainty in point-based inventories, a 3x3 pixel window is used around each event. While this is a reasonable approach, it may introduce spatial smoothing that blurs the influence of certain conditioning factors, particularly for landslide susceptibility in steep terrain. As already partially acknowledged by the authors, this may contribute to unexpected results (e.g. landslides associated with gentler slopes). The authors may consider discussing alternative or complementary strategies, such as: exclusion or buffer zones around event pixels to reduce spatial autocorrelation, or sensitivity tests on window size. If considered too technical for implementation, a clearer discussion of the implications of this choice would still be valuable.
7. The framework is described as globally applicable; however, its performance depends strongly on the availability and quality of local hazard inventories. The authors are encouraged to expand the discussion on spatial and temporal transferability, particularly regarding:

• data-rich versus data-scarce regions;
• which conditioning variables may be problematic when transferring the model (e.g. the authors mention aspect -southward facing slopes- dependence, how would that change in regions with different dominant exposure?);
• the potential role of high-resolution dynamic inputs (e.g. convection-permitting reanalysis and future projections) in improving applicability across regions and time periods.

This would help better position the framework for broader use beyond the Liguria case study.