In the manuscript the authors compare modeled and measured soil moisture in the context of landslide prediction in Switzerland. Measurements are available at 35 sites across Switzerland, while the 1D soil moisture transfer model CoupModel providing estimates of soil moisture is set up at 133 sites. The landslide prediction model used is based on different soil moisture metrics and optimized by using ROC-statistics.

The authors carry out different comparisons/experiments: comparing measured and modeled soil moisture over 14 reference sites, comparing performances of the landslide prediction model based on measured or modeled soil moisture, studying the impact of boundary conditions of the hydrological model, and the impact of soil parametrization (using soil samples, uniform texture sets, or soil properties obtained with pedotransfer functions and the SoilGrid database).

The authors find that modeled soil moisture is outperforming measured moisture, that the model is sensitive to boundary conditions, and performances worsen as the distance between soil moisture and landslide locations increases.

The manuscript is clear, well written and structured, and the topic addressed is relevant. Overall, the manuscript is worth publication in hess.

Nevertheless, there is one important aspect which is worth addressing and discussing more in details, concerning the choice and results of the hydrological model. It has been shown in previous studies that vertical flow is the dominant process leading to landsliding, compared to lateral flow (e.g., Iverson 2000), but lateral flow becomes essential for adequate description of initial soil moisture conditions (e.g., Mirus et al., 2017, Leonarduzzi et al., 2020). This idea and the fact that the model seem to reproduce mostly just event dynamics is confirmed by several of the results:

• Figure 4: the better R2 for shallowest depth (typically better matching the patterns of meteorological forcing)
• Underestimation of seasonal variations of soil moisture
• Line 382-383: "resulting even in a slight forecast goodness increase for extreme and normal coarse-grained uniform-texture profiles”. Using a highly conductive soil which drains quickly, basically reduce the model to “get rid” of initial conditions and just represent the current infiltration event (i.e., your estimate of soil moisture is basically matching P-ET). But these soils are the one giving the worst match to measured soil moisture
• 9: modeled moisture outperforming measured moisture for event dynamics but not for antecedent condition metrics.

All these aspects, lead me to think that what is happening is that the hydrological model is actually just using the information in the recent meteorology (transformed into saturation estimate using soil parameters), while the “memory” component of saturation is not well represented/useful for landslide prediction. This is sort of the opposite of what one would expect in terms of information context in a “antecedent condition” metric, as typically saturation is considered a “cause”, while recent rainfall the “trigger”. The authors checked this by comparing the results to a simple rainfall-based prediction and indeed find similar performances. It would be worth exploring, or at least addressing, if using a combination of measured soil moisture (providing antecedent conditions) and rainfall event dynamics (possibly accounting also for soil properties), would actually lead to better landslide predictions (e.g., logistic regression using antecedent conditions saturation metrics measured/observed and rainfall event metrics). This would not invalidate the work presented or any of the findings but would definitely be a more complete/objective answer to the question the authors pose in the title: “Simulated or measured soil moisture: Which one is adding more value to regional landslide early warning?".

Finally, some minor comments:

• In Figure 5, are the lines showing the average across 14 sites for each depth?
• It could be interesting (although probably worth including only in the appendix), to see the trends of soil moisture observed and measured at different depths.

Iverson, R. M. (2000). Landslide triggering by rain infiltration. Water Resources Research, 36(7), 1897–1910. https://doi.org/10.1029/2000WR900090

Leonarduzzi, E., Maxwell, R. M., Mirus, B. B., & Molnar, P. (2021). Numerical analysis of the effect of subgrid variability in a physically based hydrological model on runoff, soil moisture, and slope stability. Water Resources Research, 57, e2020WR027326. https://doi.org/10.1029/2020WR027326

Mirus, B. B., Ebel, B. A., Loague, K., & Wemple, B. C. (2007). Simulated effect of a forest road on near‐surface hydrologic response: redux. Earth Sugace Processes and Landforms, 32(1), 126–142. https://doi.org/10.1002/esp.1387

The paper deals with a hot topic for the researchers working on landslide hazard management, i.e. the potential improvement of (shallow) landslide predictive models at regional scale, offered by adding soil moisture information to the commonly used precipitation data. The idea of comparing the predictive performance improvement deriving from modelled or from measured soil moisture is surely of interest for some of the readership of HESS.

The paper is clearly written and organized, the data and methods are adequately described, and much valuable information is made available to scientists dealing with landslide hydrology. Overall, my judgement of the paper is positive, with just few minor points that could help further improving it (you can find them as comments in the attached annotated file).

However, I would like to share some points of discussion with the authors, leaving to them the decision on if, and to what extent, they could find some space in the revised paper.

To judge about the added value of soil moisture information for landslide prediction (and for the comparison between two sources of information, measurements and modelling), there are two tasks: assessing the long-term water balance of the slopes, which is mainly controlled by what happens at the boundaries (overland runoff generation, evapotranspiration, deep leakage at the bottom of the soil cover); simulating what happens during rapid rainfall or snowmelt infiltration events, for which boundaries are expected to be less important compared to soil hydraulic properties. The first task has to do with the antecedent conditions, which may predispose the slopes to failure; the second directly with the triggering of landslides.

The results clearly indicate that modelling the long-term processes affecting the slopes (i.e. water balance) can be quite different than modelling the dynamics of the short-term infiltration of rainfall or snowmelt. In fact, the adopted 1D physically based model satisfactorily reproduces slope response during short periods with several infiltration events (fig. 3, although the soil parameters used in these simulations are not specified), but in the long run there is some systematic mismatch between simulated and measured soil moisture (fig. 5, looking at which I would be careful with the conclusion that there is not “apparent trend” in simulated soil moisture). The authors recognize that there is an issue in the modelling of evapotranspiration, which can be underestimated of even up to 200 mm per year (fig. 10), but surprisingly, they do not elaborate on this in their attempts of simulation. Instead, all the attention is focused on the effects of changing soil hydraulic parameters and on the bottom boundary condition.

All the variables tested to improve landslide predictions (table 2) refer to the water balance of the soil cover (i.e. mean cover saturation, infiltration flux at the upper boundary, bottom boundary condition), and so it is not surprising that the shape parameters of van Genuchten retention model do not affect the predictions that much. Given the tested variables, only Ksat is important, and this is probably the reason why the coarse-grained homogeneous soil profile works well, as this profile is associated with the highest value of Ksat (and thus infiltration), so that the saturation variations are mostly sensitive to rainfall are the highest. I think this is also the reason why it does not seem that including soil moisture changes the predictive performance so much compared to using rainfall information alone: the possible effects of antecedent conditions on infiltration dynamics are lost, as the model fails reproducing the water balance, and so soil saturation trend simply follows precipitation trend.  

About the bottom boundary condition, the most valuable for landslide predictions are the ones which maximize the drainage. I wonder if this is a way to compensate the underestimation of evapotranspiration: in Alpine environment (high altitude, rocky bedrock, steep slopes etc…), is it plausible that there is a groundwater table, few meters below the soil cover, affecting it? It seems to me that this bottom boundary conditions becomes more conceptual than physically based.

All in all, it is not surprising that this kind of soil moisture modelling does not add useful information for landslide prediction about antecedent conditions, being outperformed even by sparse field measurements, and that so little improvement is provided by the modelling of infiltration event dynamics (compared to precipitation alone).

Given all these considerations, some questions arise: is it worth using such a sophisticated unsaturated soil model, with so many equations and so many parameters difficult to set (table A1), when only the water balance of the soil cover is needed? Would the result of the comparison between modelled and measured soil moisture give a different result if a simpler modelling approach was used? Is it possible to conclude that the aims listed at lines 85-89 have been achieved?

Indeed, after the presented detailed study, with a rich dataset (landslides, soil properties, meteorological input, soil moisture measurements) and with a complex modelling exercise (exploring also the effects of different parametrizations), little conclusions are drawn: soil moisture measurements seem to allow a better assessment of antecedent conditions, but their use is limited by spatial resolution; soil moisture modelling requires different parametrization to provide better results. In view of this, maybe the aims of the study, and the title as well, could be reformulated in a less ambitious way.

I hope that my considerations can be of some help for the authors, for this paper or for future further elaborations of their data.

Roberto Greco