Landslide is a recurring geological hazard during rainfall seasons, which causes massive destruction, loss of life, and economic damage worldwide (Klose et al., 2014). The accurate prediction and monitoring of the spatio-temporal occurrence of the landslide is key to preventing and reducing casualties and damage to properties and infrastructure. One of the most widely adopted methods for landslide prediction is based on rainfall threshold and relies on building the rainfall intensity–duration curve using the information from past landslide events (Chae et al., 2017). However, such a method is in many cases insufficient for landslide hazard assessment (Posner and Georgakakos, 2015), because in addition to rainfall, the initial soil moisture condition is one of the main triggering factors of the events (Glade et al., 2000; Crozier, 1999; Tsai and Chen, 2010; Hawke and McConchie, 2011; Bittelli et al., 2012; Segoni et al., 2018b; Valenzuela et al., 2018; Bogaard and Greco, 2018).

For landslide applications, one potential soil moisture estimation method is through satellite remote sensing technologies. Although such technologies have been improved significantly over the past decade, their retrieving accuracy is still largely affected by frozen soil conditions (Zhuo et al., 2015a) and dense vegetation coverages, particularly in mountainous regions (Temimi et al., 2010); furthermore, the acquired data only cover the top few centimetres of soil. Although the more recently launched satellites such as Sentinel-1 (1 km and 3 d resolution) has shown some promising performance of soil moisture estimation (Gao et al., 2017; Paloscia et al., 2013), its availability only covers the recent years (Geudtner et al., 2014). Those disadvantages restrict the full utilisation of satellite soil moisture products for landslide monitoring applications as discussed in our previous study (Zhuo et al., 2019). In Zhuo et al. (2019), it is discussed that both the temporal and spatial resolutions of the ESA CCI satellite soil moisture product (Dorigo et al., 2017) is too coarse for landslide applications, and its data are mostly only available after the year 2002. Moreover, the shallow depth soil moisture observation from the satellite hinders the accuracy of landslide predictions. Therefore, other alternative soil moisture estimation methods need to be explored.

One emerging area relies on modelling. Some studies have used modelled soil moisture data for landslide applications (Ponziani et al., 2012; Ciabatta et al., 2016; Zhao et al., 2019a, b). However, to our knowledge, there is a lack of existing studies using modelled soil moisture from state-of-the-art land surface models (LSMs) for landslide studies, such as the Noah LSM (Ek et al., 2003) and the Community Land Model (CLM) (Oleson et al., 2010). LSMs describe the interactions between the atmosphere and the land surface by simulating exchanges of momentum, heat, and water within the Earth system (Maheu et al., 2018). They are capable of simulating the most important subsurface hydrological processes (e.g. soil moisture) and can be integrated with the advanced numerical weather prediction (NWP) system like WRF (Weather Research and Forecasting) (Skamarock et al., 2008) for comprehensive soil moisture estimations (i.e. through the surface energy balance, the surface layer stability and the water balance equations) (Greve et al., 2013). NWP-based (i.e. with integrated LSM) soil moisture estimations have many advantages. For instance their spatial and temporal resolution can be set at different scales depending on the input datasets to fit various application requirements; their coverage is global, and the estimated soil moisture data cover multiple soil layers (from the shallow surface layer to deep root-zones); and a number of globally covered data products can provide the necessary boundary and initial conditions for running the models. Soil moisture estimated through such an approach has been widely recognised and demonstrated in many studies, which cover a broad range of applications from hydrological modelling (Srivastava et al., 2013a, 2015), drought studies (Zaitchik et al., 2013), and flood investigations (Leung and Qian, 2009), to regional weather prediction (Stéfanon et al., 2014). Therefore, NWP-based soil moisture datasets could provide valuable information for landslide applications. However, to our knowledge, relevant research has never been carried out.

The aim of this study is hence to evaluate the usefulness of NWP-modelled soil moisture for landslide monitoring. Here the advanced WRF model (version 3.8) is adopted, because it offers numerous physics options such as micro-physics, surface physics, atmospheric radiation physics, and planetary boundary layer physics (Srivastava et al., 2015), and it can be integrated with a number of LSM schemes, each varying in physical parameterisation complexities. So far there is limited literature comparing the soil moisture accuracy of different LSMs options in the WRF model. Therefore, in this study, we select three of the WRF's most advanced LSM schemes (i.e. Noah; Noah-Multiparameterization, here Noah-MP; and CLM4) to compare their soil moisture performance for landslide hazard assessment. Furthermore, since all three schemes can provide multi-layer soil moisture information, it is useful to include all those simulations for the comparison so that the optimal depth of soil moisture could be determined for the landslide monitoring application. In order to compare with the performance of our previous study on using the satellite soil moisture data (Zhuo et al., 2019), the same study area, Emilia Romagna, is used here. The study period covers 10 years from 2006 to 2015 to include a long-term record of landslide events. In addition, because slope angle is one of the major factors controlling the stability of the slope, it is hence used in this study to divide the study area into several slope groups, so that a more accurate landslide prediction model could be built.

The description of the study area and the datasets used are included in Sect. 2. Methodologies regarding the WRF model, the related LSM schemes and the adopted landslide threshold evaluation approach are provided in Sect. 3. Section 4 shows the WRF soil moisture evaluation results against the in situ observations, and the WRF rainfall evaluations over the whole study area. Section 5 covers the comparison results of the WRF-modelled soil moisture products for landslide applications. The discussions and conclusions of the study are included in Sects. 6 and 7, respectively.

In this study, the evaluation is based on the daily mean soil moisture. The reason for not using the antecedent soil moisture condition plus rainfall data on the day is because the purpose of this study is to explore the relationship between different WRF-simulated soil moisture and landslides only. In general, soil moisture is a predisposing factor for slope instability, while rainfall is the triggering factor. The same rainfall may trigger or may not a landslide depending on the soil moisture content at the time of the rainfall event. The mean soil moisture on the day of the landslide implicitly account for both the initial soil moisture and the effective rainfall absorbed by the ground, and can be a robust indicator of the hydrological condition of the slope.

As introduced at the beginning of the paper, previous works (as discussed in the introduction section) have demonstrated that in complex geomorphologic settings (e.g. in Emilia Romagna), a rainfall threshold approach is too simple, and more hydrologically driven approaches need to be established. This section is to assess whether the spatial distribution of soil moisture can provide useful information for landslide monitoring at the regional scale. Particularly, all three soil moisture products simulated through the WRF model are used to derive threshold models, and the corresponding landslide prediction performances are then compared statistically. Here the threshold is defined as the crucial soil moisture condition above which landslides are likely to happen.

Among different factors for controlling the stability of slope, the slope angle is one of the most critical ones. From the slope angle map in Fig. 2, it can be seen the region has a clear spatial pattern of high and low slope areas, with the majority of the high-slope areas (which can be as steep as around 40∘) located in the mountainous southern part and the river valleys. Based on the event data analysed, the landslides that happened during the study period are mainly located in the high-slope region, with a particularly high concentration around the central southern part. The spatial distribution of the landslide events is also in line with the overall geological characteristics of the region, i.e. the southern part mainly constitutes the outcrop of sandstone rocks that make up the steep slopes and are covered by a thin layer of permeable sandy soil, which are highly unstable. Therefore, instead of only using one soil moisture threshold for the whole study area, it is useful to divide the region into several slope groups so that within each group a threshold model is built. To derive soil moisture threshold individually under different slope conditions, all data have been divided into three groups based on the slope angle (0.4–1.86; 1.87–9.61; 9.52–40.43; since no landslide events are recorded under the 0–0.39 group, the group is not considered here). As a result, all groups have equal coverage areas. There are different ways to group the slopes. In this study, in order to have equal coverage areas, we have identified these class-break values.

In order to find the optimal threshold so that there are few overestimations (i.e. threshold is overestimated) and false alarms (i.e. threshold is underestimated), we test out 17 different exceedance probabilities from 1 % to 50 %. For each LSM scheme, the total number of threshold models is 204, which is the result of different combinations of slope groups, soil layers, and exceedance probability conditions. The calculated thresholds for all LSM schemes under three slope groups are plotted in Fig. 7. Overall there is a clear trend between the slope angle and the soil moisture threshold, i.e. the threshold becomes smaller for steeper areas. The correlation is more evident at the upper three soil layers (i.e. the top 1 m depth of soil), with only a few exceptions for Noah and CLM4 at the 1 % and the 2 % exceedance probabilities. At the deep soil layer centred at 150 cm, the soil moisture threshold difference between slope group (SG) 2 and 3 becomes very small for all three LSM schemes. This could be partially because at the deep soil layer, the change of soil moisture is much smaller than at the surface layer, and therefore the soil moisture values for SG 2 and 3 could be too similar to differentiate. However, for gentler slopes (SG 1), the higher soil moisture triggering level always applies even down to the deepest soil layer for all three LSM schemes. In this study, the results show that wetter soil is more likely to trigger landslides on gentler slopes than on steeper slopes.

All the threshold models are then evaluated under the 45 selected rainfall events (Table 5) using the ROC analysis. Each threshold determined for each of the slope class during the calibration is used for the evaluation. The period of the selected rainfall events is between 1 and 18 d, and the average rainfall intensity ranges from 5.05 to 24.69 mm d-1. The resultant Euclidean distances (d) between each scenario of exceedance probability and the optimal point for ROC analysis are listed in Table 6 for all three WRF LSM schemes at the tested exceedance probabilities. The best performance (i.e. lowest d) in each column (i.e. each soil layer of an LSM scheme) is highlighted. In addition, the d results are also plotted in Fig. 8 to give a better view of the overall trend amongst different soil layers and LSM schemes. From the figure, for all three LSM schemes at all four soil layers, there is an overall downward and then stabilised trend. Overall for Noah, the simulated surface layer soil moisture provides better landslide monitoring performance than the rest of the soil layers from 1 % to 35 % exceedance probabilities; the scheme's worst performance is observed at the third soil layer, centred at 70 cm. The values of d for Noah's second and fourth layer are quite close to each other. For Noah-MP, the simulated surface layer soil moisture gives the best performance amongst all four soil layers for most cases between the 1 % and 35 % exceedance probability range; the scheme's worst performance is observed at the fourth layer. Unlike Noah, all four soil layers from the Noah-MP scheme provide a distinct performance amongst them (i.e. larger d difference). For CLM4, the performance for the surface layer is quite similar to the second layer's, and the differences between the four layers are small. From the Table 6, it can be seen that for Noah the most suitable exceedance probabilities (i.e. the highlighted numbers) range between 35 % and 50 %; for Noah-MP they are between 30 % and 50 %, and for CLM4 it stays at 40 % for all four soil layers. For both Noah and Noah-MP, the best performance is observed at the surface layer (d=0.392 and d=0.369, respectively). For CLM4, the best performances show no distinct pattern amongst soil layers (i.e. the best performance is found at the soil Layer 3, followed by Layers 2, 1, and 4). Of all the LSM schemes and soil layers, the best performance is found for Noah-MP at the surface layer with 30 % exceedance probability (d=0.369). Based on the d results, WRF-modelled soil moisture provides better landslide prediction performance than the satellite ESA-CCI soil moisture products as shown in our previous study (Zhuo et al., 2019), i.e. d=0.51). The ROC curve for the Noah-MP scheme at the surface layer is shown in Fig. 9. In the curve, each point represents a scenario with a selected exceedance probability level. It is clear that with various exceedance probabilities, FAR can be decreased without sacrificing the HR score (e.g. 4 % to 10 % exceedance probabilities). At the optimal point at the 30 % exceedance probability, the best results for HR and FAR are observed as 0.769 and 0.289, respectively.

In this study, the best landslide prediction performance for Noah and Noah-MP follows a regular trend: the deeper the soil layer, the poorer the landslide monitoring performance. There are several potential reasons for such an outcome. First, the simulated soil moisture accuracy at the shallower layers is better than that in the deeper zones. Second, although the wetness conditions at the sliding surface are important, the soil moisture above it is also important (i.e. the loading should be heavier with more water in the upper soil layer). Third, the landslides occurring in the region are mainly in the top shallow soil layer. Fourth, the WRF-modelled soil moisture is not accurate enough in assessing the landslide events in the study region. In order to find out the exact reasons, comprehensive studies with more detailed landslide event datasets are needed in the future.

For the WRF soil moisture evaluation, clearly the evaluation work based on a single soil moisture sensor located in a plain area is not sufficient to derive conclusions about the model's performance over the whole study region. Therefore, the results here are preliminary. However, in this study, by introducing the WRF spatial soil moisture information into the landslide prediction model, the performance has indeed been improved in comparison with our previous study using the satellite remote sensing soil moisture data (Zhuo et al., 2019). A similar concept has been carried out by Segoni et al. (2018b), who implemented the soil moisture information simulated from a hydrological model into a regional landslide early warning system with clear improvements in performance with regard to false alarms or missed alarms (i.e. when a hazard occurred but no early warning was provided). Although the results shown in this study are preliminary and confined to the study area, the improved landslide prediction performance is already obtained. Therefore, it is hoped that with more globally available and dense soil moisture network data and further refinements of the method, the results could be improved further.

In addition, ideally, it will be useful if there is a dense soil moisture sensing network covering the whole study area. In reality, that is not practical, so we have to rely on the spatial soil moisture information by other means. So far, the soil moisture data with the best spatial and temporal resolution is from the WRF model. One question that arises is how representative a single soil moisture sensor can be for the whole study area. We have carried out the correlation study of a single sensor with the whole study region (using the Noah-MP top-layer soil moisture data). As seen in Fig. 10a, the study region is divided into 44 equally spaced grids (30 km apart), with the grid centres marked as black crosses. The initial assumption is that the soil moisture sensor can only represent its adjacent area, but the result was a surprise (Fig. 10b). Based on the outcome, a single-point sensor can represent a significant proportion of the region. Admittedly, there are some areas where the correlations are poor, in particular Grid 27, which has been compared with its surrounding four grids as shown in Fig. 11. It can be seen the soil moisture variation at Grid 27 is totally different in comparison with that of the four surrounding grids. The unique soil moisture variation pattern observed in Grid 27 may be caused by different land use and soil type in that area, but clearly further studies are needed to find out the exact reasons. The aforementioned work has prompted us to carry out a future study on the optimal soil moisture sensor network design for landside applications. Although there are numerous studies on the rain gauge network design by the research community, the soil moisture sensor network design has been largely ignored by the community. Hence, this study has paved a foundation for such research.

For the WRF rainfall evaluations, the results are not good. Rainfall is one of the main drivers of soil moisture change, and it is logical to think soil moisture and rainfall are highly linked. However, since rainfall is high-frequency data while soil moisture is low-frequency data, they behave differently. The results illustrate that for landslide study, it is better to use the WRF soil moisture data than its rainfall data. Clearly more studies are needed to confirm this assumption.

Here, WRF is modelled based on the ERA-Interim datasets; however, it has been found in Albergel et al. (2018) that the performance of the ERA5 has surpassed the ERA-Interim. Therefore, the ERA5 datasets will be tested in our future studies. Model-based soil moisture estimations could be affected by error accumulation issues, especially in the real-time forecasting mode. A potential solution is to use data assimilation methodologies to correct such errors by assimilating soil moisture information from other data sources. Since in situ soil moisture sensors are only sparsely available in limited regions, soil moisture measured via satellite remote sensing technologies could provide useful alternatives. Another issue is with the landslide record data, as most of them are based on human experiences (e.g. newspapers and victims) and thus a lot of incidences could be unreported. Therefore, the conclusion made here could be biased. Other ways of expanding the current landslide catalogue can depend on automatic landslide detection methods based on remote sensing images (Nichol and Wong, 2005; Chen et al., 2018), internet new sources (as all landslides with a relevant impact on society will be reported on internet new sources), and automatic web data mining methods (Battistini et al., 2013; Goswami et al., 2018).

In this study, the usability of WRF-modelled soil moisture for landslide monitoring has been evaluated in the Emilia Romagna region based on the research duration between 2006 and 2015. Specifically, the four-layer soil moisture information simulated through the WRF's three most advanced LSM schemes (i.e. Noah, Noah-MP, and CLM4) is compared for the purpose. Through the temporal comparison with the in situ soil moisture observations, it has been found that all three LSM schemes at all four soil layers can produce the general soil moisture's seasonal cycle. However, only Noah-MP is able to simulate the large soil drying phenomenon close to the observations during the drying season, and it also has the highest correlation coefficient and the lowest RMSE at most soil layers amongst the three LSM schemes. However, it should be noted, the soil moisture evaluation is only based on a single-point-based soil moisture sensor that is available in the plain region of the study area. Therefore, the WRF soil moisture performance over the whole study region, in particular at the mountainous zone, cannot be evaluated in this study. Since soil moisture is related to rainfall, we have carried out the WRF rainfall assessments, based on the comparison with the dense rainfall network in the region. The results have shown that there is no distinct difference between the three LSM schemes. The WRF rainfall performance is found to be similar to a study carried out in the central USA (Van Den Broeke et al., 2018). A landslide prediction model based on soil moisture and slope angle condition is built up, and 17 various exceedance probably levels between 1 % and 50 % are adopted to find the optimal threshold scenario. Through the ROC analysis of 612 threshold models, the best performance is obtained by the Noah-MP at the surface soil layer with 30 % exceedance probability.

In summary, this study provides an overview of the soil moisture performance of three WRF LSM schemes for landslide hazard assessment. Based on the results, we demonstrate that the surface soil moisture (centred at 10 cm) simulated through the Noah-MP LSM scheme is useful in predicting landslide occurrences in the Emilia Romagna region. With the hit rate of 0.769 and the false alarm rate of 0.289 obtained in this study, such soil moisture information has the potential to provide landslide predictions through the use of rainfall data. Further study on the soil moisture representation of a single soil moisture sensor over a large region has also been carried out. The results demonstrate that although there is a significant elevation difference in the region, a single soil moisture sensor has a high correlation with a significant proportion of the study area. Although there is still a small proportion of areas where the correlation is poor, this has prompted us to carry out a future study on the optimal design of soil moisture sensor network for landslide study.

One must bear in mind that although the results demonstrated in this study are only valid for the selected region, the methodology could be generalised to derive site-specific calibrations in other sites using the proposed approach. In order to make a general conclusion, more research is needed using the methodology described in this paper. Particularly, a considerable number of catchments with a broad spectrum of climate and environmental conditions and dense soil moisture sensor networks will need to be investigated.