In the study area, an integrated hydro-meteorological monitoring station was installed on 27 March 2012 in a test-site slope located near the village of Montuè (municipality of Canneto Pavese; Fig. 1).

The slope is characterized by a medium–high topographic gradient (between 22 and 35∘), faces the east and descends towards a rather small and narrow valley formed by a creek (Eastern Frate Creek). Its altitude ranges from 210 to 120 m a.s.l. The difference in elevation between the top of the monitored slope and the valley floor is 132 m, while the difference in elevation between the position of the monitoring station and the Rio Frate Creek is 63 m. This slope was chosen as a test-site slope within the study area by considering a series of criteria:

the presence of triggering zones of shallow landslides that occurred in April 2009;

The geotechnical characterization of the slope deposits was based on standard soil analyses carried out according to the ASTM (American Society for Testing and Materials) standards. The performed tests included (i) an assessment of the physical parameters of the materials (grain size distribution, bulk density, Atterberg limits) and (ii) triaxial tests which allowed the determination of the shear strength parameters in terms of the effective stresses.

The soil derives from the weathering of sands and poorly cemented conglomerates belonging to the Rocca Ticozzi Conglomerates. Individual soil horizons along the slope have a clayey sandy silt texture, with high silt contents ranging from 51.1 to 65.6 %, clay content between 21.3 and 29.0 % and varying amounts of gravel and sand (Table 1). By analysing the clay soil fraction (< 2 µm) through X-ray diffraction tests, the clay mostly appears to consist of smectite and chlorite. In particular, smectite constitutes approximately 50 % of the finest soil fraction and approximately 10–15 % of the solid particles in the studied soils. The weathered bedrock 1.4 m from the ground, labelled as We. Bed. in Fig. 5, consists of a sand lens with a sand content of 75.0 % (Table 1).

According to the USCS classification, the soil horizons are prevalently non-plastic or slightly plastic (CL). The liquid limit (wL) ranges from 38.5 to 41.9 %, while the plasticity index (PI) ranges from 14.3 to 17.1 %, and both remain steady along the depth (Table 1). The unit weight (γ) increases between 0.6 and 1.0 m b.g.l. (below the ground level) from 16.7 to 18.6 kN m-3 and then remains steady with depth (Table 1).

The peak shear strength parameters were reconstructed at different depths through triaxial tests. Up to 1.0 m from the ground, the soil horizons have a friction angle (φ′) between 31 and 33∘ and nil effective cohesion (c′) (Table 1). At 1.2 m from the ground, the soil level is characterized by a friction angle equal to 26∘ and effective cohesion of 29 kPa (Table 1).

The hydrological properties of the different soil horizons were determined through a laboratory reconstruction of the soil water characteristic curve (SWCC) and the hydraulic conductivity function (HCF). These functions were reconstructed through a combination of the Wind–Schindler method (WSM; Schindler, 1980; Peters and Durner, 2008) technique (Hyprop, UMS GmbH, Munich, Germany) with a vapour pressure method (VPM; Rawlins and Campbell, 1986) device (WP4T, Decagon Devices, Pullman, WA) on undisturbed soil samples. The experimental data were fitted by the Van Genuchten (1980) and Mualem (1976) models. The parameters of these models (saturated water content θs, residual water content θr, fitting parameters α and n and saturated hydraulic conductivity Ks) were then estimated using the Marquardt (1963) algorithm. All the soil levels had similar values of α (0.007–0.013 kPa-1) and θr (1–3 %). The n and θs parameters were slightly higher until 0.6 m from the ground with respect to the deeper levels. Moreover, Ks was quite steady at approximately 1.0 and 2.5 × 10-6 m s-1 except for the soil level at 1.2 m from the ground, which was less permeable (0.5 × 10-6 m s-1; Table 2).

A detailed description of the monitoring station is reported elsewhere (Bordoni et al., 2014; Fig. 4). In this paper, the necessary information required for completeness is provided. The station collects data with a time resolution of 10 min. The following meteorological parameters are measured: rainfall depth, air temperature, air humidity, atmospheric pressure, net solar radiation, wind speed and direction. Some probes are installed in the soil and weathered bedrock at different depths to measure the soil water content and soil pore water pressure. In particular, six time domain reflectometer (TDR) probes installed at 0.2, 0.4, 0.6, 1.0, 1.2 and 1.4 m from the ground level measure the soil water content, while a combination of three tensiometers and three heat dissipation (HD) sensors installed at depths of 0.2, 0.6 and 1.2 m measure the soil's pore water pressure. The HD sensors only allow pore water pressures lower than -101 kPa (Bittelli et al., 2012) to be acquired; thus, tensiometers are installed alongside the HD sensors to measure pore water pressures above -101 kPa.

The monitoring equipment allowed us to identify the soil's main hydrological behaviours, in particular the soil's response to the different seasonal rainy conditions and various rainfall intensities. The test-site slope can adequately represent the geotechnical and hydrological features of slopes affected by shallow landslides over the entire study area. For this reason, the data from continuous monitoring of the sample slope can be useful to identify the soil's hydrological conditions that can lead to the triggering mechanisms of shallow landslides in similar conditions. The field data can be used to infer the soil conditions for periods without monitoring to evaluate the prediction ability of a physically based model such as TRIGRS-unsaturated, which is used to identify the triggering zones of shallow landslides at the local scale.

To identify the triggering zones of shallow landslides using physically based models, the distributions of the main geotechnical and hydrological properties are required as input parameters to obtain the trend of the slope safety factor (Fs) in time. However, transitioning from the site-specific scale, where detailed field and laboratory test results can be available, to a local scale, where data can be available only for a limited number of sites, it is important to establish which characteristics should be considered “constant” to enable the assumption that the available data are representative.

To guess an answer to this question, the TRIGRS model was implemented in the study area using different types of mapping units (geological and pedological), the class distribution of which across the study area is represented in Fig. 3.

The main geotechnical and mechanical soil properties were assigned to each unit after performing an averaging procedure of the data collected through laboratory tests on 160 soil samples taken from different sites in the study area. In this group, 114 disturbed soil samples were used to determine the Atterberg limits and grain size distribution curves, while 52 undisturbed soil samples were used to measure the soil's unit weight. In the group of undisturbed soil samples, which were taken at a depth where the failure surface of the shallow landslides developed, 18 samples were used to measure the peak shear strength parameters through direct shear tests, and 3 samples were used for triaxial tests. As observed in the monitored slope, no significant changes in the geotechnical properties, particularly for the grain size distribution and Atterberg limits, were identified along the depth in the soil levels.

The main differences between the classes for each mapping unit type were linked to the grain size distribution. In fact, the geological and pedological units could be distinguished based on the sand and clay amounts. The soils derived from the weathering of the Monte Arzolo Sandstones and Rocca Ticozzi Conglomerates were classified as clayey-sandy silt because the amount of sand is generally more than 15 % (Table 3). In contrast, the soils derived from the weathering of the Sant'Agata Fossili Marls and Gessoso–Solfifera Formation were classified as clayey silt due to a sand content generally lower than 10 % and a prevalent silt content (Table 3). Moreover, the soils derived from the weathering of the Gessoso–Solfifera Formation exhibited a significantly higher clay content (37.1 %) than the other units (Table 3). The BRS1 and MRL1 pedological units contain groups of soils with clayey silt texture (Table 3), while the ILM1/RUM1 soils are clayey-sandy silt due to a mean sand content of 19.3 %. In contrast, the FGE1 class contains soils with a clayey-silty texture due to similar mean silt and clay contents (46.9 and 43.3 % respectively; Table 3).

According to the USCS classification, the majority of the classes in all the mapped units are grouped into non-plastic or slightly plastic soils (CL), with a mean liquid limit wL that ranges between 39.7 and 43.9 % and a mean plasticity index PI that ranges between 18.1 and 22.7 % (Table 3). Only the FGE1 pedological class presents highly plastic soils (CH) with a mean wL of 52.4 % and a mean PI of 31.8 % (Table 3).

The unit soil weight γ is similar among the different classes for each mapping unit (Tables 1–3): the mean values range between 17.0 and 18.1 kN m-3. Furthermore, the shear strength parameters of the soils, i.e. the peak friction angle φ′ and the effective cohesion c′, are quite similar for all the considered soils (Table 3). In particular, φ′ ranges between 24 and 27∘, while c′ remains between 1.2 and 2.0 kPa.

The standard deviation (SD) value of each soil property for each unit has also been provided. For both geological and pedological units, the highest values of SD were measured for the liquid limit wL, the sand and silt content and the shear strength parameters φ′ and c′ (Table 3). In particular, c′ reaches SD values on the order of 80–100 % of the mean value (Table 4).

A similar procedure was performed to assign the hydrological parameters (in terms of SWCC) to the different selected units. In particular, the Rosetta pedotransfer function model (Schaap et al., 2001) was applied to the grain size distribution of the soil samples to determine the parameters of the SWCCs and HCFs of the materials of each identified class according to the models of Mualem and Van Genuchten (Table 4). The average values of the Mualem and Van Genuchten models' parameters (θs, θr, α, n, Ks) are very similar between the classes (Table 4). In contrast, the SD values are quite high, between 20 and 70 % of the mean value (Table 4). The high SD of each mapping unit is linked to the high SD value of the soil fractions (gravel, sand, silt and clay), which are required in the Rosetta model to assess the SWCC properties.

The WRCs and HCFs were also reconstructed for eight undisturbed samples taken in the study area through the same methods (WSM and VPM) used for the soil samples from the monitored slope. For these soils, the reconstructed Mualem and Van Genuchten model parameters are confident with respect to the values modelled through the Rosetta pedotransfer function for θs, α, n and Ks: in fact, the percentage of the mean error (Er) ranges between 2.3 and 9.9 % (Fig. 5a, c, d, e). Only the θr values modelled through Rosetta are notably higher than the values measured through the Hyprop technique (Er = 50.1 %, Fig. 5b). Generally, this error is not so large considering the low values that characterize the studied soil (e.g. measured θr of 2 % against an estimated θr of 4 %).

This aspect confirms the reliability of modelling the soil's hydrological properties in the study area through the Rosetta model, correctly identifying the mean values of these properties to be assigned to the selected mapping unit classes.

The TRIGRS-unsaturated model (Baum et al., 2008) has been used for the analyses to assess the triggering zones of shallow landslides in the study area. A brief description of the main principles of this physically based model is provided in Appendix A.

This model has been applied to different rainfall events measured by the monitoring station installed in the study area during its activity. The modelled pore water pressures at two depths, 0.6 and 1.2 m from the ground, for each considered rainfall event along the monitored slope were then compared with the values measured during the same rainfall at the monitoring station. The goodness of the TRIGRS-unsaturated model's fit to the pore water pressure modelling was evaluated with the root mean square error (RMSE) statistical index, expressed in Eq. (1) as RMSE=∑i=1nψo,i-ψm,i2n, where ψo is the observed water pore water pressure at the ith hour of the considered rainfall events in Tables 7 and 8; ψm is the pore water pressure estimated by the model at the same ith hour of the same event; and n is the number of observations, which corresponds to the duration of the rainfall. As the RMSE value approaches 0 kPa, the prediction model becomes increasingly effective and accurate. Only a comparison between the measured and modelled pore water pressure trends at different depths by the monitoring station on the sample slope has been considered.

Moreover, TRIGRS-unsaturated was applied considering the benchmark rainfall event on 27–28 April 2009. Referring to a real rainfall event, these analyses provided the assessment of shallow landslide triggering zones taking into account the two different types of mapping units, namely geological and pedological. The choice to also consider the pedological unit is linked to the distinction of the soils in the study area based on their pedological features, which can be connected to different pedological processes that can directly influence the physical and hydrological behaviour of a soil (Baumhardt and Lascano, 1993). Two indexes from the receiver operating characteristic analysis (Hosmer and Lemeshow, 2000; Zizioli et al., 2013) have been used to evaluate the predictive capability of the reconstructed models: the “true positive rate” (TP) and the “false positive rate” (FP). The TP is the ratio (in percentage) between the number of elementary cells computed as unstable (safety factor < 1.0) by the model and the number of elementary digital elevation model (DEM) cells occupied by shallow landslide triggering areas. In contrast, the FP represents the ratio (in percentage) between the number of elementary cells computed as unstable that do not correspond to observed shallow landslide triggering zones and the number of elementary cells in the study area not affected by the phenomena and thus considered stable (safety factor > 1.0).

The results of the reconstructions through the TRIGRS-unsaturated model were also compared, in terms of the TP and FP, with the results obtained in a previous work through the TRIGRS-saturated (Baum et al., 2002), SINMAP (Pack et al., 1999) and SLIP (Montrasio and Valentino, 2008) models in the same study area and for the same event by Zizioli et al. (2013).

As reported in the introduction, the first step to appropriately model rainfall-induced landslides at both the site-specific and local scales is a detailed description of the physical–mechanical triggering mechanism in relation to the site-specific characteristics of the involved soils and stratigraphy. Data from the monitoring station were used to determine the dynamics of the soil's water content and soil's pore water pressure at the test site in relation to the characteristics of the different soil levels and the weathered bedrock (Fig. 6). The monitored hydrological behaviours can adequately represent the typical conditions that characterize the surrounding study area. In this work, the period between 27 March 2012 and 1 October 2014 was analysed.

The average hourly values of the water content and pore water pressure were considered. Because the tensiometer at 0.2 m from the ground level broke, the pore water pressure in the range between 0 and -101 kPa was not measured at this depth from November 2012 until the end of the analysis period. No data were acquired between 10 and 15 January 2014 because the station's alimentation system did not function correctly.

In the analysis period, the water content ranged between 10 and 45 % in the topsoil and between 15 and 38 % in the weathered bedrock. In contrast, the pore water pressure ranged from positive values to 12.7 kPa in the G horizon to values on the order of -103 kPa.

The installed tensiometers required a correction for the measured values due to the height of the water present in the column of the instrument, with an increase of 1 kPa for each 0.1 m of depth in the soil. For this reason, it is possible to measure positive pore water pressure values, as already shown in previous works (Zhan et al., 2006).

By analysing the data acquired over 2 years of monitoring, it is immediately clear that the water content and pore water pressure dynamics are strictly connected to rainfall trends and different hydrological behaviours can be identified in the soil profile (Fig. 6).

The soil horizons within 0.6–0.7 m of ground level had a faster response than the deepest soil horizons to long, dry or long, wet periods. In the summer months, the water content and pore water pressure decreased faster in the most shallow soil horizons than in the deeper ones (Fig. 6) due to evapotranspiration effects and water uptake from the roots of grass and shrubs. Changes in the hydrological parameters are less rapid in soil levels deeper than 0.6–0.7 m below the surface and in weathered bedrock (Fig. 6); this different behaviour is linked to the fact that these levels are less affected by evapotranspiration and root zone effects. The shallowest soil horizons quickly became wet (Fig. 6) during rather prolonged rainy periods following dry periods, such as in autumn months, and after rainfall events characterized by low duration and low cumulative rainfall (e.g. 34.8 mm in 21 h on 31 October–1 November 2012 and 42.2 mm in 34 h on 6–7 October 2013). In contrast, the re-wetting of soil horizons deeper than 0.6–0.7 m below the surface and the weathered bedrock is not so fast, and only prolonged rainy periods with many rainfall events in a few days or weeks can provoke an increase in the pore water pressure and water content at this level (Fig. 6).

The rapid re-wetting as a consequence of early autumn rainfalls of the soil horizons within 0.6–0.7 m and the abrupt increase in pore water pressure in a time span of 5–10 h after the start of the rain during summer concentrated events, as it occurred on 27 June 2013 (13.3 mm in 2 h), on 26 August 2013 (16.5 mm in 3 h) and on 11 September 2013 (9.1 mm in 3 h), may also be due to the presence of desiccation cracks and other macro-voids all along the soil profile where rainwater may flow preferentially. This fact could promote a quick development towards near saturated conditions in the cracks and macro-voids (Bittelli et al., 2012; Smethurst et al., 2012).

In the winter and spring months, especially between December and May, frequent precipitation can increase the soil wetness until it approaches or reaches saturated conditions (Figs. 6 and 7). The soil water content ranges between 38 and 45 %. In contrast, the pore water pressure remains between -7 and -3 kPa in soil horizons within 0.6 m of the surface, while completely saturated conditions were reached in the G horizon at 1.2 m from ground as testified by the values of pore water pressure which remained quite steady at approximately 0 kPa (Fig. 7) until it reached positive values of approximately 1–3 kPa in response to more intense rainfall events (e.g. 29.8 mm in 24 h on 24–25 March 2013; 24.6 mm in 15 h on 30 March 2013; 29.5 mm in 26 h on 4–5 April 2013; 34.6 mm in 44 h on 18–20 January 2014; 68.9 mm in 42 h on 28 February–2 March 2014). Thus, in this situation, a moderately intense rainfall could not cause a further increase of pore water pressure and water content (Figs. 6 and 7).

During wet periods, the water content in the weathered bedrock at 1.4 m below the surface was lower than in the overlying G horizon (Fig. 6a).

According to the monitored data, it might be hypothesized that during winter and spring months a perched water table can form in the test-site slope soils due to the contact between the soil and weathered bedrock and remain steady at 1.2 m below the surface until the end of the spring. The thickness of this water table is approximately 0.1 m over the contact between the soil and the weathered bedrock. Additionally, when no rain falls for many days, as between 5 and 21 March 2014 (Fig. 7), the pore water pressure remains steady within a positive value range at this depth, thus confirming the presence of a perched water table. In contrast, following particularly intense rainfalls, water table can grow up to 0.8–1.0 m as testified by the significant increase in water content at 1.0 m from ground until conditions of complete saturation are attained (water content between 39 and 42 %).

This condition was not observed following other rainfall events in wetting periods during the monitored time span. In this situation, a shallow landslide affected the test-site slope (Fig. 7). For this reason, the triggering mechanism of rainfall-induced shallow landslides in the study area could be due to the emergence of a thin (0.1–0.2 m) perched water table present in winter and spring months in the slope soils of the study area following particularly intense rainfalls. Thus, it is fundamental to take into account this mechanism on modelling shallow landslides triggering zones through physically based models.

The particular hydrogeological setting of the bedrock levels does not allow for the formation of a deep groundwater level in the slope, but the water table develops only at the interface between soil and weathered bedrock. Other evidence demonstrates the absence of a deep groundwater level. Figure 8 shows the trends of water electrical conductivity measured through the TDR probes of the monitoring station at different depths in soil and weathered bedrock levels. The electrical conductivity values range between 15 and 40 µS cm-1 in the soil profile within 1.0 m from ground level, with the highest values following wet periods characterized by more frequent rainfalls. In the same periods, the water electrical conductivity reaches values of 50 µS cm-1 at 1.2 m from the ground level. The weathered bedrock level at a depth of 1.4 m has values in a similar range with respect to the shallower soil horizons (Fig. 8). According to these data, it seems clear that there is no uprising of a deep water table from the bedrock to the superficial soil, as demonstrated by water electrical conductivity values in the range of rainfall water. Moreover, the highest values measured at a depth of 1.2 m in the G horizon during the winter and spring months suggest the temporary formation of a perched water table at this level, which can arise from more intense rainfall events up to the shallower levels (Fig. 8).

Based on monitoring observations, the development of a thin perched water table above the soil–bedrock contact seems to be the most reasonable hydrological mechanism for shallow landslide triggers in this area, even if other mechanisms (e.g. bedrock exfiltration; Brönnimann et al., 2013) cannot be completely excluded. Moreover, the soil–atmosphere interaction phenomena observed during wet periods were considered a good benchmark for the application of the TRIGRS-unsaturated model. In fact, the model allows for the modelling of shallow landslide triggers due to the emergence of a perched water table (Baum et al., 2008), as we can observe in the monitored test site. This fact implies that the lowest computed safety factors correspond to areas characterized by the presence of a natural permeability barrier, such as the soil-weathered bedrock contact where shallow landslide sliding surfaces develop, as can be observed in the study area (Zizioli et al., 2013).

A DEM acquired before the April 2009 event with a grid size of 10 m × 10 m provides the topographic basis for the study area reported in Fig. 1. The TRIGRS-unsaturated model was also applied on the monitored slope with a more detailed DEM, with a grid size of 2 m × 2 m over an area of approximately 3290 m2.

TRIGRS-unsaturated was implemented with both DEMs to evaluate the differences in the models when passing from the site-specific scale and high resolution to a local scale with lower resolution.

To analyse the role played by the types of mapping units, the study area was divided into different regions according to each zoning. All input data were acquired from a GIS database in a “raster” form. For each mapping unit, a map was generated at the same spatial resolution as the DEM.

The soil's geotechnical and hydrological parameters are required as input data by TRIGRS-unsaturated and are summarized in Table 5. To take into account the uncertainties of the soil input data, the triggering zones of the shallow landslides were modelled by considering either the mean value or the value obtained by subtracting or adding the standard deviation of each parameter. A sensitivity analysis was performed through different simulations by changing only one parameter in each simulation and keeping the others constant. The considered variables were the unit weight γ, the peak friction angle φ′, the effective cohesion c′ and the hydrological properties required by TRIGRS-unsaturated. Due to the connections between all the hydrological properties of the soil, the sensitivity analysis was made to simultaneously change all these features. For the c′ parameter, the minimum value was set to 0 kPa. The soil hydraulic diffusivity D0 was estimated based on the characteristic Ks values of each class, in agreement with Baum et al. (2011) to be ∼ 2 × Ks (Table 5).

The parameter indicated as αG corresponds to the fitting parameter of Gardner's (1958) model for the SWCC. The parameter αG was estimated based on the α and n fitting parameters of Van Genuchten's model, as reconstructed for each unit through the method proposed by Ghezzehei et al. (2007) and expressed in Eq. (2): αG=α(1.3n). The slope angle and flow direction maps required by the TopoIndex (Topographic Index) utility were derived from the considered DEM using SAGA Gis (version 2.0.8; Conrad, 2006). By assuming that the gradient of the slope angle is rather uniform along each hillslope on the study area and almost regular – slightly convex in shape – a slope proportional weighting factor (exponent = 1 in TopoIndex) was chosen for the runoff distribution. This scheme distributes the flow excess to all adjacent downslope cells with a weighting factor proportional to the slope (Baum et al., 2008).

To create a continuous map of topsoil thickness, a geomorphologically indexed model based on the local slope angle, the elevation and the topographic position was used (Zizioli et al., 2013).

An analysis of the reliability of the pore water pressure at different depths as modelled by TRIGRS-unsaturated was performed. For the monitoring station and the most intense rainfall events in the wet periods of the monitored time span (23–25 March 2013, 30 March 2013, 4–5 April 2013, 20–22 April 2013, 18–20 January 2014 and 28 February–2 March 2014; Table 6), the measured values of the pore water pressure measured between 0.6 and 1.2 m from the ground level were compared to the modelled pore water pressure values using TRIGRS-unsaturated.

For these analyses, the initial water table depth across the study area was chosen according to the information obtained during the monitoring time span in the test-site slope. The most superficial measured pore water pressure values (within 0.6 m of the surface) were considered representative of the hydrological conditions in the selected area and were used to estimate the water table depth according to Eq. (3) (Comegna, 2008): d=ψgγwcos⁡2β, where ψg is the pore water pressure near the ground surface, d is the water table depth, β is the slope angle and γw is the unit weight of water.

Hourly values of the data related to the rainfall events from the data logger of the monitoring station were included in the model (Table 6).

The modelling of the pore water pressure for the monitored rainfalls in 2013–2014 was made considering the soil's geotechnical and hydrological data reported in Table 5. In particular, the geological class for the monitoring station area is the Rocca Ticozzi Conglomerates, and the pedological class is ILM1/RUM1.

Moreover, the hourly rainfall intensities recorded by the Cigognola rain gauge during the event that occurred on 27–28 April 2009 were assumed as rainfall input data (Table 6). For these analyses, it was assumed that a thin perched water table was already present in the extended study area before the beginning of the event based on what was observed at the monitoring station during the winters and springs. It was also assumed that this water table had an upper limit located at approximately 0.1 m above the soil–bedrock contact and a parallel trend with respect to this contact.

Figures 9 and 10 show the comparison between the measured and modelled pore water pressure trends at 0.6 and 1.2 m respectively for the considered rainfall events reported in Table 6. The modelled trends for the test-site soil are the same even when considering the two different mapping unit types because the soil's hydrological properties remain constant (Table 5). Moreover, the same estimated pore water pressure trends were found for the analyses with two different DEMs.

The graphs related to the pore pressure trends show how TRIGRS-unsaturated is able to adequately model the increase in pore pressure during a rainfall event.

In particular, differences greater than 2 kPa between the measured and estimated values have never been found at both depths, except for the final phase of the 28 February–2 March 2014 event. This result can be deemed a very positive result, especially considering that the tensiometers used in the monitoring are characterized by an accuracy equal to ±1.5 kPa.

The lowest RMSEs were found for shorter events and depths of 1.2 m compared to those at shallow depths (Table 8). Generally, the RMSEs values are lower than 1.5 kPa for all the considered events and are always lower at 1.2 m from the ground than at 0.6 m (Tables 7 and 8). The differences between the modelled trends at different depths can be linked to the different hydrological properties of the soil horizons. Overall, the increasing pressure trends agree with each other. The authors of the original code (Baum et al., 2002, 2008) noted that TRIGRS was developed to model intense weather events and not long periods with low levels of precipitation.

Note that a sudden increase in the pore water pressure during the first stages of the rainfall events that occurred on 23–25 March 2013, 30 March 2013 and 4–5 April 2013, as modelled by TRIGRS-unsaturated, is clearly visible at 1.2 m from the ground (Fig. 10a–c). In particular, the field measurements show a quite negative trend in the pore water pressure, while the model shows a slightly positive trend.

Furthermore, TRIGRS-unsaturated models the highest pore water pressure value at the end of each rainfall event, while in many cases (in particular for the event that occurred on 28 February–2 March 2014) the field measurements show that the highest pore water pressure values are not reached at the end of the event (Figs. 9f and 10f). It is important to note that the Fs modelled by TRIGRS-unsaturated for these rainfall events always remained over 1.0 (stable conditions). In fact, no shallow landslides were detected after these events next to the monitoring station.

The modelling errors for the pore water pressure trends at the monitoring station can be linked to the simplification provided by the model. The layers within 0.6 m of the surface in the test-site slope soil are somewhat different in terms of the hydrological features with respect to those of the levels between 0.6 and 1.2 m (Table 2). TRIGRS-unsaturated does not consider a layered soil; thus, it is appropriate to simulate the pore water pressure by assuming the mean values of the hydrologic parameters of different layers. Indeed, the mismatch is always lower than 2 kPa, although the mismatch attained a value of 2.8 kPa for the rainfall event on 28 February 2014. It is true that, in general, the variation could have the same range for the mismatch, and then the error could be high with respect to the variation, but we must at least consider the following aspects:

the reliability of these results should be evaluated not in an absolute sense, with respect to only isolated rainfall events, but in relation to long-period analyses;

Although a comparison with field measurements is not possible in this case due to the lack of field data, this result can be considered reliable, also accounting for the presence of a perched water table 1.2 m from the ground (positive pore water pressure; Fig. 11) at the beginning of the event, which is also the typical condition observed in the test-site slope during wet months.

Figure 12 shows the April 2009 scenario at the monitored test-site slope considering DEMs with grid sizes of 2 m × 2 m and of 10 m × 10 m. The figure refers to the period when shallow landslides triggered (between 32 and 48 h since the beginning of the rainfall); after this time span, the scenarios remain steady, following the modelled pore water pressure trend (Fig. 11).

The results were the same when mapping either a geological or pedological zone. In both cases, shallow landslide scarps fell in areas modelled by TRIGRS-unsaturated as unstable (Fig. 12). There is an overestimation of the unstable areas in both the reconstructions because the slope angle is quite steady along the hillslope (between 25 and 35∘). The satisfactory result at the site-specific scale, in terms of the comparison between the modelled and observed landslides, confirms the reliability of the applied methodology.

Figure 13 shows the shallow landslide scenarios for the April 2009 event for the whole study area by taking into account the mean values of the input soil features from the selected mapping unit types. To quantify the differences between the model results based on different mapping units, which also consider the sensitivity analysis, the TP and FP indexes were computed from the Fs output maps of the study area.

Considering the mean values of the input data, the scenarios based on either geological or pedological units did not provide a significant overestimation of the unstable areas: the maps have similar FP values, ranging between 10.1 and 10.2 % (Fig. 14). The differences between the obtained maps become evident when considering the TP values: the best prediction of the effective unstable areas was obtained considering a pedological mapping unit (78.9 %) with respect to the scenarios obtained through a geological mapping unit, which has a TP index of 73.3 % (Fig. 14).

By analysing the results of the sensitivity analyses, negligible variations in both the TP and FP are linked to the role played by the γ parameter (Fig. 14a), with values on the order of 0.9 and 2.3 % for geological and pedological mapping units respectively. Regarding the hydrological parameters, the best results in terms of the TP are obtained by considering the mean values, with an improvement on the order of 12.3–21.2 % with respect to the values, including the standard deviation. The FP values remain stable in all simulations (Fig. 14d).

Greater variations affect both the TP and FP indexes by changing φ′ and c′ (Fig. 14b and c). The lowest values of these parameters indicate an improvement in the correct identification of real unstable areas, as indicated by an increase in TP up to 7.3 %, but they also lead to an increase in the areas wrongly mapped as unstable, as indicated by an increase in the FP values up to 37.1 %.

In this case, a nil cohesion value for both pedological mapping units (Fig. 14b and c) has been assumed. The highest values of φ′ and c′ cause a slight decrease in FP, ranging between 0.2 and 5.7 %, and a significant decrease in TP, ranging between 40.4 and 44.3 % (Fig. 14b and c).

The results obtained through the TRIGRS-unsaturated model have also been compared to the results obtained by Zizioli et al. (2013) for the same area using the TRIGRS-saturated, SINMAP and SLIP models (Fig. 15). These three models were applied considering different initial hydrological conditions with respect to those considered in this work (Zizioli et al., 2013). Based on the same mapping unit (geological zoning) and considering the mean values of the input soil features, the TRIGRS-unsaturated model provided a better assessment in terms of the ratio between the TP and FP with respect to the other models. In fact, the SLIP and SINMAP models have similar TP (71.7–72.4 %) and FP (30.3–32.2 %), which indicates a significant overestimation of the unstable areas, while TRIGRS-saturated has a lower FP ranging between 10.0 and 10.2 %. The TRIGRS-unsaturated scenario models the effective unstable areas slightly better, as demonstrated by the highest TP value (73.3 %; Fig. 15).