Increases in Italian Landslides

2.1 Study Area

The analysis was performed considering all the Italian territory. Italy is located in Southern Europe with a total area of 301,230 km². It is crossed by two mountainous range, the Apennines to the south and the Alps to the North, and by the large Po plain and it comprises two main islands, Sicily and Sardinia. A total of 68% of the Italian municipalities is exposed to high levels of hydrological and geological hazards, which are often caused by intense rainfall events, causing severe damage (Messeri et al., 2016b). From the geodynamic point of view, Italy is in fact an extremely active region, with frequent earthquakes and active volcanoes (Bosellini, 2017). Italy encompasses a broad range of climatic regimes: 14 of the 35 climatic regions occurring in Europe are there present. Alps and Northern Apennines are dominated by temperate climates while Southern Apennines have a so-called Mediterranean mountainous climate. Po Plain and the adjacent low hills are characterized by intermediate climates, that is, Mediterranean suboceanic to subcontinental. The former is widespread also in central Italy and it extend toward the South of Italy inlands leaving place to more characterized Mediterranean climates, reaching also Mediterranean to subtropical climates, either partly semiarid or influenced by mountains (Costantini et al., 2013).

2.2 Meteorological Data

Precipitation series over Italy was obtained from the reanalysis product, ERA5-Land (Muñoz Sabater, 2019). The data is a replay with a finer spatial resolution of the land component of the ERA5 climate reanalysis (Hersbach et al., 2020). The data set has a spatial resolution of 0.1° × 0.1° and a temporal resolution of one hour (resampled to one day for the present purposes) with a temporal coverage that spans the period from 1950 to present. In order to have an idea of the performance of the data set, the spatial distribution of temporal clustering of precipitation was compared with the one obtained using E-OBS (Cornes et al., 2018). The latter is a daily gridded observational data set covering Europe, with a 9 km spatial resolution. It is based on the blended time series of the stations collected by the European Climate Assessment and Data set (ECA&D) initiative. Finally, maximum and minimum daily temperature from E-OBS were used to explain the occurrence of some landslide phenomena. The investigated period goes from 1950-12-01 up to 2020-11-30.

Finally, the series of North Atlantic Oscillation (NAO) Index and of Mediterranean Oscillation Index (MOI) (Conte et al., 1989; Palutikof, 2003; Palutikof et al., 1996) were obtained from the NOAA CLimate Prediction Center (https://www.cpc.ncep.noaa.gov/products/precip/CWlink/pna/nao.shtml) and the Climatic Research Unit, University of East Anglia (https://crudata.uea.ac.uk/cru/data/moi/), respectively. The MOI index was computed as the normalized pressure difference between Algiers and Cairo.

2.3 Landslides Data

Landslide events over Italy were obtained from the Aree Vulnerate Italiane (AVI) database, an inventory of landslides and floods occurred in Italy until 2001 by the National Group for Prevention of Hydrological Hazards (GNDCI) of the National Research Council (CNR) (http://avi.gndci.cnr.it). It is a point data set in which landslides and related characteristics were identified from newspaper articles (Guzzetti et al., 1994).

From the database, we extracted only the landslides occurred in the period 1950–2001. We disregarded all the events with missing information about date of occurrence, type of movement, or location and with a regional or provincial spatial accuracy. For the events with multiple locations associated, the average of the coordinates was computed. This may occur since the positions of landslides were identified looking at locations' names reported in the news, therefore multiple locations may be present. In addition, when a road is identified as location, the average between the end points of the road is used. Considering the selected events, we only investigated landslides that were triggered by precipitation. We also included events with unknown trigger. Since rainfall is the main driver of landslides, we assumed that when it was missing this was the trigger. Large precipitation systems may cause multiple landslides in connected locations. The data set used reports each of them as a separate record. Since the main purpose of the analysis is the investigation of landslides triggers, keeping all of them may bias the results, adding redundant information and resulting in a biased predominance of a trigger, We therefore grouped event together when they were occurring in the same or adjacent days and closer than 55 km, considering size of small to medium precipitation systems (Zhang & Wang, 2021). For each group we than retained only one record for landslide type, chosen as the most central event in space. In this way, we retained 895 events (99 flow events, 64 debris flow, 43 complex events, 562 fall events, 127 sliding events).

The AVI inventory, despite the remarkable effort beyond its construction and the amount of data that contains, suffers of some limitations due to the available technology at the time of its collection. The spatial distribution may be biased by the availability of local newspaper reports, and some area may be more covered than others. Collecting landslides appearing in the news means that only events that attracted public attention are present, that is, events that likely resulted in some kind of damages or losses. This however does not imply that only large landslides are reported since also minor ones can cause damages. If we assume that there is not a significant difference in the mechanisms triggering landslides occurring far from the human infrastructures and landslides hitting human infrastructures (once we excluded the ones triggered by human activities) this limitation should not greatly influence the results. Also, the yearly number of landslides is influenced by an improvement in the methodologies with which events after 1990 were collected, that results in a higher number of identified events. In the present analysis therefore no considerations about the evolution of the number of landslides or the most affected areas are carried out.

2.4 Methods

2.4.1 Temporal Clustering of Precipitation

The identification of temporal clustering of precipitation follows the methodology proposed by Bevacqua et al. (2021), with the modifications of Banfi and De Michele (2022). The idea is to calculate the number of precipitation events within a specified time window and determine whether this count is the result of a Bernoulli process. If not, in this latter case, we infer the presence of temporal clustering of precipitation (Figure 1).

To apply the method correctly, a series of distinct precipitation events above a given threshold is needed. This was obtained removing the high frequency clustering with a run decluster procedure (Coles, 2001). High frequency clustering can be seen as the dependence of precipitation exceedances inside a single meteorological event, while low frequency clustering (the one we are interested in) is related to multiple subsequent precipitation events. The procedure is as follows: (a) thresholding the precipitation series, (b) clustering together events closer than r days (here r = 2 days following Barton et al. (2016)), and (c) retaining only the first exceedance in each cluster and setting to NA all other ones. From the declustered series, the probability of exceedance p was computed, disregarding the days in which precipitation events were removed, that is counting the exceedances in the series and dividing it for the total length of the series minus the days with a NA. Here, we chose a threshold equal to the 0.7 quantile of daily precipitation, considering only wet days, following Bevacqua et al. (2021). This corresponds to values between 1.16 up to 10.54 mm, with an average of 3.5 mm.

To check for the presence of temporal clustering, we selected a time window w, and we counted the number of exceedances inside w, called n. In the absence of temporal clustering, events should be independently distributed inside the window. We performed therefore a statistical test with the null hypothesis that there is no clustering, that is, the number of events inside the window is distributed like a Binomial distribution, with parameters p and w_eff. Here, w_eff is an effective window equal to w minus the days in which precipitation was removed with high frequency declustering. The test is a one side test, where the hypothesis is rejected if n is higher than what expected from a Binomial distribution. In this work, we considered a 0.05 significant level.

We checked the presence of temporal clustering in each day of the time series considering three different time windows, centered on that day: 15, 30, and 90 days. The presence of temporal clustering was tested on each cell over the Italian territory on each day, therefore a multiple testing correction was needed to keep the overall significance at 0.05. In addition, the discreteness of the p-values needed to be considered as well. Regarding the latter, we computed mid-p-values as suggested by Heller and Gur (2011). Concerning the former, we disregarded the choice of using the classical methodologies proposed in literature. Most of them are designed for continuous variables and independent tests, like the well-known Benjamini–Hochberg (Benjamini & Hochberg, 1995), and they may lose power for an increasing number of tests, like the Bonferroni correction (Armstrong, 2014). Here, not only p-values were discrete and the number of tests was large, but tests were also spatially dependent. The spatial dependence of the tests implies that several contiguous p-values relatively high in the basin are a stronger evidence of the presence of temporal clustering than few sparse very low p-values, since the probability of finding significant p-values, which are spatially contiguous, only by chance, is in fact very low. A similar consideration was presented also by García (2004) in ecological studies. Moved by this, we proceeded by considering not significant all the p-values that were lower than 0.05, but that were not adjacent (including the diagonal cells) to at least three other cells with p-values lower than 0.05.

3.1 Spatio-Temporal Distribution of Temporal Clustering of Precipitation

Figure 2 shows the spatial distribution of the seasonal number of days with temporal clustering of precipitation over a 30 days window and the seasonal precipitation amount over Italy. The area more prone to temporal clustering of precipitation is the Alpine area, mainly the eastern part, during the summer months. Temporal clustering is important also in the western coast and south of Italy during the winter months. During Autumn and Spring, the spatial distribution is more even over the territory. Compared with the total precipitation, we observe similar patterns but also some differences. For example, the western part of Piedmont, that is characterized by the highest values of total precipitation from Spring to Autumn, does not emerge when we look at the maps of temporal clustering. Also during winter, the spatial distribution has some differences, with a high number of days with temporal clustering in Sardinia and Sicily, that is not matched in the total precipitation maps. During summer, the meteorological conditions are on average more stable than in autumn and spring due to the persistence of the Azores High pressure over Italy, thus resulting in less precipitation, mainly related to convective events.

The comparison of ERA5-Land with E-OBS data set (Figure 2) shows a lower frequency of temporal clustering in the latter, in all seasons except from autumn. ERA5 data set, from which ERA5-Land is derived, is known to overestimates mean precipitation systematically in most of the domain and periods of the year, due to overestimation of wet days, with a stronger discrepancy in high mountain catchments in the convective summer period (Bandhauer et al., 2022). They also observed an underestimation of precipitation peaks in ERA5 in October and November in the Tagliamento catchment (north of Italy). These results may explain part of the difference observed in this study between the two data sets. Qualitatively, however, ERA5 reproduces the precipitation patterns well (Bandhauer et al., 2022). The performances of E-OBS are, on the other hands, much more dependent on the area considered, due to the varying spatial densities of point stations, with worse performances in areas with few meteorological stations, like the Alpine ones. Turco et al. (2013) compared E-OBS with other gridded data set over the Great Alpine Region and a subregion in northwest Italy (NWI). They concluded that E-OBS does not reproduce reliably the climatology over NWI and that the use of E-OBS in these regions should be done with caution. This brought us to prefer the use of ERA5-Land in the study analysis.

The composites of seasons with above and below average MOI show a marked difference in winter in South-Central Italy, thus suggesting a connection between MOI and temporal clustering of precipitation (Figure 3). A similar pattern was observed also using the NAO index (Figure 3). The two indexes are correlated due to the common influence of the Northeast Atlantic low systems forcing Mediterranean cyclogenesis. The MOI can be seen as a sea level pressure anomalies oscillation in the Western-Central Mediterranean. It correlates with different climatic variables, like evaporation, precipitation, and heat flux. Its negative phase is associated with a dipole of low see level pressure anomalies between Central Europe and Turkey, resulting in the movement of continental cold and dry air masses to the Mediterranean, with an increase in evaporation. During its positive phase, the dipole is located between North Africa and Central Europe, with a movement of warm and moist air masses to Central and Western Mediterranean and a decrease in evaporation (Criado-Aldeanueva & Soto-Navarro, 2020).

The association between precipitation and global scale oscillation indices in Italy was observed by other authors (Brunetti et al., 2002; Caloiero et al., 2011). For example, Caloiero et al. (2011) found a strong correlation between teleconnection patterns and precipitation in Southern Italy, that was particularly evident on the west side and in winter. From this work emerged that similar conclusions can be drawn also regarding the temporal compoundness in addition to the seasonal amount.

Moving to the synoptic conditions, we can observe that the frequency of temporal clustering associated with the different WTs is variable depending on the region and season (Figure 4). WT8 is in general the weather type associated with the highest frequency of temporal clustering of precipitation, in all seasons. This WT is characterized by a cyclonic circulation over west Europe and a ridge over the eastern Mediterranean and it causes abundant precipitation over Northern Italy. Already Messeri et al. (2016b) found that this synoptic condition is the one associated with the highest landslide and flood risk in Italy. Also WT4 is a cyclonic circulation over northern Italy, despite being associated with stable conditions over central and southern Italy due to the persistence of a subtropical high pressure. In fact, we can observe two different anomaly signs moving from south to north of Italy for this WT. An important weather type for temporal clustering of precipitation in South of Italy, mainly in winter, is WT3. This WT is characterized by a cyclonic circulation over Iceland and an anticyclonic one over northern central Europe. WT2 is instead characterized by a partial displacement of the Azores High Pressure to the Northern Atlantic Ocean that lets the maritime polar air masses to reach Central Europe and to some extent the Mediterranean area. This WT is associated with higher frequencies of temporal clustering of precipitation over central Italy and lower over northern Italy. WT5 and WT7 are both associated with anticyclonic conditions and this explains the low occurrence of temporal clustering of precipitation observed during them.

3.2 Precipitation Events Triggering Landslides

The spatial distribution of landslide events is reported in Figure 5 for each season and type of movement. Among the different classes of landslides, fall events show less evident seasonal or spatial patterns. On the contrary, flow, sliding and debris flow during summer occurred mostly in the alpine areas while during winter they were more frequent in central or southern Italy, mimicking therefore the precipitation pattern (Figure 2).

Temporal clustering of precipitation was a significant triggers for all the landslide types with percentage, excluding rock falls, of around 50% (Figures 6 and 7). The results about the characteristics of precipitation events preceding each landslide type give us also the possibility to distinguish different generating mechanisms (Figure 6). For debris flows, the temporal clustering over small windows explains a good amount of events (39% over a 15 days window). This is in line with the work of Bevacqua et al. (2021) that observed temporal clustering of rainfall over small windows before shallow movements. A high influence of temporal clustering of precipitation over large windows was found for complex and sliding events (30%–39% over a 90 days window, respectively). However, it is interesting to point out that for complex movements the presence of temporal clustering over small windows was very low compared with the others. In contrast to the other types, fall events are not predominantly associated with none of the two triggers. Looking at precipitation totals, we can observe similar patterns between them and temporal clustering for debris flow and flow. However they are fairly different for complex events. In fact we observed high precipitation totals for short duration but a very low presence of temporal clustering of precipitation, suggesting that the obtained totals are due to few intense events. A similar behavior can be observed also for slidings.

From both Figures 5 and 6 it is evident that rock falls are much less linked with precipitation events, either isolated and intense, or clustered, than the other types. In South Central Italy, in summer, they are almost the only typology observed. Different authors identified an association between the temporal distribution of rock falls and freeze–thaw cycles (Bajni et al., 2021; Nissen et al., 2022; Pratt et al., 2019), that may therefore explain winter events. The presence of a high number of events in summer, not related to precipitation, suggests that also high temperatures may play a role, for example, causing deformation of the materials, thus favoring rock fall processes in summer. In Figure 8a we report the distribution of maximum temperature (a) during summer days with rock fall not associated with precipitation and (b) during the same calendar days but for the other years. This to compare meteorological conditions driving or not rock fall. What appears is that rock fall occurrence is associated with higher maximum daily temperature with respect to normal days. The same can be observed for winter rock falls and minimum temperature (Figure 8b), with a left shift of the distribution in case of rock fall occurrence. In addition, we looked for the presence of freeze–thaw cycles in the 2 weeks preceding a rock fall in winter, or preceding the same calendar days but for the other years (Figure 8c). A clear difference in the frequency is visible, with the presence of one or more freeze–thaw in almost 80% of the weeks before a rock fall.

The occurrence of landslides is often a non isolated phenomenon, since a precipitation event may trigger movements in multiple locations. Here, we clustered together landslides close in time and space and we considered only one event for each cluster to investigate the drivers, to avoid biases due to the dependence between events (Figure 9). However, the information on cluster size was also explored in relation with temporal clustering of precipitation. The largest events occurred in the Campania region and in the Alpine area and they were driven by a temporal clustering of precipitation.