Background & Summary

The study of landslides at detailed scales necessitates the application of distributed hydrological models coupled with slope stability models. This in turn requires knowledge of the hydraulic and various geotechnical properties of the materials involved. For large and diverse areas, the lack of accurate knowledge of these parameters necessitates the use of a statistical, data-driven approach. At these scales and with this approach, the landslide triggering conditions may be influenced by less punctual variables, such as mean annual precipitation (MAP), lithology or land use1,2.

Empirical rainfall thresholds are most widely used in the prediction of shallow landslides triggered by rainfall over large areas1,3. The approach identifies an empirical relationship between rainfall and landslides, derived from historical data on past failures and triggering rainfall conditions. Together with rainfall measurements, estimates and forecasts, empirical thresholds can be integrated into territorial early warning systems4,5 to effectively predict the occurrence of rainfall-induced shallow slope failures.

The development of robust and reliable rainfall thresholds for landslide prediction requires an investment in terms of effort and resources to collect detailed information on the spatial and temporal occurrence of landslides and to define the rainfall (measured by dense rain gauge networks or estimated by radar or satellites) responsible for the initiations of the failures6,7.

In this context, Peruccacci et al.8 published ITALICA, the ITAlian rainfall-induced LandslIdes Catalogue, the most comprehensive and accurate catalogue of landslides induced by rainfall that occurred within the Italian territory. In ITALICA, the majority of landslides had a direct or indirect impact on the local population. Landslides that occurred in areas that have not been significantly altered by humans, or for which there are no documented reports, are infrequently included in the catalogue. The landslide information comes from two broad categories: (i) news reporting, including newspapers, news sites, social media and blogs, and (ii) reporting from institutions. The utilisation of news reporting is typically challenging, as the resulting information is frequently incomplete and lacking in temporal and spatial accuracy. In the absence of a minimum of information, such as the broad location and, at least, the day of occurrence of the landslide, the event is excluded from the catalogue. Institutional authorities have been demonstrated to be a particularly useful source of information, providing reliable and accurate data regarding the precise or approximate location of landslides in space and time. ITALICA provides the geographic coordinates of individual landslides, which are represented as discrete points on a map (not polygons). Each landslide in the catalogue has been assigned a level of geographic accuracy P in five classes: P0 (very high); P1 < 1 km2 (high); 1 km2 ≤ P10 < 10 km2 (medium); 10 km2 ≤ P100 < 100 km2 (low); and 100 km2 ≤ P300 < 300 km2 (very low). Regarding the temporal localisation, three different classes of temporal accuracy are assigned to the landslides. These classes vary from T1 (knowledge of the exact time at which they occurred), to T2 (knowledge of the part of the day) to T3 (only the day is known). ITALICA is publicly available9 and is particularly relevant for the highly accurate spatial and temporal localisation of the landslides. Consequently, the catalogue is adequate for the purpose of reconstructing rainfall conditions that are capable of triggering future landslides on Italian territory.

It is essential to establish objective and reproducible criteria to determining the amount of rainfall that triggered the landslides listed in ITALICA. The lack of consensus on the criteria for identifying individual rainfall events (e.g.7,10) reduces the possibility of comparing different rainfall thresholds and limits their potential use in landslide warning systems. In this regard, Melillo et al.10,11 proposed an automated algorithm for the objective definition of individual rainfall events from rainfall series and for the quantitative measurement of the rainfall conditions responsible for failures. The algorithm was subsequently updated to include additional features in a software tool named CTRL–T (Calculation of Thresholds for Rainfall-induced Landslides–Tool12) available on Zenodo13 and since then it is widely used in the field of rainfall-induced landslide prediction (e.g.14,15,16,17,18).

Here, we introduce e-ITALICA19 — the enhanced version of ITALICA — which contains new fields related to the rainfall conditions responsible for landslide triggering, calculated with a rigorous and reproducible approach. The new fields include longitude and latitude of the representative rain gauge for each individual landslide, and the reconstructed cumulated event rainfall (E, in mm) and rainfall event duration (D, in hours). Such information is rarely available to the scientific community due to the difficulty of finding sufficiently long and complete rainfall time series. e-ITALICA can be used to analyse and assess, at different geographical scales, the rainfall conditions able to trigger landslides in the Italian territory. Additional applications of the new catalogue concern the calibration and validation of data-driven, machine-based, and deep-learning landslide forecasting tools, physically-based landslide prediction models, and local to national empirical rainfall thresholds. The intersection between thematic layers, such as the MAP or the land cover, and the rainfall conditions associated with landslides allows the study of possible relationships between the chosen thematic classification and the minimum amount of rainfall that can initiate landslides. The landslide type classification — also available in the catalogue — can be used for further analyses.

The results of this work are suitable for use by civil protection authorities at local or national level to forecast and monitor rainfall-induced landslides through the implementation of early warning systems.

Methods

Over the past 15 years, many researchers of our team have been involved in the compilation of ITALICA8,9, which lists 6312 records with spatially and temporally accurate information on rainfall-induced landslides that occurred on Italian territory between January 1996 and December 2021. For these landslides (Fig. 1) — now included in e-ITALICA19— we reconstruct the rainfall triggering conditions using an objective, reproducible and automated process. A workflow is shown in Fig. 2.

Fig. 1
figure 1

Spatial distribution of the landslides listed in e-ITALICA. Background from ©Microsoft; EPSG: 4326.

Full size image
Fig. 2
figure 2

Workflow diagram used to generate e-ITALICA.

Full size image

Information on rainfall-induced landslides was gathered from various sources including institutional and news reports. Overall, about three quarters of the mass movements were classified as generic shallow landslides. Unfortunately, data availability was not uniform across the Italian territory due to different data sharing policies of provincial and regional authorities. For each landslide, the information collected and stored in ITALICA included: (i) the landslide identification and the source of information, (ii) the type of the landslide, if available, (iii) the landslide coordinates and the geographical accuracy with which the landslide could be located, and (iv) the precise or approximate date of the failures and the temporal accuracy. What makes ITALICA unique was the high spatial and temporal accuracy with which the landslides are located. About half of the landslide records have a high geographic accuracy and more than 95% of the landslides have been located with an uncertainty of less than 10 km2. For almost half of the landslides in the catalogue, the exact time of occurrence is known.

ITALICA provided the scientific community with a model for building accurate spatio-temporal catalogues worldwide. The high accuracy of the catalogue’s data makes it particularly suitable for defining models and tools to predict rainfall-induced landslides (e.g. rainfall thresholds to be implemented in landslide early warning systems). To this end, we have prepared e-ITALICA19 by adding new fields related to the land cover, the topography and the rainfall conditions that caused the slope failures in Italy. e-ITALICA preserves the high spatial and temporal accuracy characteristics of ITALICA.

In regard to the land cover, three releases of the Corine (Coordination of Information on the Environment) Land Cover (CLC) maps are employed, corresponding to the years 2006, 2012, and 2018 (henceforth designated as CLC2006, CLC2012, and CLC2018, respectively). CLC is a vector-based dataset comprising 44 land cover/use classes, derived from a range of satellite missions, providing information at a pan-European scale with a minimum mapping unit of 25 hectares and updates every six years20. Each landslide in the catalogue is associated with a land cover class in accordance with its spatial and temporal characteristics. An overlay is performed between the landslide points and the CLC maps, according to the year in which the landslides occurred. In order to consider the land cover of the year closest to the year of occurrence, landslides occurring between 2002 and 2007 are assigned to CLC2006, those occurring between 2008 and 2013 to CLC2012, and those occurring between 2014 and 2020 to CLC2018. Landslide points with very high accuracy (P0) are overlapped with the CLC maps. Points related to landslides mapped with high accuracy (P1) are converted into circles with the corresponding P1 radius of 0.6 km. Thereafter, the landslides are intersected with the CLC class layers, and the proportion of each CLC class within each circle is calculated. A CLC class is attributed to each landslide if it covers 75% or more of the circle area1.

In order to incorporate the topographic characteristics of the landslides into e-ITALICA, the average elevation and slope of the landslide locations are obtained. To achieve this, the slope units21,22 that encompass the locations of the landslides with very high and high accuracy (P0 and P1) are selected, and the average values of elevation and slope are calculated. The slope units are terrain units delimited by drainage and dividing lines, and the data used to define them is obtained exclusively from a digital elevation model. In e-ITALICA, the land cover information is available for 3956 landslides and the average elevation and slope for 3933.

To reconstruct the rainfall responsible for the landslides in e-ITALICA, we have used rain gauges from the dense national network operated by the Italian Department of Civil Protection and local networks managed by regional authorities. Overall, the rainfall database currently consists of measurements from 4033 rain gauges (Fig. 3) distributed throughout the country with an average density of one station every 75 km2 (up to one gauge per 26 km2 in some regions). Data currently covers the period from January 1996 to December 2021. Over this 27-year period, the number of stations has increased over time, allowing the reconstruction of the rainfall that triggered landslides for a greater number of otherwise discarded records. Rainfall series have a variable temporal resolution, ranging from 5 to 60 minutes. Measurements are re-aggregated to one hour to obtain homogeneous data that still maintain an adequate resolution for the reconstruction of the rainfall responsible for the failures.

Fig. 3
figure 3

Spatial distribution of the rain gauges in Italy. Background from ©Microsoft; EPSG: 4326.

Full size image

We applied CTRL-T13 to the above-described landslide and rainfall data. CTRL-T is written in R (an open-source language) and, starting from standard rainfall series and landslide information, automatically reconstructs the rainfall events, and multiple rainfall conditions responsible for the initiation of known landslides, in terms of duration and cumulated rainfall, and defines frequentist rainfall thresholds for different non-exceedance probabilities. CTRL-T adopts a bootstrapping technique to assess the statistical uncertainty associated with the parameters defining the rainfall thresholds23. The robustness of the thresholds depends primarily on the number of empirical data. The large number of records in e-ITALICA makes it possible to limit the uncertainty associated with rainfall threshold parameters even for thematic subdivisions of the territory (e.g. topographic, climatic, lithological thresholds, etc., see e.g.1). As a first step, the tool uses time series of rainfall measurements obtained from the rain gauge network (Fig. 3) to determine the duration D (in h) and the cumulated rainfall E (in mm) of individual rainfall events. To separate individual rainfall events in rainfall records, CTRL-T13 uses a set of input parameters that vary with season and latitude. Regarding the season, two periods are considered: a dry, spring–summer period, and a wet, autumn–winter period. In this work, a minimum dry interval of 48 hours and 96 hours between two consecutive rainy periods is used to separate rainfall events in the dry and wet period, respectively. A different duration of the dry and wet periods (Table 1) is established for three different areas of Italy (North, Centre and South), considering the climatic variation at the different latitudes and according to previous works (e.g.1,24,25). Figure 4 shows the spatial distribution of the landslides classified by the latitude criterion. The monthly distribution of rainfall induced landslides (inset of Fig. 4) is evidence of the different rainfall regime in the three areas.

Table 1 Dry and wet periods defined in the three areas.
Full size table
Fig. 4
figure 4

Spatial classification of landslides as a function of latitude for triggering rainfall reconstruction. Landslides are classified in the three areas: North, Centre and South. Background from ©Microsoft; EPSG: 4326.

Full size image

For each landslide in the catalogue, CTRL-T13 selects nearby rain gauges located within a circular area (buffer) centred on the landslide site and with a variable radius, depending on the morphological characteristics of the study area and the local density of rain gauges. A maximum buffer radius of 20 km is considered for this work.

For each of the rain gauges in the buffer, CTRL-T13 defines single or multiple rainfall triggering conditions (D,E) and assigns a weight to each of them. The weight is a function of the rainfall characteristics and the distance of the rain gauge from the landslide. The rainfall condition with the highest weight value (MPRC, Maximum Probability Rainfall Condition) leads to the identification of the representative rain gauge and is assumed to be the most probable for landslide initiation12. In the majority of cases (90%), the rain gauge employed is situated at a distance of less than 12.5 km from the location of the landslide.

After reconstructing the rainfall triggering conditions (D,E), CTRL-T13 uses the frequentist method23,26 to objectively and reproducibly define empirical rainfall thresholds. The method calculates ED thresholds for different non-exceedance probabilities assuming power law curves. As an example, the threshold at 5% non-exceedance probability leaves 5% of the empirical (D,E) data points below the curve.

Ultimately, the use of the algorithm, working in an automatic, objective and reproducible manner, reduces the uncertainty in the various aspects of rainfall-induced landslide prediction, from the characterization of the rainfall event responsible for the landslide using consistent decision rules, to the automatic selection of the most representative rain gauge for a given landslide from a pool of rain gauges, to the definition of reliable probabilistic rainfall thresholds. All this in a drastically reduced time compared to a manual procedure, which is also subjective.

Data Records

e-ITALICA is stored in a repository and is publicly available for download at https://doi.org/10.5281/zenodo.1420447319. The catalogue is available in Comma-Separated Values format (csv), as well as in Keyhole Markup Language (kml) and GeoPackage (gpkg) formats for a prompt visualization in a GIS environment.

The fields included in e-ITALICA are:

  • ID: unique ID for each reported rainfall-induced landslide;

  • information_source: source of record information, including news reporting (NR) and institutions reporting (IR);

  • landslide_type: type of landslides, reported if available from the source: debris flow (DF), earth flow (EF), mud flow (MF), rock fall (RF), generic shallow landslide (SL);

  • lon: longitude (EPSG: 4326) of the reported failures;

  • lat: latitude (EPSG: 4326) of the reported failures;

  • municipality: municipality in which the landslide occurred;

  • province: province in which the landslide occurred;

  • region: region in which the landslide occurred;

  • geographic_accuracy: this field assigns a qualitative level for the landslide geographic accuracy based on the area over which the landslide occurred, described as a radius from the coordinates of the failures (in kilometres), in five classes:

    • very high, P0 exact landslide location;

    • high, P1 < 1 km2;

    • medium, 1 ≤ P10 < 10 km2;

    • low, 10 ≤ P100 < 100 km2;

    • very low, 100 ≤ P300 < 300 km2;

  • land_cover: land cover described by the three-level code based on Corine Land Cover legend20;

  • elevation: average elevation (m a.s.l.) of the slope unit containing the landslide location point;

  • slope: average slope (decimal degrees) of the slope unit containing the landslide location point;

  • day: reported day of the landslide;

  • month: reported month of the landslide;

  • year: reported year of the landslide;

  • local_time: reported hour and minute of the landslide, recorded as HH:MM (24 clock, local time);

  • local_date: this field summarises the day, month, year, and local time of the reported landslides.

  • UTC_date: this field summarises the date and the UTC time of the reported landslides;

  • temporal_accuracy: this field assigns a qualitative level for the landslide temporal accuracy, in three classes:

    • T1 when the time (minute to hour) of the failure is known;

    • T2 when the part of the day (a period of 3–5 h) is known (i.e., early or late morning, midday, early or late afternoon, middle of the night); for details see Peruccacci et al. (2023);

    • T3 when only the day of occurrence is known;

  • lon_raingauge: longitude (EPSG: 4326) of the representative rain gauges;

  • lat_raingauge: latitude (EPSG: 4326) of the representative rain gauges;

  • duration: duration (h) of the rainfall condition responsible for the slope failure;

  • cumulated_rainfall: cumulated event rainfall (mm) of the rainfall condition responsible for the slope failure;

For half of the records in the catalogue, the representative rain gauge was identified within 4.5 km, while for 90% of the records the distance was less than 12.5 km. In several cases where many landslides occurred in a relatively small area and on the same day, the same triggering rainfall condition is associated with all the failures. This is because a single rain gauge is representative of all the landslides.

The landslide-triggering ED conditions are reconstructed for 5974 landslides. Figure 5a shows the distribution of the (D,E) conditions that caused one or more slope failures in northern (2556), central (2055) and southern (1363) Italy represented in logarithmic coordinates. In the catalogue, rainfall conditions that resulted in landslides in Italy are in the wide range of duration 1 h ≤ D ≤ 1615 h. Nevertheless, in the three areas the median and the third quartile of the reconstructed rainfall durations do not exceed 44 h and 106 h, respectively (Fig. 5b). The cumulated rainfall is in the range 5.6 mm ≤ E ≤ 1779.6 mm with median values decreasing from North (132.4 mm) to South (76.2 mm) (Fig. 5c).

Fig. 5
figure 5

Landslide-triggering rainfall conditions in Italy. (a) Rainfall duration D (h) vs. cumulated event rainfall E (mm) conditions that have resulted in landslides in northern, central and southern Italy. Data shown in log-log coordinates. (b,c) Box plots of the duration and the cumulated rainfall of the rainfall conditions that caused landslides in the three areas (North, Centre and South).

Full size image

The triggering conditions for landslides are calculated using hourly rainfall measurements. Consequently, if the duration of a triggering rainfall lasts less than one hour (short, intense rainfall mainly from convective events) it is approximated to one hour, which may lead to a possible underestimation of the mean rainfall intensity (E/D).

Regarding the temporal localisation, the landslides are classified according to the three distinct classes of temporal accuracy, T1 (the exact occurrence time is known), T2 (the part of the day is known) and T3 (only the day is known). The temporal accuracies T2 and T3 give rise to triggering conditions (D,E), which may be affected by a certain uncertainty. For example, in the case of a landslide that occurred during a full day of rainfall, but of which only the day is known from the information source (T3), the associated duration and cumulated rainfall may be overestimated. Nevertheless, almost half of the landslides included in the catalogue have a precise time of occurrence (T1).

With regard to the spatial location, for landslides in classes with low geographic accuracy (P100 and P300), the assignment of the selected rain gauge may not be the most representative for the failure. However, it is notable that about half of the landslides have a high geographic accuracy (P1) and more than 95% of the landslides were located with an uncertainty of less than 10 km2.

Technical Validation

The technical quality and reliability of the records included in e-ITALICA19 are ensured by the accuracy of data collection, the objective and reproducible approach used to reconstruct the triggering rainfall conditions, and the numerous checks carried out during the compilation of the catalogue.

It is worth noting that for all the landslides included in e-ITALICA the source of information indicates that their trigger is rainfall. In particular, landslides are excluded if the triggering factors are unknown or different from rainfall, and if there is evidence of other triggers acting in conjunction with rainfall (freeze-thaw cycles, snow melt, seismic vibration, human influence, etc.).

A first check concerns the landslide data of ITALICA (see workflow diagram in Fig. 1). Landslide information is collected and entered into the catalogue by several operators. In order to limit subjectivity in the compilation of the catalogue, several workshop and training courses were organized to ensure the adoption of uniform criteria by all the operators. A validation of the landslides added in the catalogue is carried out by assigning a random sample of the records from one operator to another one. In most cases the records are filled in with the same details. In case of discrepancies, they are double-checked by the team.

Moreover, a double-check on the definition of landslide location and timing is conducted, in order to avoid discrepancies between the information gathered from the sources and the details listed in the catalogue. Through the application of combined filters (e.g. landslide location and date of occurrence), duplicate records are eliminated.

A subsequent verification and validation of the rainfall measurements is performed. Several rainfall series have single or multiple measurement gaps due to technical and operational problems. CTRL-T13 is used to automatically check the continuity of the records and to identify gaps or errors (negative or anomalous values). In order to reduce uncertainties, we decide not to fill the gaps with interpolation procedures but to label them as ‘NA’ (meaning ‘not available’).

The landslide-triggering ED conditions automatically reconstructed by CTRL-T are checked and validated to eliminate rainfall conditions that are definitely not capable of triggering slope failures (rainfall less than 1 mm/h).

A further check of the coordinates of landslides and representative rain gauges, with calculation of relative distances, allows possible errors in the location of a few landslides to be identified and corrected.

The validation chain results in the reconstruction of ED conditions for 5974 landslides out of 6312 records previously listed in ITALICA. The remaining 338 slope failures are not associated with rainfall due to a lack of rainfall data or a time lag of more than 48 hours between the end of the rainfall and the occurrence of the landslide. The lack of rainfall data can be explained by: (i) absence of rainfall measured by rain gauges, possibly due to a very localized (convective) storm that is difficult to detect even by a dense rain gauge network; (ii) the lack of rain gauges within the buffer, which is typical of areas where the density of the rain gauge network is particularly low, (iii) malfunctioning rain gauges, especially in the case of very severe thunderstorms, and/or (iv) unreliable rainfall conditions. In these cases, the fields “Rain gauge longitude”, “Rain gauge latitude”, “Rainfall duration”, and “Cumulated rainfall” are marked with “NA”.

Finally, to complement and support the technical validation, we replicated the analysis performed in Peruccacci et al.1 on the definition of frequentist rainfall thresholds for different mean annual precipitation (MAP, in mm) classes (T5,VL, T5,LO, T5,ME and T5,HI in Fig. 6). The results obtained are in line with those previously found, i.e. the higher the MAP value, the higher and steeper the threshold, confirming the validity of the data included in e-ITALICA.

Fig. 6
figure 6

Rainfall thresholds for different MAP classes. (a) Subdivision of Italy into five MAP classes as a proxy for the regional rainfall regime, modified after Desiato et al.27. Background from ©Microsoft; EPSG: 4326. Legend: VL, very low (MAP ≤ 800 mm); LO, low (800 mm < MAP ≤ 1200 mm); ME, medium (1200 mm < MAP ≤ 1600 mm); HI, high (1600 mm < MAP ≤ 2000. mm); VH, very high (MAP > 2000 mm). (b) Frequentist ED rainfall thresholds for the four classes: T5,VL (light blue curve), T5,LO (blue curve), T5,ME (purple curve) and, T5,HI (magenta curve), in log–log coordinates. It should be noted that for this subdivision, the relative uncertainty in the threshold parameters is below 10%, and is visually represented by the shaded area around the curves.

Full size image