Technical Note

Landslides (2023) 20:2725–2739 Pierpaolo Distefano · David J. Peres · Luca Piciullo · Nunziarita Palazzolo ·
DOI 10.1007/s10346-023-02132-5
Received: 11 April 2023 Pietro Scandura · Antonino Cancelliere
Accepted: 11 August 2023
Published online: 6 September 2023
© Springer-Verlag GmbH Germany,
Hydro‑meteorological landslide triggering
part of Springer Nature 2023 thresholds based on artificial neural networks
using observed precipitation and ERA5‑Land
soil moisture

Abstract Landslide prediction is key for the development of early to landslide occurrences in the past (Aleotti 2004; Guzzetti et al.
warning systems. In this work, we develop artificial neural networks 2007; Peruccacci et al. 2012, 2017; Gariano et al. 2015; Piciullo et al.
(ANNs) that can identify landslide triggering conditions using soil 2017; Segoni et al. 2018). Duration (D), cumulative rainfall (E), and
moisture data in addition to precipitation. In particular, we use rainfall intensity (I = E/D) have been the variables typically used
observed precipitation and ERA5-Land reanalysis soil moisture for the definition of the thresholds.
data at four different depth layers at the beginning and end of the In recent years, there has been a growing interest in using a
precipitation events. Two different case studies, Sicily region (Italy), hydro-meteorological approach for deriving landslide triggering
and a group of catchments in the Bergen area (Norway), are used thresholds (Bogaard and Greco 2018). According to this approach,
to test the proposed approach against different climatic and geo- the x-axis of thresholds (traditionally presenting rainfall duration)
morphological conditions. As a first step, traditional power law contains a “hydrological” variable such as antecedent rainfall, soil
thresholds based on cumulative precipitation and duration (E-D) moisture, groundwater level, or catchment storage deemed to rep-
are derived by maximizing the true skill statistic (TSS) as a bench- resent some of the “causes” predisposing slopes to failure. On the
mark. For both study areas, ANNs using 87 different input com- other side, on the y-axis, the magnitude of the rainfall event, typi-
binations of precipitation characteristics and soil moisture data cally measured by either its intensity (I) or cumulative precipitation
at multiple depth layers are analyzed. The developed ANN classi- (E), is reported representing the “triggering” factor. This method
fiers using soil moisture information in addition to precipitation has gained popularity in the literature, with studies by Mirus et al.
outperform those using precipitation data only. Specifically, while (2018a, b), Conrad et al. (2021), and Palazzolo et al. (2023) support-
power law E-D thresholds lead to a TSS maximum of 0.50 for both ing its advantages.
areas, the use of single-layer soil moisture yields a maximum TSS Attempts have also been made to use more than two variables.
of 0.76 (0.78) for Sicily (Bergen area), while the use of multilayer Rosi et al. (2021) proposed a 3D threshold, using rainfall intensity,
soil moisture taken at both the start and the end of precipitation duration, and mean areal rainfall (MeAR). The introduction of the
events yields a TSS = 0.79 (0.89). These results demonstrate that third variable led to an improvement in the prediction of triggering
the proposed methodology is particularly promising for improving conditions with a reduction of the rate of false alarms. Within this
landslide prediction. context, in a previous study by the authors (Distefano et al. 2022),
artificial neural networks (ANNs) identifying landslide trigger-
Keywords Landslide early warning · Geohazards · Machine ing conditions have been introduced, and a 3D threshold consid-
learning · Sicily · Norway ering rainfall duration, depth, and maximum event intensity was
found to have higher performances that the one based just on the
Introduction first two variables. ANN thresholds based on precipitation have a
The report of the United Nations Office for Disaster Risk Reduction significantly better predictive performance than E-D power law
(UNDRR 2020) highlighted that the frequency of natural disasters thresholds, even when using just the same rainfall variables. Other
observed worldwide in the period 2000–2019 increased compared studies have investigated the use of ANNs and other machine learn-
to the period 1980–1999, together with the number of casualties, ing algorithms for dynamic landslide prediction at both local and
people affected, and funds allocated to deal with the emergency regional scales, demonstrating their advantages (Collini et al. 2022;
conditions. Landslides triggered by rainfall events must certainly Pota et al. 2022; Steger et al. 2022).
be acknowledged among the most impacting geohazards (Froude In this work, we extend our mentioned previous study, by inves-
and Petley 2018; Haque et al. 2019). tigating the possibility of developing landslide triggering classifi-
Landslide early warning, which is one of the most important ers inspired to the hydro-meteorological approach, i.e., taking into
risk mitigation measures, has seen improvements throughout the account both precipitation and hydrological information (specifi-
last decades, improving from the first attempts to develop empirical cally soil moisture data). In particular, we investigate whether the
landslide triggering thresholds in the 1970s (Campbell 1975; Caine use of soil moisture data (in addition to rainfall data) at different
1980). Until a few years ago, thresholds have mostly been statisti- depth layers can improve prediction performance. Furthermore,
cally determined by examining the rainfall characteristics that led in addition to the majority of the abovementioned studies on

Landslides 20 • (2023) 2725

 2726 Technical Note
hydro-meteorological thresholds, here we also investigate the pos- positives” (FP, NTE classified as TE), and “false negatives” (FN, TE
sible use of soil moisture either at the beginning or at the end of classified as NTE). As in many classification problems (Fawcett
rainfall events. 2006), the confusion matrix has been here used to evaluate ANNs
Our research examines two study regions with significantly dif- predictive capability. In particular, the TSS (Eq. 4) is given by the
ferent climatic and geomorphological characteristics: the island difference of the true positive rate (TPR) (Eq. 2) and the false nega-
of Sicily (Italy), and a group of watersheds near Bergen (Norway). tive rate (FPR) (Eq. 3):
TP
Methodology TPR = (2)
TP + FN
Overview
FP
The proposed methodology involves three main steps (Fig. 1): data FPR = (3)
TN + FP
collection and preprocessing, data analysis, and performance evalu-
ation. The first step involves the acquisition of rainfall and landslide
data. The preprocessing procedure aims at removing incorrect data TSS = TPR − FPR (4)
and double entries. The TSS is equal to 1 in case of a prefect prediction, while it is
The second step is to create a database of triggering (TE) and equal to zero, when the prediction is completely random. ANNs
non-triggering (NTE) events. Here, the CTRL-T software, proposed designed for pattern recognition have been then applied to derive
by Melillo et al. (2018), was used for the purpose. CTRL-T software landslide triggering thresholds (more precisely, a classifier able to
provides also traditional E-D power law rainfall thresholds (Eq. 1) predict when a landslide is triggered). The considered architecture
based on triggering events only is that of a 3-layer feedforward neural network with a variable num-
ber of hidden neurons, similarly to Distefano et al. (2022).
E = 𝛼D𝛾 , (1)
Finally, a comparison is made between the thresholds deter-
where α and γ are the threshold parameters, which in a double- mined with the power law and those obtained with ANNs in terms
logarithmic plot correspond, respectively, to the intercept and the of the confusion matrix.
slope of the threshold line. Moreover, power law thresholds using
both TE and NTE and maximizing the receiver operating character-
istics (ROC) true skill statistic (TSS) index have also been derived Reconstruction of precipitation events
(Peres and Cancelliere 2021). The TSS is a measure of the ability of
a classifier to distinguish TE from NTE and is computed using the The use of replicable and objective criteria for the identification
confusion matrix that distinguishes “true positives” (TP, correctly of precipitation events is of key importance. Melillo et al. (2015,
classified TE), “true negatives” (TN, correctly classified NTE), “false 2016, 2018) developed the CTRL-T (calculation of thresholds for

Fig. 1  Flow chart illustrating the proposed methodology

2726 Landslides 20 • (2023)

 2727

rainfall-induced landslides-tool) code in R, for an automated recon-

struction of precipitation events from a rainfall time series and a
landslide inventory. The reconstruction is based on the following
set of climatic and spatial parameters: the minimum period of zero
precipitation (Pw) (hours) in the hot season (cw), and the analogous
period (Pc) in the cold season (cc); the sensitivity of the rain gauge
(gs), and the radius of the circle (Rb) centered on the rain gauge,
within which a landslide falls to its area of influence. The program
includes a module that selects the representative rain gauge to be
associated with the landslide for the computation of triggering pre-
cipitation. The precipitation conditions which are most likely respon-
sible for the trigger are then identified (multiple rainfall conditions,
i.e., associated to multiple rain gauges). The MRC are defined by
a pair (DL, EL) of rainfall event duration (DL) and cumulative pre- Fig. 2  Schematization of the considered artificial neural network
cipitation (EL) or a different set of two or more pairs. Each MRC is inputs, outputs, and architecture
assigned a weight to select the representative rain gauge and a weight
for the rainfall conditions associated with the landslide. The weight
is proportional to the inverse square distance between the rain gauge in this paper. All input data were entered using their natural loga-
and the landslide and proportional to the cumulative rainfall and rithms, which allows for better classification. To find the optimal
the average rainfall intensity. The MRC condition with the heavi- value, or the highest TSS, a different number of hidden neurons
est weight is promoted to MPRC (most probably rainfall condition) were heuristically tested. The number of hidden neurons has been
and it is considered TE. At the same time, the remaining precipi- varied from 5 to 20. The whole rainfall events dataset (TE and NTE)
tation events reconstructed by the algorithm and identified as not has been split in 3 different subsets: training, validation, and test,
most likely to trigger landslides, are then considered NTE. Finally, with proportions of 70%, 15%, and 15% of the total dataset. ANN’s
the algorithm subsequently provides rainfall power law thresholds original output was then transformed into binary: 1 when a land-
in terms of cumulative precipitation and duration (E-D), using the slide is predicted; 0 otherwise. The networks were trained 30 times
method proposed by Brunetti et al. (2010) including the “bootstrap” to assess the impact of the database’s random subdivision on the
technique for uncertainty assessment (Peruccacci et al. 2012), rely- training, validation, and test samples. The range of TSS values is
ing only on TE and on a fixed value of non-exceedance probability. shown by a box and whiskers plot. Also, an uncertainty range ∆,
equal to the difference between the maximum and minimum TSS
associated to the best performing ANNs is shown.
Artificial neural networks for hydro‑meteorological threshold
derivation
ERA5‑Land soil moisture data
ANNs belong to artificial intelligence or machine learning tech-
niques and are universal approximators, capable of reproducing any Soil moisture is a hydrological variable that is affected in a complex
continuous function (Haykin 1999). As demonstrated in Distefano way by all components of the hydrological cycle (Jung et al. 2010;
et al. (2022), choosing a predetermined equation to identify thresh- Crow et al. 2012). Information on soil moisture can be used for
olds, such as the power law, may imply a limitation to prediction reservoir management, irrigation planning, crop yield prediction
performance. The ANNs can be useful in developing predictors that as well as to better understand landslide triggering conditions
overcome the limitations imposed by the choice of a predetermined (Piciullo et al. 2022; Marino et al. 2020; Mirus et al. 2018a, b). There
form of equations for the threshold and easily allow multiple pre- are three possible types of soil moisture data that can be used for
dictor variables to be included. the identification of hydro-meteorological thresholds: (i) in situ
Hydro-meteorological conditions for landslide triggering may measures (Wicki et al. 2020; 2021), (ii) remote sensing (Brocca
consider different additional variables to represent the predispos- et al. 2016), and (iii) reanalysis projects (Reder and Rianna 2021;
ing dynamic factors of initiation such as antecedent rainfall (Glade Palazzolo et al. 2023).
et al. 2000), catchment storage (Bogaard and Greco 2018), or soil As already mentioned in the previous section, soil moisture
moisture (Marino et al. 2020; Palazzolo et al. 2023). data used in this study was retrieved from the ERA5-Land reanal-
In this study, we consider the following input variables: (i) pre- ysis project (Muñoz-Sabater et al. 2021) provided by the European
cipitation duration D, (ii) cumulative precipitation E, (iii) mean pre- Centre for Medium-Range Weather Forecasts (ECMWF). Reanaly-
cipitation intensity I, (iv) volumetric soil water content S at different sis data provide the most complete picture currently available
depths retrieved from ERA5-Land reanalysis. The ANNs have been of past weather and climate and are generally available free of
trained through the scaled conjugate gradient backpropagation charge on dedicated web platforms, released in normal grid for-
algorithm (Møller 1993). ANNs performance was evaluated through mats, with a delay of a few months from the present (Pelosi et al.
“cross-entropy” function (Kline and Berardi 2005). For more infor- 2020). The ERA5-Land soil moisture data have been already used
mation on the features of ANN used in this study, the reader is in some studies (Reder and Rianna 2021; Palazzolo et al. 2023)
referred to Distefano et al. (2022). Figure 2 shows the ANNs used showing the usefulness of this dataset to better understand the
for data processing and input variables employed and presented conditions leading to slope failure. However, ERA5-Land soil

Landslides 20 • (2023) 2727

 2728 Technical Note
Table 1  Some characteristics of the ERA5-Land dataset
Description Features
Data type Gridded
Projection Regular latitude–longitude grid
Horizontal coverage Global
Horizontal resolution 0.1° × 0.1°; Native resolution is 9 km
Vertical coverage From 2 m above the surface level, to a soil depth of 289 cm
Vertical resolution 4 levels of the ECMWF surface model: layer 1: 0–7 cm, layer 2:
7–28 cm, layer 3: 28–100 cm, layer 4: 100–289 cm
Temporal coverage January 1950 to present
Temporal resolution Hourly
File format GRIB
Update frequency Monthly with a delay of about three months relatively to actual date

moisture data have some accuracy limitations, as revealed by The attribution of the soil moisture values to both TE and NTE was
comparison with in situ measurements (Beck et al. 2021). The carried out using a ­MATLAB® code developed for the purpose. Specifi-
schematization of the soil in ERA5-Land includes 4 different lay- cally, the soil moisture of the ERA5-Land cells where the landslides
ers (0–7 cm, 7–28 cm, 28–100 cm, and 100–289 cm). A summary occurred was associated to the TE. Soil moisture values of the ERA5-
of ERA5-Land features is shown in in Table 1. Land cells where there are no landslides were associated to the NTE.
Soil moisture data from the ERA5-Land reanalysis project (four Given that ERA5-Land soil moisture data are derived from
depth layers) were then retrieved and the values at the beginning hydrological modeling using the precipitation of the same rea-
(denoted as S) and end (S’) of each rainfall event were selected. nalysis model, an issue of consistency between it and observed
More precisely, regarding the end, in the case of a TE, soil moisture rainfall (which we use here) may raise. In general, the uncertainty
at the triggering hour was considered, while for NTEs soil moisture of ERA5/ERA5-Land precipitation is significant but still within
was taken at the end of storms. a reasonable range, as several studies evaluating the product by

Fig. 3  Map showing topography, location of rain gauges, and landslides for the Bergen watersheds (Norway), study area

2728 Landslides 20 • (2023)

 2729

Fig. 4  Plot of observed precipitation and ERA5-Land reanalysis soil moisture for case study area of Bergen watersheds (Norway)

comparison with observational datasets indicate a Pearson cor- 2023). Visual consistency checks have been made (see next sec-
relation coefficient around 0.6 for daily data (Sharifi et al. 2019; tions) by plotting the observed precipitation signal versus the
Longo-Minnolo et al. 2022; Xu et al. 2022; Gomis-Cebolla et al. ERA5-Land soil moisture.

Fig. 5  Map showing topography, location of rain gauges, and landslides for Sicily (Italy)

Landslides 20 • (2023) 2729

 2730 Technical Note

Fig. 6  Plot of observed precipitation and ERA5-Land reanalysis soil moisture for case study area of Sicily (Italy)

Study areas and data territory as gridded files with a mesh of 1 km × 1 km (https://​thred​ds.​
met.​no/​thred​ds/​catal​og/​senor​ge/​seNor​ge_​2018/​catal​og.​html, last
Bergen watersheds (Norway) accessed 14/07/2023). The period of analysis goes from January 2000
Norwegian study area is composed of a group of 5 adjacent catch- to December 2020. To simplify the analyses, the number of cells to
ments located east of Bergen, on the western coast of the country. The be analyzed was reduced by choosing 175 cells, which are positioned
­ m2. Altitudes vary
total extent of the study area is about 18 thousand k on a regular grid, with a distance between each individual cell center
between zero and 1600 m a.s.l., as shown in Fig. 3. The average annual equal to 5 km (Fig. 3).
rainfall for the study area is between 3500 and 4700 mm (Meyer et al. Landslide data in the study area comes from the Norwegian Water
2012). The precipitation data used for the elaborations are collected Resources and Energy Directorate (NVE) database, an office under the
from the “seNorge_2018” database (Lussana et al. 2019), which cov- administration of the Ministry of Petroleum and Energy. The NVE
ers a time interval ranging from 1957 to 2020 and takes into account landslide database (https://​nedla​sting.​nve.​no/​gis/, last access 31/01/2022)
solid precipitation. Precipitation data considered, rainfall and contains over 73,528 entries. All database instances are accompanied
snowfall, are on a daily scale and available for the entire Norwegian by various information such as landslide type, trigger date, and spatial

Fig. 7  E-D thresholds identified using TE only and TE and NTE (max TSS)

2730 Landslides 20 • (2023)

 2731

Table 2  List of different combinations of variables for threshold determination by ANNs

Input variables Description Number of Number of Notation
combinations variables
Precipitation duration, intensity, and cumula- Only precipitation data 6 1÷3 DEI
tive amount
Precipitation duration and cumulative Precipitation event duration and cumu- 12 3÷4 DES
amount, and soil moisture lative amount (depth), and soil mois- DES’
ture considering a single layer (initial, DESS’
final, initial and final instants)
Precipitation duration, intensity, and soil Precipitation event duration and inten- 12 3÷4 DIS
moisture sity, and soil moisture considering a DIS’
single layer (initial, final, initial and DISS’
final instants)
Precipitation duration, intensity and cumula- Precipitation event duration, intensity 12 4÷5 DEIS
tive amount, and soil moisture and cumulative amount, and soil DEIS’
moisture considering a single layer DEISS’
(initial, final, initial and final instants)
Precipitation duration and cumulative Precipitation event duration and cumu- 15 4 ÷ 10 DES
amount, and soil moisture lative amount, and soil moisture using DES’
combinations of several layers (initial, DESS’
final, initial and final instants)
Precipitation duration and intensity, and soil Precipitation event duration and inten- 15 4 ÷ 10 DIS
moisture sity, and soil moisture using combina- DIS’
tions of several layers (initial, final, DISS’
initial and final instants)
Precipitation duration, intensity and cumulative Precipitation event duration, inten- 15 5 ÷ 11 DEIS
amount, and soil moisture sity and cumulative amount, and DEIS’
soil moisture using combinations of DEISS’
several layers (initial, final, initial and
final instants)

Fig. 8  Performance comparison of neural classifiers that consider only precipitation characteristics

Landslides 20 • (2023) 2731

 2732 Technical Note
and temporal accuracy. For the analyses, rainfall-induced landslides

0.07

0.02
0.02
0.02
0.03
0.01
Δ
belonging to the following types have been selected: loose material
(LM), debris flow (DF), and debris slide (DS). A further filter has been

mean
applied to remove from the database landslides with a spatial uncer-

FPR

0.42
0.39
0.34

0.36
0.34
0.33
tainty of more than 5 km and temporal uncertainty of more than 24 h.
More information about landslide database can be found at https://​gis3.​
mean nve.​no//​metad​ata/​produ​ktspe​sifik​asjon​er/​produ​ktspe​sifik​asjon_​skred​
hende​lser.​pdf (in Norwegian, last accessed on 13 July 2023). The number
TPR

0.84

0.92
0.87
0.82

0.93
0.91
of landslides matching the above-described characteristics dropped to
298, which are distributed as follows: 147 LM, 51 DF, and 100 DS. The
mean

following values of CTRL-T parameters have been adopted: gs = 1 mm,

test

0.39
0.57
0.41

0.57
0.57
0.53
TSS

Pw = 48 h, Pc = 24 h, and Rb = 16 km. The application of CTRL-T allowed

to reconstruct 247 TE (117 LM, 48 DF, and 82 DS). The location of the
landslides reconstructed by the algorithm is shown in Fig. 3.
mean
valid

As mentioned in the “Methodology” section, a visual consist-

0.42

0.56
0.57
0.57
0.37
0.53
TSS
Table 3  Performances of ANN classifiers considering only precipitation event characteristics for Sicily (Italy) and Bergen watersheds (Norway)

ency check between observed rainfall and reanalysis soil mois-

ture has been carried out. As can be seen in Fig. 4, there is a
mean
train

satisfactory consistency between the two datasets as soil moisture

0.60
0.47

0.42

0.59
0.59
0.55
TSS

increases are usually found just a few instants after precipitation

amounts. Some small inconsistencies are more related to the fact
Bergen watersheds (Norway)

that for the Bergen case study area, precipitation is available only
mean
TSS tot

at the daily scale and not at the hourly as soil moisture.

0.39
0.45

0.58
0.58
0.57
0.53

Sicily (Italy)
mean
neu-
rons
Input Num.

16.27
16.83

16.33

15.87
13.93

16.13

The study area in Italy corresponds to the entire territory of Sic-

­ m2 and is the largest island in
ily, which covers an area of 25,708 k
the Mediterranean Sea. The territory is mainly hilly and moun-
0.01 DEI
0.02 DE
0.01 DI
0.02 D
0.01 E
0.01 I

tainous, with altitudes ranging from sea level to 3320 m a.s.l. on

Mount Etna, the highest active volcano in Europe. The climate is
Δ

temperate-warm, with hot and dry summers, especially on the

southern coasts, and more frequent rain events during the cold-
mean

est winter months, in the mountainous inland areas. The average

0.44
0.44

0.20
0.20
0.20
FPR

0.27

annual rainfall is around 700–800 mm, mainly distributed dur-

ing the autumn and winter seasons. The most abundant rainfall
is recorded in the eastern part of the island on the Ionian side
mean
TPR

0.80

0.80

between Mount Etna and the Peloritani mountains. The hourly

0.79
0.87
0.75

0.71

precipitation data, used in the analysis, comes from 306 rain

gauges managed by the Regional Water Observatory (Osserva-
mean.

torio delle Acque, OdA), the Sicilian Agrometeorological Infor-

test

0.40
0.42
0.28

0.56
0.57
0.55
TSS

mation Service (Servizio Informativo Agrometeoreologico Sicili-

ano, SIAS), and by the Regional Department of Civil Protection
(Dipartimento Regionale di Protezione Civile, DRPC). Figure 5
mean
valid

shows the location of the available rain gauges for the period
0.29

0.56
0.41

0.57
0.41
0.55
TSS

January 2009–October 2018 (white circles) and for the period

January 2014–October 2018 (black circles).
mean

An issue with information on landslides is that in many cases

train

0.48
0.45

0.63
0.63
0.63
0.35
TSS

an uncertain knowledge of the time of triggering may be available

(e.g., morning, afternoon, night) by implying an important impact
on threshold identification (Peres et al. 2018). In the case that hourly
mean
TSS tot

rainfall data is used (Sicily), when exact indication on the trigger

0.60
0.60
0.60
0.44
0.43
0.31

time was not available, it was assumed to coincide with the time of
peak rainfall. If no information was available, the triggering instant
mean

was set at the end of the day (23:00).

neu-
rons
Input Num.
Sicily (Italy)

16.70
16.67
13.20

13.10
15.83

11.67

Landslide information are collected from the FraneItalia data-

base created by Calvello and Pecoraro (2018). From the FraneItalia
database, we retrieved historical landslides observed from January
DEI
DE
DI

2010 to October 2018.

D
E
I

2732 Landslides 20 • (2023)

 2733

Fig. 9  Performance comparison of neural classifiers that consider precipitation event duration and cumulative precipitation and soil mois-
ture data of single layers

In the database, there are two types of landslide events defined Results
as single landslide event (SLE) and areal landslide event (ALE).
Both types were used to reconstruct the largest number of land- Power law thresholds
slide events. SLEs are typically accompanied by more accurate Using the data presented and the CTRL-T code, the duration-
information regarding the place and time of failure. ALEs are a cumulative rainfall thresholds were determined using a power law
set of several landslides affecting the same area and less detailed with a 5% exceedance probability. These thresholds, that consider only
information is often available (e.g., only the municipality where TE, are described by Eqs. (5) and (6) for Italian and Norwegian case
the landslide occurred). Even though the accuracy of the loca- study, respectively, and are shown in Fig. 7.
tion of the SLEs and ALEs can be subject to uncertainty, we kept
those events for deriving the thresholds, as otherwise the dataset E5% = 4.9D0.26 (5)
size would not been sufficient to calibrate, validate, and test the
neural networks. E5% = 5D0.44 (6)
Even for Sicily case study, only rainfall-induced landslides were
selected or those that had mobilized materials attributable to land- On the other hand, thresholds corresponding to the maximum
slides triggered by rainfall (for example, rockfalls were withdrawn value of the TSS (Eq. 4) have been derived. The search of such best
from the analysis), resulting in 207 landslides. The following values performing threshold has been carried out through the particle swarm
of CTRL-T parameters have been adopted: gs = 0.1 mm, Pw = 96 h, algorithm of the global optimization M ­ ATLAB® toolbox. Thresholds
Pc = 48 h, and Rb = 16 km. The use of the software CTRL-T made it for Sicily and Bergen area are respectively (see also Fig. 7)
possible to reconstruct the triggering conditions of 144 out of the 207 E = 2.4D0.68 (7)
landslides for Sicily region. For 126 of these, no information is available
on the type of movement (87.5%). The type of the remaining events
E = 9.05D0.33 (8)
is distributed as follows: 10 (6.9%) rock fall areal landslides, 4 (5.6%)
flows, and 4 (5.6%) slides. Concerning the temporal accuracy of land- The TSS value for the thresholds of both study areas is ­TSS0 = 0.5.
slide data for 103 events, only the day of triggering is known, while for This value of TSS can be considered a benchmark value for the
the remaining a more precise indication—hour or part of the day (e.g., assessment of improvements obtained by our study with ANNs
morning, afternoon, evening)—of the time of slope failure is available. using soil moisture data.
In the first case, the triggering instant was attributed to the end of the
day, in the second case to the peak rainfall of the rainfall event. The
location of the landslides is shown in Fig. 5. ANN thresholds
Even here, the visual consistency check between observed rain-
fall and reanalysis soil moisture (Fig. 6) is satisfactory. In this case, In total, 87 combinations of input variables were analyzed for both study
respect to Bergen case study, consistency is even better due the fact areas (always considering 16 different networks with the number of
that hourly precipitation is available.

Landslides 20 • (2023) 2733

2734
2734

Landslides 20
•
(2023)
Table 4  Performances of ANN classifiers using precipitation event duration and cumulative precipitation and soil moisture at single layers for Sicily (Italy) and Bergen watersheds (Norway)
Sicily (Italy) Bergen watersheds (Norway)
Input Num. TSS tot TSS TSS TSS TPR FPR Δ Input Num. TSS tot TSS TSS TSS TPR FPR Δ
mean mean mean mean mean mean mean mean mean mean mean mean mean mean
neurons train valid test neurons train valid test
14.90 0.75 0.78 0.74 0.72 0.88 0.12 0.02 15.07 0.62 0.64 0.61 0.61 0.87 0.25 0.05
Technical Note

DES1 DES1
DES’1 15.27 0.68 0.71 0.64 0.63 0.85 0.17 0.02 DES’1 16.63 0.70 0.72 0.69 0.68 0.89 0.19 0.06
DES1S’1 15.83 0.75 0.78 0.73 0.72 0.88 0.13 0.02 DES1S’1 15.73 0.75 0.77 0.72 0.72 0.87 0.13 0.04
DES2 13.10 0.76 0.78 0.75 0.74 0.89 0.13 0.02 DES2 16.50 0.64 0.66 0.62 0.61 0.85 0.21 0.04
DES’2 13.70 0.72 0.75 0.70 0.68 0.88 0.16 0.02 DES’2 15.80 0.62 0.65 0.60 0.60 0.83 0.20 0.03
DES2S’2 12.63 0.77 0.79 0.74 0.73 0.89 0.12 0.02 DES2S’2 15.50 0.68 0.70 0.67 0.65 0.85 0.17 0.02
DES3 14.77 0.67 0.70 0.66 0.66 0.85 0.17 0.02 DES3 15.97 0.60 0.62 0.59 0.59 0.88 0.28 0.04
DES’3 16.27 0.69 0.72 0.70 0.67 0.85 0.15 0.02 DES’3 16.40 0.60 0.62 0.59 0.59 0.89 0.29 0.03
DES3S’3 14.97 0.70 0.73 0.69 0.67 0.88 0.18 0.04 DES3S’3 16.37 0.70 0.72 0.69 0.68 0.91 0.21 0.05
DES4 16.67 0.62 0.65 0.61 0.59 0.84 0.21 0.02 DES4 15.90 0.66 0.68 0.64 0.64 0.89 0.23 0.03
DES’4 15.50 0.63 0.66 0.60 0.60 0.83 0.20 0.04 DES’4 15.17 0.70 0.71 0.69 0.68 0.93 0.23 0.02
DES4S’4 15.17 0.63 0.67 0.61 0.60 0.82 0.19 0.03 DES4S’4 15.80 0.78 0.79 0.78 0.76 0.94 0.17 0.03
 2735

Fig. 10  Performance comparison of neural classifiers that consider precipitation event duration and cumulative precipitation and soil mois-
ture data of multiple layers

neurons in the hidden layer varying between 5 and 20) for a total of Figure 10 presents the results of the analysis where the ANNs
(2 × 30 × 16 × 87) 83,520 trained ANNs as described in Table 2. use the combination of several soil moisture layers as input. Also,
The first implementation of ANNs considers only the precipita- in this case, the soil moisture conditions at the beginning of the rain
tion characteristics. The results of the comparison are shown in event (Si), at the possible triggering instant (Si’), and a combination
Fig. 8 and Table 3. (SiSi’) thereof are considered. The notation Sall indicates that all the
As can be seen in the graphs, the best TSS values are obtained by layers in both initial and final moments (i.e., a total of 8 variables)
considering the precipitation duration-depth (D, E) and precipita- are used as input to the ANN. The performances of the classifiers
tion duration-intensity (D, I) thresholds for both study areas. The that consider rainfall duration and depth and soil moisture are pre-
introduction of a third variable does not improve the recognition sented in Fig. 10, while the results of the proposed elaborations are
of the conditions that can lead to the triggering of the landslide, as reported in Table 5. The results obtained using other input variables
already shown in Distefano et al. (2022). (DIS; DEIS) in graphical and numerical form are shown in the sup-
Table 3 shows the input data for the Italian and Norwegian case plementary material.
studies, and in particular, the average number of neurons for which
the best TSS value is obtained, together with the mean TSS value, Discussion
the TSS value of the training validation and test phases, the aver- Table 6 shows a summary of the characteristics of the neural net-
age values of TPR and FPR, as well as the uncertainty range Δ. works tested reporting for each configuration the best TSS value,
The mean TSS values for both study areas, considering two input the mean TSS value, and the ∆ value corresponding to the maxi-
variables, are between 0.57 and 0.6, and in any case, they are sig- mum TSS. For the data in the table, it is not distinguished whether
nificantly higher than ­TSS0, which is equal to 0.50. they are obtained with S or S’ or a combination thereof.
Thresholds have also been determined including ERA5-Land soil For the Italian case study, the combinations considering sin-
moisture data. The use of these additional variables improves the pre- gle layers of soil moisture that lead to the best prediction per-
dictive capabilities of ANNs as shown in Fig. 9. The cumulative rain- formances are those including the soil moisture information of
fall, duration, and ERA5-Land soil moisture data at each single layer at the first and second layer, and in particular, the combinations
the initial instant (Si) or at the final instant (S’i), and both at the initial of precipitation characteristics with S1, S2, S1S’1, and S2S’2. The
and final instants (SiS’i) are considered input of the ANNs. The results performance of the neural predictors stands at an average TSS
are presented in Fig. 9 and Table 4. Remaining input combinations value close to 0.75. With reference to the Norwegian case study,
considering duration-intensity-soil moisture, duration-cumulative the best results are obtained with the combinations of precipita-
precipitation-intensity-soil moisture data are shown in the supple- tion characteristics with S1S’1 and S4S’4. The good performance
mentary material (Figs. S1 and S3, Tables S1 and S3). obtained using the data of the fourth soil layer turns out to be
somewhat counterintuitive but still compatible with the possible

Landslides 20 • (2023) 2735

2736
2736

Landslides 20
•
Table 5  Performances of ANNs using precipitation event duration and cumulative precipitation and soil moisture at multiple layers for Sicily (Italy) and Bergen watersheds (Norway)

(2023)
Sicily (Italy) Bergen watersheds (Norway)
Input Num. TSS tot TSS TSS TSS TPR FPR Δ Input Num. TSS tot TSS TSS TSS TPR FPR Δ
mean mean mean mean mean mean mean mean mean mean mean mean mean mean
neurons train valid test neu- train valid test
rons
DES1S2 13.90 0.77 0.79 0.74 0.73 0.88 0.11 0.01 DES1S2 16.23 0.64 0.66 0.62 0.61 0.85 0.21 0.03
DES’1S’2 15.03 0.73 0.75 0.70 0.71 0.88 0.15 0.04 DES’1S’2 16.50 0.74 0.76 0.73 0.72 0.90 0.16 0.07
Technical Note

DES12S’12 14.90 0.77 0.80 0.75 0.74 0.88 0.11 0.03 DES12S’12 16.70 0.77 0.79 0.74 0.75 0.88 0.11 0.03
DES123 15.67 0.77 0.80 0.76 0.74 0.89 0.12 0.04 DES123 15.00 0.63 0.65 0.61 0.61 0.83 0.20 0.04
DES’123 14.17 0.75 0.77 0.73 0.73 0.90 0.16 0.04 DES’123 15.10 0.73 0.75 0.72 0.71 0.86 0.12 0.12
DES123S’123 15.97 0.79 0.81 0.77 0.76 0.89 0.10 0.03 DES123S’123 16.63 0.80 0.82 0.77 0.78 0.86 0.06 0.04
DES1234 14.63 0.78 0.81 0.74 0.75 0.89 0.11 0.03 DES1234 16.10 0.66 0.69 0.65 0.65 0.87 0.21 0.05
DES’1234 15.87 0.74 0.76 0.72 0.72 0.90 0.16 0.05 DES’1234 15.20 0.81 0.83 0.79 0.78 0.89 0.08 0.05
DESall 15.03 0.79 0.81 0.77 0.76 0.89 0.10 0.04 DESall 16.60 0.89 0.90 0.87 0.86 0.94 0.05 0.05
DES23 14.97 0.78 0.80 0.75 0.74 0.88 0.10 0.03 DES23 13.93 0.63 0.65 0.61 0.61 0.81 0.18 0.04
DES’23 14.67 0.74 0.76 0.70 0.72 0.90 0.16 0.03 DES’23 16.30 0.65 0.68 0.63 0.62 0.84 0.19 0.06
DES23S’23 14.93 0.79 0.81 0.75 0.76 0.88 0.10 0.03 DES23S’23 17.23 0.74 0.76 0.72 0.71 0.88 0.14 0.06
DES234 14.77 0.78 0.81 0.75 0.75 0.89 0.11 0.02 DES234 16.20 0.66 0.68 0.65 0.64 0.87 0.21 0.05
DES’234 14.17 0.73 0.76 0.72 0.71 0.89 0.16 0.03 DES’234 16.07 0.76 0.78 0.75 0.75 0.91 0.14 0.07
DES234S’234 15.23 0.79 0.82 0.76 0.77 0.89 0.09 0.03 DES234S’234 16.17 0.88 0.89 0.86 0.86 0.95 0.07 0.06
 2737

Table 6  Summary table of TSS and variability values of various neu- in prediction performance compared to triggering thresholds based
ral network models on precipitation variables only. Specifically, considering soil moisture
data of single layers, there is an increase in the average values of the
TSS max TSS mean ∆
TSS from 0.58 ÷ 0.59 (ANNs using precipitation data only) to values
DES single layer Sicily 0.77 0.70 0.02 ranging between 0.65 ÷ 0.78 for both study areas. The use of multi-
layer soil moisture information at both the beginning and the end
Bergen 0.78 0.67 0.03
of the event yields even a better performance as TSS = 0.79 and 0.89,
DIS single layer Sicily 0.77 0.67 0.03 respectively for Sicily (Italy) and Bergen area (Norway). It has been
Bergen 0.78 0.67 0.03 seen that the highest values of TSS are obtained by considering the
first and second soil moisture layers in the Italian case study, and the
DEIS single layer Sicily 0.76 0.70 0.02 first and fourth layers for the Norwegian case study. Results indicate
Bergen 0.78 0.67 0.03 that the best performances and the least uncertainty are obtained
with networks using precipitation duration, depth, and volumetric
DES multilayer Sicily 0.79 0.77 0.03–0.04 soil water content at all the four depth layers. Overall, these results
Bergen 0.89 0.73 0.05 demonstrate that the proposed methodology is particularly promis-
ing in improving landslide prediction. However, in some cases, the
DIS multilayer Sicily 0.79 0.76 0.06 use of the whole bulk of information does not bring to a significant
Bergen 0.88 0.73 0.07 improvement compared to some other simpler input combinations.
In such cases, there could be opportunities for simplification of the
DEIS multilayer Sicily 0.79 0.77 0.03–0.04 ANNs, which can be explored, for example, by techniques such as
Bergen 0.89 0.73 0.08 principal component analysis that exploit the fact that soil moisture
at different depths exhibits some correlation (Palazzolo et al. 2023).

Acknowledgements
range of failure depths of shallow landslides, as this can be up to Pierpaolo Distefano expresses his sincere gratitude to NGI mem-
2 or 3 m deep (Perry 1989; Kayyal 1991; Dai et al. 2003; Dahal et al. bers, especially Natural hazards department, where he had the priv-
2009; Huang et al. 2022), and thus influenced by soil moisture ilege of being a PhD visiting student for a six-months period. Their
changes in the fourth ERA5-Land layer (average depth 1.89 m, support, guidance, and collaboration have contributed significantly
range 1–2.89 m). The variability in TSS value, expressed by the to his academic and personal growth.
parameter Δ and visually by the width of the box plots, appears to
be greater in the Norwegian case study, probably due to the type Author contribution
of rainfall that characterizes the study area, climatic conditions, Conceptualization was done by DJP and LP; formal analysis by PD,
and the freeze–thaw cycles that can occur there. Overall, includ- LP, NP, and DJP; investigation by PD, LP, and DJP; methodology by
ing soil moisture data increases the performance of the predictor PD, LP, and DJP; coding and mapping by PD, NP, and DJP; supervi-
compared to the case where only precipitation data is included for sion by LP, DJP, PS, and AC; writing the original draft by PD and
both case studies (Figs. 9, S1 and S3). DJP; and the writing, review, and editing by PD, DJP, LP, and NP. DJP,
When considering predictors using moisture information from LP, PS, and AC supervised the research. All authors have read and
multiple soil layers, different trends are observed for the study agreed to the published version of the paper.
areas. For the Italian case study, it is possible to observe a lower
variation of the TSS mean values, which range between 0.73 ÷ 0.79.
The best average TSS values are obtained with the combinations Funding
S234S’234, whatever the combination of precipitation characteristics. Pierpaolo Distefano’s doctoral program is funded by the “Notice
The variation of ANN performance is much more pronounced 2/2019 for financing the PhD regional grant in Sicily” as part of
in the Norwegian case study, where the TSS mean values range the Operational Programme of European Social Funding 2014–
between 0.62 ÷ 0.89. The best mean TSS value is obtained with the 2020 (PO FSE 2014–2020, CUP E65E19000830002). Nunziarita
combination Sall that considers all soil layers (Fig. 10, S2 and S4). Palazzolo is supported by a post-doctoral program funded by
the project “Autorità di Bacino del Distretto Idrografico della
Conclusion Sicilia—Interventi per il miglioramento dei corpi idrici CUP:
In this work, the classification capabilities of ANNs in the develop- F62G16000000001.” This research was partially carried out within
ment of hydro-meteorological thresholds for landslides forecasting the projects HydrEx—Hydrological extremes in a changing
have been investigated. ANNs have the advantage of allowing the deri- climate— and VARIO—VAlutazione del Rischio Idraulico in sistemi
vation of thresholds without a specific parametric equation and are cOmplessi—Piano di incentivi per la ricerca di Ateneo (Pia.ce.ri.),
also suitable for developing thresholds that use several input variables. 2020–2022, Università di Catania, and the Ministero dell’Università
We considered soil moisture information at different depths from the e della Ricerca (Programma Operativo Nazionale Ricerca e
ERA5-Land reanalysis project. The analyses were conducted refer- Innovazione 2014–2020—Progetto “reCITY—Resilient City
ring to two study areas characterized by completely different climatic Everyday Revolution”—grant agreement no. ARS01_00592—CUP
regimes and geomorphological characteristics. For both areas, the B69C21000390005). Pierpaolo Distefano’s post-doc is also funded
inclusion of soil moisture data resulted in significant improvements by the project “reCITY—Resilient City—Everyday Revolution.”

Landslides 20 • (2023) 2737

 2738 Technical Note
Data availability Distefano P, Peres DJ, Scandura P, Cancelliere A (2022) Brief communi-
Landslide data are available at https://​frane​italia.​wordp​ress.​com/ cation: introducing rainfall thresholds for landslide triggering based
on artificial neural networks. Nat Hazards Earth Syst Sci 1151–1157
(Italy) and https://​nedla​sting.​nve.​no/​gis/ (Norway). Precipitation
Fawcett T (2006) An introduction to ROC analysis. Pattern Recognit Lett
data for Norway are available at https://​thred​ds.​met.​no/​thred​ds/​ 27:861–874. https://​doi.​org/​10.​1016/j.​patrec.​2005.​10.​010
catal​og/​senor​ge/​seNor​ge_​2018/​catal​og.​html, while for Sicily are Froude MJ, Petley DN (2018) Global fatal landslide occurrence from 2004
available from several sources (http://​www.​sias.​regio​ne.​sicil​ia.​ to 2016. Nat Hazards Earth Syst Sci 18:2161–2181. https://​doi.​org/​10.​
it/, https://​www.​prote​zione​civil​esici​lia.​it:​8443/​aegis/​map/​map2d). 5194/​nhess-​18-​2161-​2018
Gariano SL, Brunetti MT, Iovine G, Melillo M, Peruccacci S, Terranova
Codes and datasets can be made available upon request to the O, Vennari C, Guzzetti F (2015) Calibration and validation of rainfall
authors. thresholds for shallow landslide forecasting in Sicily, southern Italy.
Geomorphology 228:653–665. https://​doi.​org/​10.​1016/J.​GEOMO​RPH.​
Declarations 2014.​10.​019
Glade T, Crozier M, Smith P (2000) Applying probability determination
Competing interests All authors declare no competing interests. to refine landslide-triggering rainfall thresholds using an empirical
“antecedent daily rainfall model.” Pure Appl Geophys 157:1059–1079.
https://​doi.​org/​10.​1007/​s0002​40050​017
Gomis-Cebolla J, Rattayova V, Salazar-Galán S, Francés F (2023) Eval-
uation of ERA5 and ERA5-land reanalysis precipitation datasets
over Spain (1951–2020). Atmos Res 284. https://​d oi.​o rg/​1 0.​5 194/​
References hess-​23-​207-​2019
Guzzetti F, Peruccacci S, Rossi M, Stark CP (2007) Rainfall thresholds for
Aleotti P (2004) A warning system for rainfall-induced shallow failures. the initiation of landslides in central and southern Europe. Meteorol
Eng Geol 73:247–265. https://​doi.​org/​10.​1016/J.​ENGGEO.​2004.​01.​007 Atmos Phys 98:239–267. https://​doi.​org/​10.​1007/​s00703-​007-​0262-7
Beck HE, Pan M, Miralles DG, Reichle RH, Dorigo WA, Hahn S, Sheffield J, Haque U, da Silva PF, Devoli G, Pilz J, Zhao B, Khaloua A, Wilopo W,
Karthikeyan L, Balsamo G, Parinussa RM, van Dijk AIJM, Du J, Kimball Andersen P, Lu P, Lee J, Yamamoto T, Keellings D, Jian-Hong W, Glass
JS, Vergopolan N, Wood EF (2021) Evaluation of 18 satellite- and GE (2019) The human cost of global warming: deadly landslides and
model-based soil moisture products using in situ measurements from their triggers (1995–2014). Sci Total Environ 682:673–684. https://​doi.​
826 sensors. Hydrol Earth Syst Sci 25:17–40. https://​doi.​org/​10.​5194/​ org/​10.​1016/j.​scito​tenv.​2019.​03.​415
hess-​25-​17-​2021 Haykin S (1999) Neural networks- a comprehensive foundation 2nd rd.
Bogaard T, Greco R (2018) Invited perspectives: hydrological perspec- Prentice Hall
tives on precipitation intensity-duration thresholds for landslide Huang W, Loveridge F, Satyanaga A (2022) Translational upper bound
initiation: proposing hydro-meteorological thresholds. Nat Hazards limit analysis of shallow landslides accounting for pore pressure
Earth Syst Sci 18:31–39. https://​doi.​org/​10.​5194/​nhess-​18-​31-​2018 effects. Comput Geotech 148:104841. https://​d oi.​o rg/​1 0.​1 016/j.​
Brocca L, Ciabatta L, Moramarco T, Ponziani F, Berni N, Wagner W (2016) Use of compg​eo.​2022.​104841
satellite soil moisture products for the operational mitigation of landslides Jung M, Reichstein M, Ciais P, Seneviratne SI, Sheffield J, Goulden ML,
risk in Central Italy, satellite soil moisture retrieval: techniques and appli- Bonan G, Cescatti A, Chen J, De Jeu R, Dolman AJ, Eugster W, Gerten
cations. Elsevier Inc. https://​doi.​org/​10.​1016/​B978-0-​12-​803388-​3.​00012-7 D, Gianelle D, Gobron N, Heinke J, Kimball J, Law BE, Montagnani L, Mu
Brunetti MT, Peruccacci S, Rossi M, Luciani S, Valigi D, F.G., (2010) Rain- Q, Mueller B, Oleson K, Papale D, Richardson AD, Roupsard O, Running
fall thresholds for the possible occurrence of landslides in Italy. Nat S, Tomelleri E, Viovy N, Weber U, Williams C, Wood E, Zaehle S, Zhang
Hazards Earth Syst Sci 47:633–635. https://​doi.​org/​10.​1016/​S1387-​ K (2010) Recent decline in the global land evapotranspiration trend
6473(03)​00110-6 due to limited moisture supply. Nature 467:951–954. https://​doi.​org/​
Caine N (1980) The rainfall intensity-duration control of shallow land- 10.​1038/​natur​e09396
slides and debris flows. Geogr Ann Ser A Phys Geogr 62:23–27 Kayyal MK (1991) Investigation of long-term strength properties of
Calvello M, Pecoraro G (2018) FraneItalia: a catalog of recent Italian Paris and Beaumont clays in earth embankments. Research Report
landslides. Geoenvironmen Disasters 5. https://​d oi.​o rg/​1 0.​1 186/​ 1195–2F, Center for Transportation Research, University of Texas at
s40677-​018-​0105-5 Austin, Austin, Texas
Campbell RH (1975) Debris flows originating from soil slips during rain- Kline DM, Berardi VL (2005) Revisiting squared-error and cross-entropy
storms in Southern California. Q J Eng Geol 7:339–349. https://​doi.​org/​ functions for training neural network classifiers. Neural Comput Appl
10.​1144/​GSL.​QJEG.​1974.​007.​04.​04 14:310–318. https://​doi.​org/​10.​1007/​s00521-​005-​0467-y
Collini E, Palesi LAI, Nesi P, Pantaleo G, Nocentini N, Rosi A (2022) Predicting Longo-Minnolo G, Vanella D, Consoli S, Pappalardo S, Ramírez-Cuesta JM
and understanding landslide events with explainable AI. IEEE Access (2022) Assessing the use of ERA5-land reanalysis and spatial interpolation
10:31175–31189. https://​doi.​org/​10.​1109/​ACCESS.​2022.​31583​28 methods for retrieving precipitation estimates at basin scale. Atmos Res
Conrad JL, Morphew MD, Baum RL, Mirus BB (2021) Hydromet: a new 271:106131. https://​doi.​org/​10.​1016/j.​atmos​res.​2022.​106131
code for automated objective optimization of hydro-meteorological Lussana C, Einar Tveito O, Dobler A, Tunheim K (2019) SeNorge_2018,
thresholds for landslide initiation. Water (Switzerland) 13. https://​doi.​ daily precipitation, and temperature datasets over Norway. Earth Syst
org/​10.​3390/​w1313​1752 Sci Data 11:1531–1551. https://​doi.​org/​10.​5194/​essd-​11-​1531-​2019
Crow WT, Berg AA, Cosh MH, Loew A, Mohanty BP, Panciera R, De Rosnay Marino P, Peres DJ, Cancelliere A, Greco R, Bogaard TA (2020). Soil mois-
P, Ryu D, Walker JP (2012) Upscaling sparse ground-based soil mois- ture information can improve shallow landslide forecasting using the
ture observations for the validation of coarse-resolution satellite soil hydro-meteorological threshold approach. https://​doi.​org/​10.​1007/​
moisture products. Rev Geophys 50:1–20. https://​doi.​org/​10.​1029/​2011R​ s10346-​020-​01420-8
G0003​72 Melillo M, Brunetti MT, Peruccacci S, Gariano SL, Guzzetti F (2016) Rainfall
Dahal RK, Hasegawa S, Nonomura A, Yamanaka M, Masuda T, Nishino K thresholds for the possible landslide occurrence in Sicily (Southern
(2009) Failure characteristics of rainfall-induced shallow landslides Italy) based on the automatic reconstruction of rainfall events. Land-
in granitic terrains of Shikoku island of Japan. Environ Geol 56:1295– slides 13:165–172. https://​doi.​org/​10.​1007/​s10346-​015-​0630-1
1310. https://​doi.​org/​10.​1007/​s00254-​008-​1228-x Melillo M, Brunetti MT, Peruccacci S, Gariano SL, Guzzetti F (2015) An
Dai FC, Lee CF, Wang SJ (2003) Characterization of rainfall-induced algorithm for the objective reconstruction of rainfall events respon-
landslides. Int J Remote Sens 24:4817–4834. https://​doi.​org/​10.​1080/​ sible for landslides. Landslides 12:311–320. https://​doi.​org/​10.​1007/​
01431​16011​31000​082424 s10346-​014-​0471-3

2738 Landslides 20 • (2023)

 2739

Melillo M, Brunetti MT, Peruccacci S, Gariano SL, Roccati A, Guzzetti F Piciullo L, Gariano SL, Melillo M, Brunetti MT, Peruccacci S, Guzzetti F,
(2018) A tool for the automatic calculation of rainfall thresholds for Calvello M (2017) Definition and performance of a threshold-based
landslide occurrence. Environ Model Softw 105:230–243. https://​doi.​ regional early warning model for rainfall-induced landslides. Land-
org/​10.​1016/J.​ENVSO​FT.​2018.​03.​024 slides 14:995–1008. https://​doi.​org/​10.​1007/​s10346-​016-​0750-2
Meyer NK, Dyrrdal AV, Frauenfelder R, EtzelmÃller B, Nadim F (2012) Pota M, Pecoraro G, Rianna G, Reder A, Calvello M, Esposito M (2022)
Hydro-meteorological threshold conditions for debris flow initiation Machine learning for the definition of landslide alert models: a case
in Norway. Nat Hazards Earth Syst Sci 12:3059–3073. https://​doi.​org/​ study in Campania region, Italy. Discov Artif Intell 2. https://​doi.​org/​
10.​5194/​nhess-​12-​3059-​2012 10.​1007/​s44163-​022-​00033-5
Mirus BB, Becker RE, Baum RL, Smith JB (2018a) Integrating real-time Reder A, Rianna G (2021) Exploring ERA5 reanalysis potentialities for
subsurface hydrologic monitoring with empirical rainfall thresholds supporting landslide investigations: a test case from Campania region
to improve landslide early warning. Landslides 15:1909–1919. https://​ (Southern Italy). Landslides 18:1909–1924. https://​doi.​org/​10.​1007/​
doi.​org/​10.​1007/​s10346-​018-​0995-z s10346-​020-​01610-4
Mirus BB, Morphew MD, Smith JB (2018b) Developing hydro-meteorological Rosi A, Segoni S, Canavesi V, Monni A, Gallucci A, Casagli N (2021) Defi-
thresholds for shallow landslide initiation and early warning. Water (swit- nition of 3D rainfall thresholds to increase operative landslide early
zerland) 10:1–19. https://​doi.​org/​10.​3390/​W1009​1274 warning system performances. Landslides 18:1045–1057. https://​doi.​
Møller MF (1993) A scaled conjugate gradient algorithm for fast super- org/​10.​1007/​s10346-​020-​01523-2
vised learning, neural networks. https://​d oi.​o rg/​1 0.​1 016/​S 0893-​ Segoni S, Piciullo L, Gariano SL (2018) A review of the recent literature
6080(05)​80056-5 on rainfall thresholds for landslide occurrence. Landslides 22:478–494.
Muñoz-Sabater J, Dutra E, Agustí-Panareda A, Albergel C, Arduini G, Balsamo https://​doi.​org/​10.​1007/​s10346-​018-​0966-4
G, Boussetta S, Choulga M, Harrigan S, Hersbach H, Martens B, Miralles Sharifi E, Eitzinger J, Dorigo W (2019) Performance of the state-of-the-art
DG, Piles M, Rodríguez-Fernández NJ, Zsoter E, Buontempo C, Thépaut gridded precipitation products over mountainous terrain: a regional
JN (2021) ERA5-land: a state-of-the-art global reanalysis dataset for land study over Austria. Remote Sens 11:1–20. https://​doi.​org/​10.​3390/​
applications. Earth Syst Sci Data 13:4349–4383. https://​doi.​org/​10.​5194/​ rs111​72018
essd-​13-​4349-​2021 Steger S, Moreno M, Crespi A, Zellner PJ, Gariano SL, Brunetti T, Melillo M,
Palazzolo N, Peres DJ, Creaco E, Cancelliere A (2023) Using principal Peruccacci S, Marra F, Kohrs R, Goetz J, Mair V, Pittore M (2022) Deci-
component analysis to incorporate multi-layer soil moisture informa- phering seasonal effects of triggering and preparatory precipitation
tion in hydro-meteorological thresholds for landslide prediction: an for improved shallow landslide prediction using generalized additive
investigation based on ERA5-land reanalysis data. Nat Hazards Earth mixed models 1–38
Syst Sci 1–22 UNDRR (2020) UN office for disaster risk reduction - human cost of dis-
Pelosi A, Terribile F, D’Urso G, Chirico GB (2020) Comparison of ERA5-land asters - an overview of the last 20 years 2000–2019
and UERRA MESCAN-SURFEX reanalysis data with spatially interpo- Wicki A, Jansson PE, Lehmann P, Hauck C, Stähli M (2021) Simulated or
lated weather observations for the regional assessment of reference measured soil moisture: which one is adding more value to regional
evapotranspiration. Water (Switzerland) 12. https://​doi.​org/​10.​3390/​ landslide early warning? Hydrol Earth Syst Sci 25:4585–4610. https://​
W1206​1669 doi.​org/​10.​5194/​hess-​25-​4585-​2021
Peres DJ, Cancelliere A (2021) Comparing methods for determining Wicki A, Lehmann P, Hauck C, Seneviratne SI, Waldner P, Stähli M (2020)
landslide early warning thresholds: potential use of non-triggering Assessing the potential of soil moisture measurements for regional
rainfall for locations with scarce landslide data availability. Landslides. landslide early warning. Landslides 17:1881–1896. https://​doi.​org/​10.​
https://​doi.​org/​10.​1007/​s10346-​021-​01704-7 1007/​s10346-​020-​01400-y
Peres DJ, Cancelliere A, Greco R, Bogaard TA (2018) Influence of uncer- Xu J, Ma Z, Yan S, Peng J (2022) Do ERA5 and ERA5-land precipitation
tain identification of triggering rainfall on the assessment of land- estimates outperform satellite-based precipitation products? A com-
slide early warning thresholds. Nat Hazards Earth Syst Sci 18:633–646. prehensive comparison between state-of-the-art model-based and
https://​doi.​org/​10.​5194/​nhess-​18-​633-​2018 satellite-based precipitation products over mainland China. J Hydrol
Perry J (1989) A survey of slope condition on motorway earthworks in 605:127353. https://​doi.​org/​10.​1016/j.​jhydr​ol.​2021.​127353
England and Wales. Research Report 199. Transport & Road Research
Laboratory, Crowthorne Supplementary Information The online version contains
Peruccacci S, Brunetti MT, Gariano SL, Melillo M, Rossi M, Guzzetti F supplementary material available at https://​doi.​org/​10.​1007/​
(2017) Rainfall thresholds for possible landslide occurrence in Italy. s10346-​023-​02132-5.
Geomorphology 290:39–57. https://​doi.​org/​10.​1016/J.​GEOMO​RPH.​
2017.​03.​031
Peruccacci S, Brunetti MT, Luciani S, Vennari C, Guzzetti F (2012) Litho- Pierpaolo Distefano · David J. Peres (*) · Nunziarita Palazzolo ·
logical and seasonal control on rainfall thresholds for the possible ini- Pietro Scandura · Antonino Cancelliere
tiation of landslides in central Italy. Geomorphology 139–140:79–90. Department of Civil Engineering and Architecture, University
https://​doi.​org/​10.​1016/J.​GEOMO​RPH.​2011.​10.​005 of Catania, Catania, Italy
Piciullo L, Capobianco V, Heyerdahl H (2022) A first step towards a IoT- Email: [email protected]
based local early warning system for an unsaturated slope in Nor-
way, natural hazards. Springer, Netherlands. https://​doi.​org/​10.​1007/​ Luca Piciullo
s11069-​022-​05524-3 Norwegian Geotechnical Institute, Oslo, Norway

Landslides 20 • (2023) 2739