Content uploaded by Danilo Godone
Author content
All content in this area was uploaded by Danilo Godone on Apr 08, 2024
Content may be subject to copyright.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
Available online 5 April 2024
1569-8432/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
The unsuPervised shAllow laNdslide rapiD mApping: PANDA method
applied to severe rainfalls in northeastern appenine (Italy)
Davide Notti , Martina Cignetti
*
, Danilo Godone , Davide Cardone , Daniele Giordan
National Research Council of Italy, Research Institute for Geo-Hydrological Protection (CNR IRPI), Torino 10135, Italy
ARTICLE INFO
Keywords:
Semi-automatic processing
Sentinel-2
Extreme event
Emergency management
Residual risk
Change detection
Emilia-Romagna Region
ABSTRACT
Shallow landslides, frequently triggered by extreme events such as heavy rainfall, snowmelt, or earthquakes,
affect vast areas with remarkable density. In the immediate aftermath of such events, it becomes crucial to
rapidly assess landslides distribution and pinpoint the most severely affected areas to prioritize damage as-
sessments and guide eld survey operations effectively. Once the emergency phase subsides, the attention can
shift to enhancing the accuracy of landslide inventory. In this work, we introduce the two-phase methodology
“PANDA”, the unsuPervised shAllow laNdslide rapiD mApping, for the low-cost mapping of the potential
landslides, rstly in the emergency phase and then, with an improved version, in the post-emergency one. This
approach utilizes variations in NDVI derived from Sentinel-2 satellite imagery and geomorphological lters. We
applied PANDA to rainfall events in the northeastern Apennine range, Italy, occurred in May 2023, causing
dramatic social and economic consequences for this mountain territory. Within just ve days of obtaining
Sentinel-2 post-event imagery, we produced a reliable, ready-to-use map covering a vast area (~4000 km
2
). The
map tested during emergency eld mapping shows positive feedback. In the post-emergency phase, accuracy was
enhanced using completely cloud-free imagery, a lter to identify false positives associated with land use
changes, a higher resolution digital terrain model (DTM), and an iterative approach to optimize NDVI and slope
thresholds. Potential landslide density related with rainfall, indicating that the most severely affected region
attained a density of approximately 50 landslides/km
2
. Validation against an independent manual inventory
based on high-resolution imagery demonstrated encouraging accuracy results from both inventories, with a
noticeable increase in the F1 score for the post-emergency version.
1. Introduction
Extreme rainfall events are one of the primary triggers of shallow
landslides (Guzzetti et al., 2008). They usually initiate almost simulta-
neously, densely distributed and involving large territory (Bessette-
Kirton et al., 1950; Cignetti et al., 2019; Guzzetti et al., 2004), causing
severe damage, especially, to the road network (Bordoni et al., 2018;
Giordan et al., 2017). After those events, one of the rst activities
necessary to evaluate the damage is a rapid mapping of landslides.
Landslide event inventory mapping is one of the main approaches to
estimating the level of risk posed by geo-hydrological hazards, playing a
crucial role in recording the date of triggered events and evaluating the
spatial extent and the magnitude of landslides (Corominas et al., 2014;
Guzzetti et al., 2005; van Westen et al., 2008). Many approaches have
been proposed for landslide event inventory mapping (Novellino et al.,
2024). Traditional methods rely on 3D visual interpretation of aerial
photos or satellite images, coupled with extensive eld surveys (Brar-
dinoni et al., 2003; Santangelo et al., 2015). The increased availability of
high-resolution images (from Unmanned Aerial Vehicle (UAV), aerial or
satellite platforms) (Casagli et al., 2017; Giordan et al., 2020) allowed
accurate mapping of small landslides. Unfortunately, these products are
rarely free-cost and require an acquisition plan and/or only cover some
of the area affected. Moreover, detailed mapping needs a time-
consuming expert manual approach (Donnini et al., 2023; Guzzetti
et al., 2012) or the development of a machine learning algorithm that
requires a high level of expertise (Đuri´
c et al., 2017; Galli et al., 2008; Lu
et al., 2019; Meena et al., 2023; Murillo-García et al., 2015). The advent
of remote sensing technologies has opened opportunities for the devel-
opment of methodologies and techniques for automating landslides
mapping through the utilization of various high-resolution optical
* Corresponding author.
E-mail addresses: davide.notti@irpi.cnr.it (D. Notti), martina.cignetti@irpi.cnr.it (M. Cignetti), danilo.godone@irpi.cnr.it (D. Godone), davide.cardone@irpi.cnr.it
(D. Cardone), daniele.giordan@irpi.cnr.it (D. Giordan).
Contents lists available at ScienceDirect
International Journal of Applied Earth
Observation and Geoinformation
journal homepage: www.elsevier.com/locate/jag
https://doi.org/10.1016/j.jag.2024.103806
Received 19 January 2024; Received in revised form 18 March 2024; Accepted 29 March 2024
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
2
sensor products, signicantly reducing the timing for the mapping
(Casagli et al., 2016; Guzzetti et al., 2012). Recent studies (Nava et al.,
2022; Santangelo et al., 2022) use also the amplitude of Synthetic
Aperture Radar (SAR) derived products to overcome the problem of
cloud cover of multi-spectral satellite data. However, SAR techniques
introduce some other limitations (e.g., vegetation and snow coverage,
shawdowing, decorrelation noise), necessitating the expertise of users to
process the data.
The Sentinel-2 satellites, with a 5-day revisiting pass, allow medium-
resolution images over wide areas and the possibility to compare the
post-event images with the same condition of sunlight direction angle
(Drusch et al., 2012). The availability of these images allows a rapid,
fully or partially, automatic mapping of potential landslides (PL) over
vast areas as soon as cloud-free images are available using the variation
of NDVI (Ghorbanzadeh et al., 2022b; Notti et al., 2023b). These maps
aim to estimate the severity of the event and its impact, and to focus the
high-resolution mapping on the most affected areas.
Despite NDVI-based landslide mapping methods offering rapid re-
sults, and low computational and economic costs, ensuring a user-
friendly approach even for non-expert users, diverse drawbacks need
to be taken into account. Final products may be inuenced by shadowed
areas, cloud coverage, NDVI variations associated with agricultural ac-
tivities and deforestation, as well as river erosional processes, poten-
tially leading to an increased number of false positives in the mapped
results. In recent years, various methods that aim to solve the limitation
related to the use of NDVI images have emerged from different re-
searchers. Fiorucci et al., (2019) proposed a manual mapping approach
that leverages NDVI images coupled with digital stereoscopy for the 3D
visualization of landslide events mapping, even in shadowed areas.
Other authors have developed algorithms that exploit the slow regrowth
of vegetation in landslide detachment areas to automatically detect
landslides using NDVI time series data following the triggering event.
However, this approach faces challenges with cloud coverage, making it
unsuitable for rapid landslide mapping after a catastrophic event (Mil-
ledge et al., 2022). The recent advancement of new machine learning
techniques, in particular, the algorithms that use the convolutional
Neural Network (CNN) and Fully Convolutional Network (FCN) for
image segmentation (Ghorbanzadeh et al., 2021), has facilitated the
analysis of extensive datasets, enabling the observation of NDVI trends
over extended periods and facilitating the identication of various fac-
tors inuencing these uctuations, such as seasonality, landslide trig-
gers, and human activities (Doan et al., 2023).This work presents a two-
phase methodology aimed at the low-cost mapping of shallow land-
slides: the unsuPervised shAllow laNdslide rapiD mApping - PANDA
method. The rst step (PANDA-E) is applied in the emergency phase as a
survey management tool. It is based on an improved version of the
procedure proposed in (Notti et al., 2023b), to obtain a rst assessment
map of the area affected by PL in the aftermath of the rainfall event,
immediately available for survey eld operations and risk management.
The second phase is post-emergency mapping (PANDA-PE), in which the
PL inventory is progressively improved with the availability of new
images and ancillary data (e.g., high-resolution DTMs, eld survey in-
formation, UAV surveys), of which acquisition is usually arranged in the
stages immediately following the event. The availability of progressively
more detailed and higher-resolution data, allows the initial mapping to
be rened, identifying the false positive and resolving the eventual gap
of cloud cover of the rst map. The PANDA-PE aims to obtain a reliable
inventory tool for the subsequent phase of validation, risk mitigation
and action planning.
The proposed methodology was applied both in the emergency and
post-emergency phases in a portion of northeastern Apennine chain
(Central Italy) to map the dramatic consequences of two severe rainfall
events that occurred in May 2023. The maps obtained were also vali-
dated by a third-party (Ferrario, 2023) manual inventory made on high-
resolution satellite images.
2. Study area
2.1. Geological and geomorphological settings
The study area covers the central-eastern sector of the Tuscan-
Emilian Apennines, within the provinces of Bologna, Ravenna, and
Forlì-Cesena, Emilia-Romagna Region, and a small portion of the Tus-
cany Region (Florence Province) (Fig. 1). The area of interest (AOI) was
identied as the mountainous and hilly regions predominantly impacted
by the dual rainfall occurrences in May 2023 and was pinpointed with a
particular focus on the sector where the second event (16–17 May)
recorded precipitation exceeding 150 mm or the cumulative rainfall of
two events surpassed 250 mm (Fig. 2b).
This territory pertaining to the Tuscan-Emilian Apennine segment is
mainly characterized by a mountainous to hilly morphology, with ele-
vations ranging from 1300 m a.s.l. in the Apennine portion to 50 m a.s.l.
towards the Po Plain. This Apennine sector, a fold and thrust belt, is
mainly constituted by turbiditic deposits (ysch) with the alternation of
mudstone and massive rocks as sandstones and calcarenites (Pini, 1999)
(Fig. 1b). Considering the high-medium slope values and the presence of
lithologies prone to erosion, this territory is extremely susceptible to
slope instabilities. According to the IFFI catalogue (Trigila et al., 2008)
the most widespread landslide typologies are earth ows, translational/
rotational slides and complex landslides, generally triggered by snow
melting and moderate but exceptionally prolonged (even up to 6
months) periods of rainfalls. On the other hand, shallow landslides and
debris ows are generally less prevalent; however, their occurrence
rapidly increases during high-intensity rainfalls or by prolonged low-
intensity rainfalls (Benedetti et al., 2005; Ibsen and Casagli, (2004);
Martelloni et al., 2013; Martina et al., 2010). This territory generally
displays a high degree of urbanization characterized by numerous small
inhabited areas scattered in the hilly zones connected with the main
urban centers by a dense road and rail network. Consequently, geo-
hazard related to landslides is one of the major issues affecting the entire
Region.
2.2. The rainfall events
During May 2023, an extended portion of the Central Apennine,
specically the Emilian Apennine side and the Po Plain of the Emilia-
Romagna Region, was affected by two consecutive rainstorms that
occurred in less than twenty days: i) 1-3th May 2023 weather event, and
ii) 16–17 May 2023 weather event (Brath et al., 2023).
In the rst event, a low-pressure area generated a convergence of
moist air, interacting with the Apennine chain in the central-eastern
sector of the Region (i.e., provinces of Bologna, Ravenna and Forlì-
Cesena), and a small portion of the Tuscany Region included in the
Adriatic river basins (Reno and Lamone). Referring to the spatial grid-
ded data (5 km of spatial resolution) of rainfall available on the SIMC
platform (https://dati-simc.arpae.it/opendata/erg5v2/timeseries/ma
ppa.html), this rst event led to rainfall accumulations in the hilly
area even exceeding 200 mm (Arpae-SIMC, 2023a), in the central area
(cells ID 1628, ID 1505 Fig. 2a) and 150 mm in the western limit (cell ID
1302) calculated on the whole event (about 40 continuous hours). The
average hourly intensity ranged from 3 to 6 mm/h, with some peaks up
to 10 mm/h (Fig. 2a). This event was weaker in the eastern part with
cumulated rainfall of 75 mm (cell ID 2031, Fig. 2b). Other minor rainfall
events contributed up to 50 mm of rainfall between the two events in the
same area.
In the second event, a cyclone transiting the Mediterranean affected
approximately the same area impacted by the previous event, with peak
rainfall slightly shifted to the east (i.e., provinces of Bologna and Forlì-
Cesena). Most of the precipitations were concentrated in the central-
eastern hills and foothills areas, with accumulations of ~ 250 mm in
35 h and peaks recorded during May 16 exceeding 200 mm (Arpae-
SIMC, 2023b; Brath et al., 2023). The intensity of the second event was
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
3
also slightly higher than that of the rst event, especially in the central-
eastern part (7 mm/h), where some peaks reached 20 mm/h, especially
in cell ID 2031(Fig. 2a).
These amounts of precipitation poured over an area close to satu-
ration (according to the soil water index – SWI content available on
Copernicus Global Land Service https://land.copernicus.eu/global/pro
ducts/swi) by the intense rainfall events in the preceding days, mak-
ing the soil unable to absorb part of the precipitations and triggering
many shallow and deeper landslides.
3. Materials and methods
Following the methodological ow chart proposed in Fig. 3, we
dened two main phases of PANDA approach for shallow landslide
mapping:
•Emergency mapping (PANDA-E): It aims to create a rst assessment
map helpful for support in eld surveys to estimate the most affected
area in terms of landslide density and a preliminary estimate of
damage. In this step, rapidity has priority on accuracy. The emer-
gency mapping is based only on a semi-automatic methodology
derived from Notti et al., (2023) that uses Sentinel-2 satellites.
•Post-Emergency mapping (PANDA-PE): This step aims to improve
the rst asset mapping and to make a more rened mapping of
shallow landslides. This step could improve the semi-automatic
methodology using the land use change lter and enhanced
geomorphological lter. Here, high-resolution images from aerial or
satellite can be used for validation. The mapping aims to create an
accurate event map for research, land-use planning, or susceptibility
studies.
It is not possible to quantify the time of the emergency phase, which
may last from a few days up to several months, depending on many
factors. For the Emilia case, we considered the emergency phase to have
ended when the eld survey for the rst assessment and residual risk
evaluation ended, approximately 40 days after the event.
3.1. The PANDA-E mapping steps and parameters
Performing a shallow landslide map over large areas during an
emergency context, available immediately after a rainstorm event, re-
quires a rapid and unsupervised mapping of the affected area. We
implemented PANDA to obtain a prompt map of the ground effects
based on free-cost optical satellite images.
The emergency methodology is based on those proposed by (Notti
et al., 2023b), a GIS-based and user-friendly approach to map PLs, based
on the difference between pre- and post-event Normalized Difference
Vegetation Index (NDVI) obtained from satellite images, e.g., free
Sentinel-2 dataset, available on the Copernicus Dataspace Service (https
://dataspace.copernicus.eu/) and geomorphological ltering. The vari-
ation in the NDVI (NDVI
var
), mainly linked to the processes that occur
during the rainstorm, allows the functional detection of surface changes
and signs of PLs. Considering the emergency context, the need for im-
ages immediately following the event may result in some constraints in
the post-event image selection. After a rainstorm, the availability of
entirely could-free images can be challenging to achieve. However, if the
cloud percentage is below an empirical threshold that allows an almost
complete mapping (e.g., <20 % of the study area), it is possible to
exclude from PL mapping the areas covered by the clouds using the mask
available on the Sentinel-2 dataset (Table 1), joined with a manual
renement to include cloud shadow. In our case, the rst image was
acquired on 23–05-2023, a week after the event, compared with the
previous year image (13/05/2022). It is worth noting that The Google
Earth Engine (GEE) platform has undergone advancements in cloud
detection capabilities, enabling precise identication of both clouds and
shadows (https://developers.google.com/earth-engine/tutorials/comm
unity/sentinel-2-s2cloudless).
Then, we combine two Boolean lters based on the NDVI
var
and slope
gradient. The rst one selected all the areas with an NDVI
var
below a
certain threshold (based on visual evaluation of NDVI
var
pattern) that
Fig. 1. (a) Location of the AOI that identies the sector most affected by the rainstorm events that occurred in May 2023, and (b) simplied lithological map based
on 1:500.000 scale lithological map of Italy (MASE, 2009).
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
4
identify a PL. The second is a geomorphological lter based on slope
gradient (derived from digital terrain model (DTM)). This lter is based
on an empirical threshold (based on local geomorphology), below which
an NDVI change is most likely unrelated to a shallow landslide occur-
rence. In the Boolean raster, the pixels where both thresholds are
satised have a value of 1, and such pixels are extracted and converted
into vectors: i.e. the PL polygons. In our study area, we used the
thresholds NDVI
var
<=-0.3 and slope >15◦. Such thresholds are
empirically based on the pattern NDVI
var,
previous experience, the local
geomorphological conditions, and literature (Ghorbanzadeh et al.,
2022a; Notti et al., 2023b); such thresholds could be changed in iterative
procedure in order to obtain the most likely ground truth scenario. In
this case, we used a conservative NDVI
var
threshold for the problem
related to cloud coverage.
In addition, the polygons have been intersected with the
hydrographic network (polygon or buffered layer geometry) to exclude
bank erosion processes or sediment deposition along the river system.
The ltered polygons exemplify the area where a shallow landslide is
most probably located. Consequently, the PL density map allows us to
identify the most affected areas to focus emergency activities.
3.2. The PANDA-PE steps and parameters
After the preliminary release, in the post-emergency phase, it is
possible to improve the PL database by acquiring new images and
ancillary data. The aim is to obtain a more accurate mapping of the event
impact and its consequences. Among the several improvements that can
be made, the rst is to use an utterly cloud-free image (cloud cover <3
%) to have information on the whole AOI. In this case study, we found
the rst cloud-free image on 17 July 2023, compared with a pre-event
Fig. 2. Rainfall data of May 2023 events. a) Hourly (left Y-axis) and cumulated (right Y-axis) rainfall data (computed respectively for the May 1-3rd and the May 16-
17th rainfall events. The precipitation time series utilized have been extrapolated from the grid of the SIMC platform (https://dati-simc.arpae.it/opendata
/erg5v2/timeseries/mappa.html) over 4 representative cells. b) Location of selected cells and the rainfall contour lines used to delimit the AOI.
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
5
image of 22 July 2022.
The false positive land use change (LUC) lter based on NDVI
var
is a
second improvement. The change of land use, and vegetation activity,
that intercurred between the pre-event images and the event is one of
the leading causes of false positive (e.g., a forest cut or change in crop
cultivation). If available, a cloud-free image acquired slightly (i.e. in the
same agricultural cycle) before the event can be compared with a syn-
chronous image from the previous year. Leveraging the NDVI difference,
it is possible to calculate a pre-event NDVI
var
. If the mean LUC NDVI
var
in
the PL is below an empirical threshold (<-0.15, in our AOI), we classify
it as a possible false positive. In our case study, we could make this
improvement only on a portion of the study area because the closest LUC
image to the event (22–02-2023) features partial snow cover. The
NDVI
var
LUC lter and cloud-free post-event images allowed us to use a
less conservative threshold of NDVI
var
for PANDA-PE, which was xed at
−0.15.
Finally, the geomorphological lter could be improved from the
emergency phase using a more detailed hydrographic network or slope-
DTM. In our case study, a few months after the events, a 5 m DTM was
available on the whole AOI and allowed us to improve the accuracy of
the slope lter.
3.3. Comparison and validation with a manual inventory
In the aftermath of the May 2023 rainstorm, the authors conducted
eld surveys and rapid reports to aid regional and local authorities in
identifying and mapping critical landslides. This allowed for testing the
PANDA-E method, although on a limited scale, just a few days after the
event. Moreover, we used data from geolocated reports, videos, and
photos on the web to validate the rst PANDA-E map locally.
Then, during the post-emergency phase, the results obtained from
both methodology phases (i.e. PANDA-E and PANDA-PE) were validated
by exploiting the manual landslide (ML) inventory made by Ferrario,
2023. This manual mapping, available on the Zenodo data repository, is
based on Planet Scope satellite images (about 3 m of spatial resolution)
published about 40 days after the event, overlapping most of our AOI
(Ferrario and Livio, 2024). We also considered only the ML and PL with
a median slope ≥15◦to make a homogeneous comparison and discard
the process more related to the river erosion. We used the same vali-
dation schema described in (Notti et al., 2023b) where ve intersection
types are dened: True Positive (i.e. PL and ML complete overlapping),
Partial Positive (i.e. PL that intersects a ML), False Positive (i.e. PL that
not intersect or overlap a ML), Partial Detection (i.e. ML that intersect a
PL) and False Negative (i.e. ML that not intersect or overlap a PL). We
also calculate the statistics like the Detection Rate (DR) =(TP +PD)/
(TP +PD +FN) and the False Positive Rate (FPR) =FP / (FP +TP +PP)
and the standard performance of accuracy: precision (P) =TP / (TP +
FP), recall R =TP / (TP +FN) and F
1
score F
1
=2TP / (2TP +FP +FN).
For the Emilia-Romagna region, we also intersected the validation test
with the land use dataset to evaluate the effect of land cover on the
performance.
Additional local validation was made using the very-high-resolution
(0.2 m) post-event images acquired by CGR S.p.A and available on the
Emilia-Romagna Region web portal, specically acquired for the 2023
rainfall events (https://geoportale.regione.emilia-romagna.it/approfon
dimenti/emergenza-maggio-23/emergenza-rer-maggio-2023-servizi).
4. Results
4.1. Application of PANDA-E
The PANDA method, used in the emergency related to the May 2023
Emilia-Romagna rainstorms, rapidly mapped potential ground effects in
the aftermath of the second event. The unsupervised mapping of shallow
landslides provided an overall territorial framework, that can be very
useful for a rst evaluation of the impact of the occurred event, the
Fig. 3. The two-phase owchart of PANDA methodology proposed for shallow
landslide mapping.
Table 1
Used data for the shallow landslide mapping with PANDA-E and PANDA-PE
steps related to the Emilia Romagna Rainstorm events.
PANDA-E PANDA-PE
Sensor and
products
pre-
event
post-
event
pre-
event
post-
event
NDVI
var
for
PL
detection
Sentinel-2, Bottom-
of-Atmosphere
reectance in
cartographic
geometry (L2A). B4,
NIR (10 m)
13/05/
2022
23/05/
2023
22/
07/
2022
17/
07/
2023
Cloud and shadow
mask
Cloud probability
mask available on
the Sentinel-2
quality assessment
folder (layer:
MSK_CLDPRB_20m)
Cloud cover <3
% cloud mask
not necessary
NDVI
var
lter
for false
positive
related
land-use
change
(LUC)
Sentinel-2, Bottom-
of-Atmosphere
reectance in
cartographic
geometry (L2A). B4,
NIR (10 m)
Not used pre-
LUC
post-
LUC
17/
02/
2022
22/
02/
2023
Filtering Slope derived from
DTM
10 m DTM TIN Italy 5 m DTM Emila-
Region
Hydrographic
network
Available shapele from Geoportale of
Emilia-Romagna Region (https://geoport
ale.regione.emilia-romagna.it/download)
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
6
distribution of damage and the denition of priorities in the scheduling
of eld activities (Notti et al., 2023a). The mapping of ground effects
performed within the investigated areas revealed about 30,000 PLs
(covering an area of 20 km
2
, 0.5 % of the whole AOI), among them, a
few thousand are most probably related to hydrological processes.
About 10 % of PLs intersect the road network within 10 m. This map
partially suffers from cloud cover affecting about 8 % of AOI (related to
the available Sentinel-2 image acquired on 23rd May 2023), even if
mainly in the southern portion of the observed area, close to the
Adriatic-Tyrrhenian drainage divide, less affected by rainfall. Never-
theless, it is immediately possible to detect the sectors most affected by
PL, on which to focus risk assessment and mitigation activities. Fig. 4a
shows the Kernel density distribution of the PL centroid; the density was
computed on a 100 m resolution grid with a search radius of km and
uniform interpolation. It is possible to see that the zone featuring higher
values (about 150 km
2
) is nearby Modigliana municipality (more than
3000 PL) and matches rainfall data and ground evidence reported day-
by-day by eld operators during the emergency. The most affected
municipalities, in terms of PL abundance and density, are reported in
Table 2.
The histogram in Fig. 5 a shows the distribution of PL areas. Such
distribution agrees with the power law distribution of shallow landslide
areas (Bellugi et al., 2021; Guzzetti et al., 2002). Approximately 80 % of
potential landslides (PL) exhibit sizes ranging from 100 m
2
(equivalent
to one pixel of Sentinel-2) to 1000 m
2
, as depicted in Fig. 5b. Some
outlier values up to 1.2⋅105 m2 (Fig. 5 inset plot) are probably related to
false positives due to the residual unltered cloud cover or land use and
vegetation changes. Due to the spatial resolution of both Sentinel −2 and
DTM, there are no PLs smaller than 100 m
2
.
Because the rst aim of the PANDA methodology is to rapidly detect
the most affected area and the preliminary location of PL, it was almost
impossible to validate it with a manual inventory, high-resolution
images or extensive eld surveys. However, leveraging the collected
geolocated information and videos available on the web (e.g., national
newspapers or Localteam website https://www.localteam.it/), it was
possible to carry out a local ground-truth comparison of PANDA results
(Fig. 6 C-D). Moreover, during the eld survey operated by our research
group and started a few days after the event, several positive feedbacks
were observed. For instance, Fig. 6 A-B shows a high correspondence
between extrapolated PL polygons and phenomena observed.
4.2. The improved PANDA-PE inventory
The improved PANDA-PE version was made in September 2023
(Fig. 7). The better quality of the post-event image, combined with the
LUC lter and higher resolution DTM, allowed us to obtain a complete
database of PL with about 54,000 polygons corresponding to an area of
Fig. 4. The Kernel density map made with PL centroid (parameter: cell size 100 m; search radius 2.5 km; uniform interpolation; scaled) overlapped with the mask of
cloud and shadow areas. The location of ground validation (Fig. 6) is also added to the same map.
Table 2
PL centroid density in the top 10 most affected municipalities.
Municipality Inhabitants (dati.
istat.it)
PL PL Density
(PL/km
2
)
area
affected (%)
Dovadola 1573 1708 43.6 3.17
Modigliana 4288 3717 36.6 2.32
Fontanelice 1913 907 24.4 1.87
Casola Valsenio 2498 2044 24.1 1.66
Predappio 6296 2010 22 1.38
Roncofreddo 3422 1125 22 1.31
Brisighella 7186 3551 18.4 1.12
Borgo
Tossignano
3198 469 15.9 1.17
Civitella di
Romagna
3639 1821 15.4 0.83
Mercato
Saraceno
6797 1489 14.9 1.2
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
7
about 40 km
2
. The intersection with the hydrographic network allowed
the classication of about 2700 polygons as a hydrological process,
while the LUC NDVI
var
lter allowed the detection of about 1600 false
positives (Fig. 7a). The relative Kernel density of the PL classied as
shallow landside (Fig. 7b) shows a similar distribution of the PANDA
inventory. In Fig. 7b, it is possible to observe the low density around
rainfall cell ID 1505 and an isolated peak of PL density in correspon-
dence with cell ID 2031, despite a similar amount of rainfall (Fig. 2).
Fig. 8 shows, over a test area, some details on the improvements of
the PANDA-PE version. The higher resolution of 5 m DTM and derived
slope (Fig. 8b) better identies the critical slope than the previous 10 m
resolution (Fig. 8a). Then, the NDVI
var
made with the post-event cloud-
free image (2023–07-17) (Fig. 8c) and corrected with LUC NDVI
var
and
hydrographic network layer (Fig. 8d) allowed the classication of the
PL. The post-event high-resolution image of the Emilia-Romagna region
(Fig. 8e) conrmed the validity of classication; in particular, the PLs
classied as false positive with LUC NDVI
var
corresponds to a change in
the cultivated eld.
4.3. PANDA-PE PL density and distribution and correlation with rainfall,
slope lithology and land use
Using the PANDA inventories, we made some rapid evaluations on
the main factors that are commonly used in literature (Reichenbach
et al., 2018) for shallow landslide distribution analysis or susceptibility
modelling: the trigger factor (rainfall) and predisposal factors (slope,
lithology and land use). The most interesting results are related to the
triggering factor, the rainfall. We compared the PL density with rainfall
distribution in Fig. 9 for the rst event (panels a and a’), the second
event (panels b and b’), and the summarized rainfall (panels c and c’).
The rainfall and landslide density are calculated for each cell of the
rainfall grid of the Emilia-Romagna region (SIMC platform https://dati
-simc.arpae.it/opendata/erg5v2/timeseries/mappa.html). Panel b’ of
Fig. 9 shows a box plot of the PL density for each rainfall class of the
second event (16–17th May 2023). It is important to note the relation-
ship between cumulated rainfall and density, notably where the second
event reached values up to 150 mm (300–400 mm, considering the
cumulated precipitation of the two events). Conversely, the relation is
less evident with the cumulated rainfall of the rst event (Panel a’ of
Fig. 9). Most probably, while the rst rainfall saturated the soil (https://
land.copernicus.eu/global/products/swi) the second (also with high
hourly intensity) triggered most of the processing. Moreover, the ante-
cedent rainfall (see paragraph 2.2) also affected shallow landslide dis-
tribution. Considering the intensity-duration thresholds (Guzzetti et al.,
2008; Mondini et al., 2023), both single and the combination of these
two are above the trigger thresholds worldwide found. Compared to
nearby historical events, the total rainfall of these two events (up to 500
mm and average intensity of 6 mm/h) is similar to the 2013 Marche
events, while it is very different from the event that hit the nearby
Marche region in September 2022 (Donnini et al., 2023), which showed
much higher intensity and short duration.
As already known, the slope gradient is one of the main factors
related to landslide distribution (Lee and Min, 2001). In our case, the
landslide density related with the slope angle is a partial analysis, as we
consider only a slope above 15◦to extract the PL. Considering this there
is a linear relationship between the percentage of the area affected by
landslide and the slope angle, which ranges from 1 % for the with a slope
between 15◦and 20◦up to 4 % for the area with slope gradient class 40◦-
45◦.
The lithology also played a role in the distribution of shallow land-
slides density and distribution, with higher concentration in the arena-
ceous turbiditic unit, especially in the “Marnoso-Areaneacea
formations”. In contrast, shale and chaotic formations show less density
of landslides. Considering the percentage of the area affected by land-
slides, the Marnoso-Areaneacea formations reach a peak of 20 Kernel
Density against the 4 KD of chaotic shale. The lithology has a strong
inuence on the slope degree distribution. The lithology also explains
some incongruency with rainfall distribution: for instance, the low PL
density (KD 8) of cell ID 1505 with cumulated rainfall of 350 mm located
in chaotic shale, by contrast, the higher PL density (KD 15) of cell ID
2031 in arenaceous turbiditic formation with low rainfall (170 mm). The
role of lithology in landslide density was already documented in the
nearby Marche 2022 event (Donnini et al., 2023).
The land use (based on the 2020 Land cover database of Emilia-
Romagna Region, 2023) already inuences the landslide density dis-
tribution but less reliably compared to the previous factors because the
methodology introduces possible bias in this comparison. For instance,
the higher PL density on oak-hornbeam and chestnut forest types is
related to the fact that this type of land use is located mainly on the
arenaceous turbiditic formation and in the area most affected by rainfall.
Meanwhile, beech forests in the southern limit of the study area, which
are less affected by rainfall, show low PL density. At the same time, the
distribution of PL shallow landslides may suffer an underestimation
detection bias in the area where the vegetation cover is already lower
(vineyard, badland), or an overestimation where human activity in-
terferes with vegetation (cultivated land), as already reported in previ-
ous work (Notti et al., 2023a) and in paragraph 5.2.
5. Discussions
5.1. PANDA-E mapping time and accuracy
During the emergency phase, like in the Emilia-Romagna event, it is
crucial to quickly establish an overall framework for the most affected
areas. An important point to stress here is that, due to the relevant
emergency occurred, the rapid mapping refers to the ground variation.
The obtained product does not correspond to a geomorphological
Fig. 5. A) histogram of pl area frequency distribution up to 1000 m2. Inset plot
shows the whole distribution, including the outliers. b) Cumulative distribution,
up to 10,000 m2.
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
8
landslide inventory in the strict sense because it is the output of a semi-
automatic feature extraction based on middle-resolution images.
Consequently, the proposed approach may generate potential false-
positive cases mainly related to land use changes during the observed
period (e.g., cultivation areas) or the effect of cloud cover. The semi-
automatic cloud lter allows the partial reduction of false positives.
However, local residual effects related to cloud coverage remain at this
early stage. Another limitation is related to the spatial resolution and
Fig. 6. Comparison between the potential landslide polygons overlapped to Google Earth Satellite images and the on-eld observations and web-based images or
videos. Montiano municipality, Forlì-Cesena Province, affected by a rotational sliding evolved in an earth ow: A) PL overlapped to Google Earth pre-event image
(13/09/2022 Credit ©Maxar, 2022); B) Post-event ground photo taken a few days after the events (Credit Davide Notti). SP306 close to Casola Valsenio hamlet,
Ravenna, affected by several shallow landslides along the upstream slope: C) PL overlapped to Google Earth pre-event image (13/09/2022 Credit ©Maxar, 2022); D)
Same PL on the frame from Localteam video ().
Source: https://www.localteam.it/video/grandi-frane-a-casola-valsenio-la-collina-appare-sventrata
Fig. 7. A) the centroids of improved panda-pe pl classied in a shallow landslide (shl), hydrological process (hyp) and false positive related land-use changes (fp); b)
relative kernel density (parameter: cell size 100 m; search radius 2.5 km; uniform interpolation; output value scaling: scaled) of PL classied shallow landside,
associated to the selected cells used in Fig. 3 for analysis of rainfall events.
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
9
accuracy of the DTM, which hamper the slope ltering. Despite the
above limits, the obtained map provides a tool, freely and immediately
available after a catastrophic event, useful for the eld survey opera-
tions devoted to residual risk assessment. The use of Sentinel-2 satellite
images allows to a totally free-of-charge approach, ensuring a cost-
effective and user-friendly downloading and management with a free
cloud computing platform as Google Earth Engine. Nevertheless,
PANDA is structured to operate with any optical spaceborne imagery,
regardless of revisiting time or resolution. However, the exploitation of
satellite images with shorter revisit times, like PlanetScope, is, in any
case, inherently constrained by cloud coverage, particularly after a
rainstorm event. Surely, higher-resolution sensors improve detection
accuracy in terms of landslide boundaries and extent, but generally have
costs not affordable for several users, misaligning with the objectives of
PANDA. Leveraging on the “Open Data Program” (https://www.maxar.
com/open-data), a comparative analysis of potential shallow landslides
mapped using Sentinel-2 and Maxar data (0.5 m resolution) has been
carried out. The availability of Maxar imagery, limited to the ooded
areas, did not allow extensive analysis over the hilly area, which was
only sporadically covered by the available imagery. The comparison
displayed, as the resolution increases, an improved accuracy and quality
of mapping, but, excluding those exceptional cases of data availability,
they commonly entail a related cost. Conversely, the PANDA method-
ology aims to be a free and user-friendly tool for timely use immediately
after a calamitous event, to identify occurred landslides in terms of
number and density. While the reduced time to generate the ground
effects map might compromise the accuracy of landslide outlines
compared to high-resolution imagery, it still offers valuable support for
initial emergency response actions. The rst draft map obtained was
generated on 26th May 2023, just three days after the Sentinel-2
acquisition and after one week, the map was published on the Zenodo
repository. The need to have such an instrument within a few days,
immediately after a rainstorm, obviously precludes the possibility of a
large-scale data validation with a manual landslide inventory. A manual
mapping on high-resolution images or by eld survey operations re-
quires, on such a large area, several weeks/months of work or many
operators (Galli et al., 2007; Milledge et al., 2022), an incompatible
timing during the emergency phase. In the case of the May 2023 events,
the rst manual inventory was produced about one month after the
event and released in vector format in the Zenodo repository (Ferrario,
2023), and a technical report, including a landslide map, was produced
about six months later (Brath et al., 2023). Directly comparing the ac-
curacy, time, and cost of this study methodology to others is complicated
or even biased due to variations in study area size and event intensity, as
well as the aim of the nal users of the method. However, some insights
can be gleaned. For example, Mondini et al., (2011) spent ve days
creating and validating their inventory, with two experts in remote
sensing and geomorphology, using PCA, covering an area of 9.4 km
2
.
PANDA could potentially reduce processing and validation time by two
days for larger areas, and it does not need remote sensing expertise, but
at the cost of lower accuracy. In agreement with PANDA also Mondini
et al., (2013) found that the use of a limited number of morphometric
parameters (e.g., slope and hydrographic network) reduces the
complexity of the classication framework and the computation time.
Fig. 8. PANDA-PE PL overlapped to: a) 10 m spatial resolution slope-DTM; b) 5 m spatial resolution slope-DTM. c) PANDA-PE PL overlapped to cloud-free NDVI
var;
d) Classied PANDA-PE - PL overlapped to LUC NDVI
var;
and hydrographic network, e) Classied PL overlapped to post-event high-resolution images.
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
10
(Li et al., 2016), compared simple change detection with more sophis-
ticated techniques that improve results. In this work, using geomor-
phological lters and LUC in the post-emergency phase could improve
the performance, avoiding expertise in processing. Similarly, Lu et al.,
(2019) demonstrated the effectiveness of PCA in small forested areas
using Landsat-8 and Sentinel-2 data compared to NDVI variation.
However, their study area was small and required remote sensing
expertise. Compared to the other approaches, PANDA provides some
important advantages, ensuring a methodology capables to process wide
areas in a short time (e.g., two days for Sentinel-2 scene), with a reduced
number of input data, commonly available worldwide, that are needed
to run the model (NDVI, cloud cover, slope, hydrographic network),
through a GIS-based procedure completely cost-free and suitable for
users with different degree of experience.
5.2. Comparison and validation with a manual inventory. PANDA-E vs
PANDA-PE
As shown in Notti et al., (2023b) the PL dataset could be validated
with manual-based landslide inventory (ML) to evaluate its accuracy. In
the emergency phase, nding data for validation is almost impossible,
except from local cases such as reports or geolocated video (Fig. 6). In
the weeks following the event, it could be possible to obtain or create a
manual inventory based on different sources (e.g., satellite, aerial,
ground-based).
The PANDA-E PL inventories made during the Emergency (v 1.0)
phase and the improved PANDA-PE (v 2.0) version were validated using
the manual inventory made by Ferrario, (2023), available on the Zenodo
data repository. The overall results for the validation study area are
resumed in Table 3. It is possible to observe that the improved PL in-
ventory (PANDA-PE) increased the surface of detected shallow land-
slides about two times (7.6 to 13.4 km
2
) with a slight decrease of False
Positive (11.3 to 10.9 km
2
) more evident in percentage (24 % to 18 %).
The overall performances of PL inventories calculated as reported in
paragraph 3.3 are compared in Table 4. Fig. 10 shows, on a test area, a
comparison between the validation of PANDA-E (Fig. 10a) and the
PANDA-PE (Fig. 10b). It is possible to observe the improvement of the
PANDA-PE PL version, considering the area: the overall F
1
-score
increased from 0.40 to 0.59, which is a satisfactory result, considering
the complex land use. Considering the validation test only for broadleaf
forest land-use type, the F
1
-score reached a value of 0.72 (Fig. 10c),
similar to studies that use high-resolution images (Meena et al., 2023) or
tested in tropical forests land-use (Ghorbanzadeh et al., 2022a), or with
more complex multi-sensor analysis (Lu et al., 2019) On the other hand,
the F
1
value decreased to 0.15 on cultivated land (Fig. 10e) also urban
areas and badlands show lower performance. The effect of land use on
Fig. 9. Maps of interpolated rainfall distribution and associated bar plots of PL Kernel Density for rainfall class respectively for: the 1 −3rd May rainfall event (a and
a’), 16 – 17th May rainfall event (b and b’) and the sum of the two events (c and c’). The Kernel Density and rainfall are computed on the rainfall grid of SIMC
platform (https://dati-simc.arpae.it/opendata/erg5v2/timeseries/mappa.html).
Table 3
Intersection cases summary with ML inventory (Ferrario, 2023) by polygon
count and total area for the whole AOI.
PL Inventory version PANDA-E PANDA-PE
Intersection case Polygon N Area Km
2
Polygon N Area Km
2
True Positive (TP) 20,230 7.6 30,224 13.4
Partial Positive (PP) 16,534 3.0 24,598 15.3
False Positive (FP) 11,329 11.3 28,871 10.9
Partial Detection (PD) 14,704 13.8 24,449 11.5
False Negative (FN) 24,482 11.4 13,750 8.0
Total 87,279 47.1 121,892 59.1
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
11
landslide detection performance has also been observed in other studies
(Dias et al., 2023; Mondini et al., 2011). In areas with low vegetation,
the use of stereoscopic or 3D data is found to be the most effective so-
lution (Fiorucci et al., 2018). A possible solution could be to use SAR
amplitude images; however, this requires a complex process, and they
could be partially affected by the geometry distortion related to topog-
raphy. For instance, to map shallow landslides in a forested area in
Hokkaido, Japan, Nava et al., (2022) obtained a score of 0.61 using
INSAR with U-NET methods, while from the same research group,
Meena et al. (2023) in the same area, using optical data of Planet and U-
Net and f1 score of 0.81. Moreover In the same area, using SAR, it was
noticed the effect of land use on performance detection (Jung and Yun,
2020; Mondini et al., 2021).
It was possible to verify that most FNs are related to shallow
landslides smaller than 100 m
2
, an intrinsic limit related to Sentinel-2
spatial resolution (10 m) and the badland area (Fig. 10d), especially in
the marl-shale formation. As mentioned before, FPs correspond to land
use changes not detected by the LUC lter because they occurred later in
the image used for LUC NDVI
var
(2023–02-22) or were located in an area
covered by snow. By random checking on the very high-resolution im-
ages of CGR, we also noted that in several cases, some FP were correctly
detected, most probably because of some cloud cover, and lower reso-
lution (3 m vs 0.2 m) of the Planet images used for ML inventory. On the
other hand, as reported by (Ferrario, 2023) manual inventory could also
be affected by the false negative and false positive errors that inuence
the comparison. Finally, validation using high-resolution images also
allowed the detection of some cases in which the PANDA-E performed
better than PANDA-PE, where, for instance, the accumulation material
was removed (along the main roads) or was interested in the regrowth of
herbaceous vegetation occurring from May to July (e.g., the small
landslides in or the north sector of Fig. 10a and Fig. 10b).
6. Conclusions
The primary purpose of the PANDA is to serve as an initial and swift
emergency assessment map, allowing users to identify the most heavily
impacted areas. It has shown reliability, particularly in quickly mapping
large areas with cost-free inputs, leveraging a user-friendly processing.
Table 4
Performance of PL inventories considering the area compared to ML inventory
(Ferrario, 2023).
Parameters PANDA-E PANDA-PE
DR 65 % 76 %
FPR 52 % 28 %
Precision 40 % 55 %
Recall 40 % 62 %
F
1
-Score 0.40 0.59
Fig. 10. A detail over a test area of the validation of PL Inventories: a) PANDA-E; b) PANDA-PE improved inventory. The performance of PANDA-PE mapping
depends on land use: c) broadleaf forest shows best performance; d) shallow landslides on badlands are rarely detected (False Negative); e) cultivated land still
produces False Positive, despite ltering. The post-event high-resolution post-event images created by CGR spa (© AGEA - ALL RIGHTS RESERVED).
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
12
PANDA in emergency phase, applied to the May 2023 rainfall events of
northeastern Apennine, allowed us to map, in a few days after the event
and at a low-cost, tens of thousands of PL over a wide area (>4000 km2)
and identied the most affected sector (about 150 km2) where PL
density reached values up to 50 km2.The whole map was immediately
shared on a public data repository to be used and improved by the expert
community. The obtained PL inventory was tested directly during the
emergency eld surveys with positive feedback. In contrast, emergency
rapid mapping is limited by the possibility of having a completely cloud-
free image, increasing the probability of false positives. Such limitations
during the post-emergency mapping (PANDA-PE) can be partially solved
by improving PL using a cloud-free image, a high-resolution slope-DTM
and a land use change lter also based on NDVI variation, thus
decreasing False Positive and increasing inventory accuracy as
conrmed by validation employing an independent manual inventory
(F1-score improved 0.4 to 0.59). The performance of the semi-automatic
mapping based on NDVI is related to land use, working well on vege-
tated areas, and worse on bare soils and cultivated land, and this should
be considered in emergency and post-emergency mapping. Potential
enhancements to the PANDA methodology may involve rening the
ltering capabilities to address cloud cover and variations in land use.
Moreover, the conversion of the proposed methodologies into a software
would make PANDA a user-friendly tool accessible for everyone. Over-
all, this application conrmed the suitability of the PANDA methodol-
ogy for shallow landslide detection in several contexts worldwide. The
enhanced PANDA-PE methodology further renes landslide inventory
accuracy, paving the way for improved disaster management strategies.
CRediT authorship contribution statement
Davide Notti: Writing – original draft, Formal analysis, Data cura-
tion. Martina Cignetti: Writing – original draft, Methodology, Formal
analysis, Data curation. Danilo Godone: Writing – review & editing,
Validation, Data curation. Davide Cardone: Writing – original draft,
Formal analysis, Data curation. Daniele Giordan: Writing – review &
editing.
Declaration of competing interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Data availability
Data are available at the link https://zenodo.org/records/10706716.
Acknowledgements
We thank the two anonymous reviewers for their comments and
suggestions that greatly improved the quality of the paper.
References
Arpae-SIMC, L’evento meteo-idrogeologico del 1–4 Maggio 2023 (technical analysis
related to the weather, hydrogeological and hydraulic event) 2023 Agenzia
Regionale per la Protezione Ambientale Emilia-Romagna.
Arpae-SIMC, L’evento meteo-idrogeologico e idraulico del 16–18 Maggio 2023 (technical
analysis related to the weather, hydrogeological and hydraulic event) 2023 Agenzia
Regionale per la Protezione Ambientale Emilia-Romagna.
Bellugi, D.G., Milledge, D.G., Cuffey, K.M., Dietrich, W.E., Larsen, L.G., 2021. Controls on
the Size Distributions of Shallow Landslides. Proceedings of the National Academy of
Sciences. https://doi.org/10.1073/pnas.2021855118/-/DCSupplemental.y.
Benedetti, A.I., Casagli, N., Dapporto, S., Palmieri, M., Zinoni, F., 2005. Modello
statistico per la previsione operative dei fenomeni franosi nella regione Emilia-
Romagna. Ital. J. Geosci. 124, 333–344.
Bessette-Kirton, E.K., Coe, J.A., Schulz, W.H., Cerovski-Darriau, C., Einbund, M.M., 1950.
Mobility characteristics of debris slides and ows triggered by hurricane Maria in
Puerto Rico. Hungr Evans. https://doi.org/10.1007/s10346-020-01445-z.
Bordoni, M., Persichillo, M.G., Meisina, C., Crema, S., Cavalli, M., Bartelletti, C.,
Galanti, Y., Barsanti, M., Giannecchini, R., D’Amato Avanzi, G., 2018. Estimation of
the susceptibility of a road network to shallow landslides with the integration of the
sediment connectivity. Nat. Hazards Earth Syst. Sci. 18 (6), 1735–1758. https://doi.
org/10.5194/nhess-18-1735-2018.
Brardinoni, F., Slaymaker, O., Hassan, M.A., 2003. Landslide inventory in a rugged
forested watershed: a comparison between air-photo and eld survey data.
Geomorphology 54, 179–196. https://doi.org/10.1016/S0169-555X(02)00355-0.
A. Brath N. Casagli M. Marani P. Mercogliano R. Motta Rapporto Della Commissione
Tecnico-Scientica Istituita Con Deliberazione Della Giunta Regionale n 2023
Region Emila Romagna Bologna (Italy).
Casagli, N., Cigna, F., Bianchini, S., H¨
olbling, D., Füreder, P., Righini, G., Del Conte, S.,
Friedl, B., Schneiderbauer, S., Iasio, C., Vlcko, J., 2016. Landslide mapping and
monitoring by using radar and optical remote sensing: Examples from the EC-FP7
project SAFER. Remote Sens. Appl. Soc. Environ. 4, 92–108. https://doi.org/
10.1016/j.rsase.2016.07.001.
Casagli, N., Frodella, W., Morelli, S., Tofani, V., Ciampalini, A., Intrieri, E., Raspini, F.,
Rossi, G., Tanteri, L., Lu, P., 2017. Spaceborne, UAV and ground-based remote
sensing techniques for landslide mapping, monitoring and early warning.
Geoenvironmental Disasters 4, 9. https://doi.org/10.1186/s40677-017-0073-1.
Cignetti, M., Godone, D., Giordan, D., 2019. Shallow landslide susceptibility, Rupinaro
catchment, Liguria (northwestern Italy). J. Maps 15 (2), 333–345. https://doi.org/
10.1080/17445647.2019.1593252.
Corominas, J., van Westen, C., Frattini, P., Cascini, L., Malet, J.P., Fotopoulou, S.,
Catani, F., Van Den Eeckhaut, M., Mavrouli, O., Agliardi, F., Pitilakis, K., 2014.
Recommendations for the quantitative analysis of landslide risk. Bull. Eng. Geol.
Environ 73, 209–263. https://doi.org/10.1007/s10064-013-0538-8.
Dias, H.C., H¨
olbling, D., Grohmann, C.H., 2023. Rainfall-induced shallow landslide
recognition and transferability using object-based image analysis in Brazil. Remote
Sens. 15, 5137. https://doi.org/10.3390/rs15215137.
Doan, V.L., Nguyen, B.-Q.-V., Pham, H.T., Nguyen, C.C., Nguyen, C.T., 2023. Effect of
time-variant NDVI on landside susceptibility: a case study in Quang Ngai province,
Vietnam. Open Geosci. 15 https://doi.org/10.1515/geo-2022-0550.
Donnini, M., Santangelo, M., Gariano, S.L., Bucci, F., Peruccacci, S., Alvioli, M.,
Althuwaynee, O., Ardizzone, F., Bianchi, C., Bornaetxea, T., Brunetti, M.T., 2023.
Landslides triggered by an extraordinary rainfall event in Central Italy on September
15, 2022. Landslides 20 (10), 2199–2211. https://doi.org/10.1007/s10346-023-
02109-4.
Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B.,
Isola, C., Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F.,
Bargellini, P., 2012. Sentinel-2: ESA’s optical high-resolution Mission for GMES
operational Services. Remote Sens. Environ. 120, 25–36. https://doi.org/10.1016/j.
rse.2011.11.026.
Đuri´
c, D., Mladenovi´
c, A., Peˇ
si´
c-Georgiadis, M., Marjanovi´
c, M., Abolmasov, B., 2017.
Using multiresolution and multitemporal satellite data for post-disaster landslide
inventory in the republic of Serbia. Landslides 14, 1467–1482. https://doi.org/
10.1007/s10346-017-0847-2.
Ferrario, M.F., Livio, F., 2024. Rapid mapping of landslides induced by heavy rainfall in
the Emilia-Romagna (Italy) region in may 2023. Remote Sens. 16 https://doi.org/
10.3390/rs16010122.
Ferrario, M.F., 2023. Inventory of landslides triggered by heavy rainfall in the Emilia-
Romagna region (Italy) in May 2023. Doi: 10.5281/ZENODO.8102429.
Fiorucci, F., Giordan, D., Santangelo, M., Dutto, F., Rossi, M., Guzzetti, F., 2018. Criteria
for the optimal selection of remote sensing optical images to map event landslides.
Nat Hazards Earth Syst Sci 18, 405–417. https://doi.org/10.5194/nhess-18-405-
2018.
Fiorucci, F., Ardizzone, F., Mondini, A.C., Viero, A., Guzzetti, F., 2019. Visual
interpretation of stereoscopic NDVI satellite images to map rainfall-induced
landslides. Landslides 16, 165–174. https://doi.org/10.1007/s10346-018-1069-y.
Galli, M., Ardizzone, F., Cardinali, M., Guzzetti, F., Reichenbach, P., 2007. Comparing
Landslide Inventory Maps. Geomorphology. https://doi.org/10.1016/j.
geomorph.2006.09.023.
Galli, M., Ardizzone, F., Cardinali, M., Guzzetti, F., Reichenbach, P., 2008. Comparing
landslide inventory maps. Geomorphology, GIS Technology and Models for
Assessing Landslide Hazard and Risk 94, 268–289. https://doi.org/10.1016/j.
geomorph.2006.09.023.
Ghorbanzadeh, O., Crivellari, A., Ghamisi, P., Shahabi, H., Blaschke, T., 2021.
A comprehensive transferability evaluation of U-net and ResU-net for landslide
detection from Sentinel-2 data (case study areas from Taiwan, China, and Japan).
Sci. Rep. 11, 14629. https://doi.org/10.1038/s41598-021-94190-9.
Ghorbanzadeh, O., Gholamnia, K., Ghamisi, P., 2022a. The application of ResU-net and
OBIA for landslide detection from multi-temporal Sentinel-2 images. Big Earth Data
1–25. https://doi.org/10.1080/20964471.2022.2031544.
Ghorbanzadeh, O., Shahabi, H., Crivellari, A., Homayouni, S., Blaschke, T., Ghamisi, P.,
2022b. Landslide detection using deep learning and object-based image analysis.
Landslides 19, 929–939. https://doi.org/10.1007/S10346-021-01843-X/TABLES/2.
Giordan, D., Cignetti, M., Baldo, M., Godone, D., 2017. Relationship between man-made
environment and slope stability: the case of 2014 rainfall events in the terraced
landscape of the Liguria region (northwestern Italy). Geomat. Nat. Hazards Risk 8,
1833–1852. https://doi.org/10.1080/19475705.2017.1391129.
Giordan, D., Adams, M.S., Aicardi, I., Alicandro, M., Allasia, P., Baldo, M., De
Berardinis, P., Dominici, D., Godone, D., Hobbs, P., Lechner, V., Niedzielski, T.,
Piras, M., Rotilio, M., Salvini, R., Segor, V., Sotier, B., Troilo, F., 2020. The use of
unmanned aerial vehicles (UAVs) for engineering geology applications. Bull. Eng.
Geol. Environ. 79, 3437–3481. https://doi.org/10.1007/s10064-020-01766-2.
D. Notti et al.
International Journal of Applied Earth Observation and Geoinformation 129 (2024) 103806
13
Guzzetti, F., Malamud, B.D., Turcotte, D.L., Reichenbach, P., 2002. Power-law
correlations of landslide areas in central Italy. Earth Planet. Sci. Lett. 195, 169–183.
https://doi.org/10.1016/S0012-821X(01)00589-1.
Guzzetti, F., Cardinali, M., Reichenbach, P., Cipolla, F., Sebastiani, C., Galli, M.,
Salvati, P., 2004. Landslides triggered by the 23 november 2000 rainfall event in the
Imperia Province, Western Liguria, Italy. Eng. Geol. 73, 229–245. https://doi.org/
10.1016/j.enggeo.2004.01.006.
Guzzetti, F., Reichenbach, P., Cardinali, M., Galli, M., Ardizzone, F., 2005. Probabilistic
landslide hazard assessment at the basin scale. Geomorphology 72, 272–299.
https://doi.org/10.1016/j.geomorph.2005.06.002.
Guzzetti, F., Peruccacci, S., Rossi, M., Stark, C.P., 2008. The rainfall intensity–duration
control of shallow landslides and debris ows: an update. Landslides 5, 3–17.
https://doi.org/10.1007/s10346-007-0112-1.
Guzzetti, F., Mondini, A.C., Cardinali, M., Fiorucci, F., Santangelo, M., Chang, K.-T.,
2012. Landslide inventory maps: new tools for an old problem. Earth Sci. Rev. 112,
42–66. https://doi.org/10.1016/j.earscirev.2012.02.001.
Ibsen, M.L., Casagli, N., 2004. Rainfall patterns and related landslide incidence in the
Porretta-Vergato region, Italy. Landslides 1, 143–150. https://doi.org/10.1007/
s10346-004-0018-0.
Jung, J., Yun, S.-H., 2020. Evaluation of coherent and incoherent landslide detection
methods based on synthetic aperture Radar for rapid response: a case study for the
2018 Hokkaido landslides. Remote Sens. 12, 265. https://doi.org/10.3390/
rs12020265.
Lee, S., Min, K., 2001. Statistical analysis of landslide susceptibility at yongin, Korea. Env
Geol 40, 1095–1113. https://doi.org/10.1007/s002540100310.
Li, Z., Shi, W., Myint, S.W., Lu, P., Wang, Q., 2016. Semi-automated landslide inventory
mapping from bitemporal aerial photographs using change detection and level set
method. Remote Sens. Environ. 175, 215–230. https://doi.org/10.1016/j.
rse.2016.01.003.
Lu, P., Qin, Y., Li, Z., Mondini, A.C., Casagli, N., 2019. Landslide mapping from multi-
sensor data through improved change detection-based Markov random eld. Remote
Sens. Environ. 231, 111235 https://doi.org/10.1016/j.rse.2019.111235.
Martelloni, G., Segoni, S., Lagomarsino, D., Fanti, R., Catani, F., 2013. Snow
accumulation/melting model (SAMM) for integrated use in regional scale landslide
early warning systems. Hydrol. Earth Syst. Sci. 17, 1229–1240. https://doi.org/
10.5194/HESS-17-1229-2013.
Martina, M., Pignone, S., Berti, M., Simoni, A., 2010. Un approccio bayesiano per
individuare le soglie di innesco delle frane. In: Proceedings of the First Italian
Workshop on Landslides.
MASE, 2009. Carta geolitologica d’Italia. Risoluzione 1:500.000.
Meena, S.R., Nava, L., Bhuyan, K., Puliero, S., Soares, L.P., Dias, H.C., Floris, M.,
Catani, F., 2023. HR-GLDD: a globally distributed dataset using generalized deep
learning (DL) for rapid landslide mapping on high-resolution (HR) satellite imagery.
Earth Syst. Sci. Data 15, 3283–3298. https://doi.org/10.5194/ESSD-15-3283-2023.
Milledge, D.G., Bellugi, D.G., Watt, J., Densmore, A.L., 2022. Automated determination
of landslide locations after large trigger events: advantages and disadvantages
compared to manual mapping. Nat. Hazards Earth Syst. Sci. 22 (2), 481–508.
https://doi.org/10.5194/NHESS-22-481-2022.
Mondini, A.C., Guzzetti, F., Reichenbach, P., Rossi, M., Cardinali, M., Ardizzone, F.,
2011. Semi-automatic recognition and mapping of rainfall induced shallow
landslides using optical satellite images. Remote Sens. Environ. 115, 1743–1757.
https://doi.org/10.1016/j.rse.2011.03.006.
Mondini, A.C., Marchesini, I., Rossi, M., Chang, K.-T., Pasquariello, G., Guzzetti, F., 2013.
Bayesian framework for mapping and classifying shallow landslides exploiting
remote sensing and topographic data. Geomorphology 201, 135–147. https://doi.
org/10.1016/j.geomorph.2013.06.015.
Mondini, A.C., Guzzetti, F., Chang, K.T., Monserrat, O., Martha, T.R., Manconi, A., 2021.
Landslide failures detection and mapping using synthetic aperture Radar: past,
present and future. Earth-Sci. Rev. 216, 103574 https://doi.org/10.1016/J.
EARSCIREV.2021.103574.
Mondini, A.C., Guzzetti, F., Melillo, M., 2023. Deep learning forecast of rainfall-induced
shallow landslides. Nat. Commun. 14, 2466. https://doi.org/10.1038/s41467-023-
38135-y.
Murillo-García, F.G., Alc´
antara-Ayala, I., Ardizzone, F., Cardinali, M., Fiourucci, F.,
Guzzetti, F., 2015. Satellite stereoscopic pair images of very high resolution: a step
forward for the development of landslide inventories. Landslides 12, 277–291.
https://doi.org/10.1007/s10346-014-0473-1.
Nava, L., Bhuyan, K., Meena, S.R., Monserrat, O., Catani, F., 2022. Rapid mapping of
landslides on SAR data by attention U-net. Remote Sens. 14, 1449. https://doi.org/
10.3390/rs14061449.
Notti, D., Cignetti, M., Godone, D., Giordan, D., 2023. Semi-automatic mapping of
shallow landslides using free Sentinel-2 images and Google Earth engine. Nat.
Hazards Earth Syst. Sci. 23, 2625–2648. https://doi.org/10.5194/NHESS-23-2625-
2023.
Notti, D., Cignetti, M., Cardone, D., Godone, D., Giordan, D., 2023a. Rapid mapping of
potential ground effects of the May 2023 Emilia-Romagna rainstorms. (Version 1).
Novellino, A., Pennington, C., Leeming, K., Taylor, S., Alvarez, I.G., McAllister, E.,
Arnhardt, C., Winson, A., 2024. Mapping landslides from space: a review. Landslides.
https://doi.org/10.1007/s10346-024-02215-x.
Pini, G.A., 1999. Tectonosomes and olistostromes in the Argille Scagliose of the Northern
Apennines, Italy, Vol. 335. Geological Society of America.
E.-R. Region 2020 - coperture vettoriali uso del suolo di dettaglio - edizione 2023 [WWW
document] Geoportale. 2023 accessed 1.19.24 https://geoportale.regione.emilia-ro
magna.it/catalogo/dati-cartograci/pianicazione-e-catasto/uso-del-suolo/layer
-14.
Reichenbach, P., Rossi, M., Malamud, B.D., Mihir, M., Guzzetti, F., 2018. A review of
statistically-based landslide susceptibility models. Earth-Sci. Rev 180, 60–91.
https://doi.org/10.1016/j.earscirev.2018.03.001.
Santangelo, M., Cardinali, M., Bucci, F., Fiorucci, F., Mondini, A.C., 2022. Exploring
event landslide mapping using Sentinel-1 SAR backscatter products. Geomorphology
397, 108021. https://doi.org/10.1016/j.geomorph.2021.108021.
Santangelo, M., Marchesini, I., Bucci, F., Cardinali, M., Fiorucci, F., Guzzetti, F., 2015. An
approach to reduce mapping errors in the production of landslide inventory maps.
Nat. Hazards Earth Syst. Sci. 15 (9), 2111–2126. https://doi.org/10.5194/nhess-15-
2111-2015.
Trigila, A., Iadanza, C., Spizzichino, D., 2008. IFFI project (italian landslide inventory)
and risk assessment. First World Landslide Forum. 18–21.
van Westen, C.J., Castellanos, E., Kuriakose, S.L., 2008. Spatial data for landslide
susceptibility, hazard, and vulnerability assessment: an overview. eng. geol.
Landslide Susceptibility, Hazard and Risk Zoning for Land Use Planning 102,
112–131. https://doi.org/10.1016/j.enggeo.2008.03.010.
D. Notti et al.
