Available via license: CC BY 4.0
Content may be subject to copyright.
1
Is higher resolution always better? Open-access DEM comparison for
1
Slope Units delineation and regional landslide prediction
2
Mahnoor Ahmeda, Giacomo Tittib*, Sebastiano Trevisanic, Lisa Borgattib, Mirko Francionia
3
aDepartment of Pure and Applied Sciences, University of Urbino Carlo Bo, Urbino, 61029, Italy
4
bDepartment of Civil, Chemical, Environmental and Materials Engineering, ALMA MATER STUDIORUM
5
University of Bologna, Bologna, 40126, Italy
6
cDipartimento di Culture del Progetto, University Iuav of Venice, Venezia, 30135, Italy
7
*Correspondence to: Giacomo Titti (giacomo.titti@unibo.it)
8
Abstract
9
Digital Elevation Models (DEMs) play a key role in slope instability studies, ranging from landslide detection and
10
recognition to landslide prediction. DEMs assist these investigations by reproducing landscape morphological
11
features and deriving relevant predisposing factors, such as slope gradient, roughness, aspect, and curvature.
12
Additionally, DEMs are useful for delineating map units with homogeneous morphological characteristics, such
13
as Slope Units (SUs).
14
In many cases, the selection of a DEM depends on factors like accessibility and resolution, without considering
15
its actual accuracy. In this study, we compared freely available global DEMs (ALOS, COP, FABDEM) and a
16
national DEM (TINITALY) with a reference DEM (local airborne LiDAR) to identify the most suitable DEM for
17
representing fine-scale morphology and delineating SUs in the Marche Region, Italy, for landslide susceptibility
18
mapping. Furthermore, we proposed a novel approach for selecting the optimal SUs partition.
19
The DEM comparison was based on several criteria, including elevation, residual DEMs, roughness indices, slope
20
variations, and the ability to delineate SUs. TINITALY, resampled at a 30x30m pixel size, was found to be the
21
most suitable DEM for representing fine-scale terrain morphology. It was then used to generate the optimal SUs
22
partition among 18 combinations. These combinations were evaluated using both existing and newly integrated
23
metrics alongside mapped landslide inventories to optimize terrain delineation and produce landslide
24
susceptibility maps.
25
26
Introduction
27
Open-access global Digital Elevation Models (DEMs) have been commonly used for a vast range of
28
geomorphological studies, which have required modelling or analysis of terrain surface in mountain environments,
29
where these DEMs have been characterized by a marked quality deterioration (Guth et al., 2024; Trevisani et al.,
30
2023b). One of the many uses of DEMs has been to serve as the base input for analyzing landslides morphological
31
features, state and style of activity and generating landslide susceptibility models (Brock et al., 2020). Among
32
multiple methods of data-driven (Ahmed et al., 2023; Lombardo et al., 2020; Lombardo and Tanyas, 2020; Titti
33
et al., 2021a) and physical-based models (Van den Bout et al., 2021) to predict, investigate (Brenning, 2005;
34
Pirasteh and Li, 2017; Steger et al., 2023) and detect landslides (Qin et al., 2013), the elevation model has been
35
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
2
of essential use. DEMs are utilized to derive terrain-based characteristics (Brock et al., 2020; Mahalingam and
36
Olsen, 2016) which have been conditioned by their resolution. In the literature, DEM resolution and its influence
37
have been tested in several aspects such as; in landslide modelling and hazard assessment (Catani et al., 2013;
38
Claessens et al., 2005; Fenton et al., 2013; Huang et al., 2023) , in 3D physical models (Qiu et al., 2022), as well
39
as morphological quality assessment explored at regional scales (Grohmann, 2018; Hawker et al., 2019; Trevisani
40
et al., 2023b).
41
Comparisons among DEMs to evaluate the most suitable product are based on different criteria and the results
42
have likely varied as per the test site. Thus, even if the same criteria have been used to rank DEMs, regional
43
topography has influenced the preference of the elevation model in different areas (Florinsky et al., 2019; Zhang
44
et al., 2019). Landcover has been specifically important when global DEMs (Bielski et al., 2024), such as
45
Copernicus DEM and ALOS AW3D30, have been used for deriving a Digital Terrain Model (DTM), given that
46
most of the times these products resembled more a Digital Surface Model (DSM: Guth & Geoffroy, 2021).
47
An ongoing initiative, the Digital Elevation Model Inter-comparison eXercise (DEMIX; Strobl et al., 2021), has
48
aimed to align methodologies allowing for criteria-based ranking of global DEMs. In the first application (Bielski
49
et al 2024), metrics related to slope and roughness have been considered in addition to those related to elevation
50
differences; the approach has further developed, adopting new metrics and a wide range of geomorphometric
51
derivatives (Guth et al., 2024). Global DEMs have been commonly used in geoscientific research due to their
52
spatial extent and public accessibility whereas national DEMs (Gesch et al., 2018; Muralikrishnan et al., 2013;
53
Tarquini et al., 2007) have generally been tailored to represent country-specific land surface and morphology at a
54
higher spatial resolution and accuracy to serve geoscience applications. Shuttle Radar Topography Mission
55
(SRTM; Jarvis et al., 2008), Advanced Land Observing Satellite (ALOS; (Takaku et al., 2014), Terra Advanced
56
Spaceborne Thermal Emission and Reflection Radiometer Global DEM (ASTER GDEM; Abrams et al., 2010)
57
have been among the most widely used, freely accessible and initial global DEMs popularized in geomorphic
58
analysis (Becek, 2014; Florinsky et al., 2019; Mahalingam and Olsen, 2016; Trevisani et al., 2023b; Zhang et al.,
59
2019). However, there are many considerations to be considered for implementing these global datasets to a
60
localized area in the frame of landslide recognition, mapping and assessment.
61
Landslide inventories and elevation models have been essential inputs for data-driven landslide models, for which
62
the DEM has been used to derive morphological parameters such as slope angle and slope aspect. For these
63
derivatives to be as accurate as possible in a model, the DEM quality (Claessens et al., 2005; Mahalingam and
64
Olsen, 2016; Saleem et al., 2019) should satisfy the representation of fine-scale morphology (Chaplot et al., 2006;
65
Florinsky, 1998). In other words, the DEM quality has significantly affected the prediction capacity of a model.
66
The errors contained within a DEM, even when small, propagate in derivatives of elevation (Karakas et al., 2022;
67
Mahalingam and Olsen, 2016; Pawluszek and Borkowski, 2017; Saleem et al., 2019) which have been weighed
68
as important factors in landslide occurrence. The various available DEMs have been generated using a range of
69
technologies. While significant efforts have been made to improve DEMs over time, the accuracy of these models
70
has remained a critical issue. Selecting an appropriate DEM has proven to be more important than the number of
71
DEM-derived factors used in landslide assessment (Kamiński, 2020).
72
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
3
Another use of DEMs has been the delineation of mapping units (Schlögel et al., 2018). Mapping units have been
73
used to subdivide the study area in homogeneous, elemental units such as: administrative units (Lombardo et al.,
74
2019), terrain units (Van Westen et al., 1997), unique condition units (Titti et al., 2021b), grid cells (Reichenbach
75
et al., 2018) or Slope Units (SUs; Ahmed et al., 2023). SUs were initially introduced by Carrara et al. (1991) as
76
portions of territory, presenting homogeneous morphological characteristics for landslide identification and
77
susceptibility mapping. The SU is, according to the scale adopted, has served as a solution that adequately
78
represents unstable slopes.
79
To assess the suitability of DEMs for landslide susceptibility and prediction, it has been essential to conduct a
80
quality assessment of these models which has commonly referred to the spatial resolution alone. Therefore, global
81
DSMs and a national Italian Digital Terrain Model (DTM) has been compared with a local accurate elevation
82
model (Airborn LiDAR) in the context of terrain representation and its delineation. The Italian DTM has been
83
already investigated in some studies, mainly focusing on hydrogeomorphology studies (Pulighe and Fava, 2013;
84
Zingaro et al., 2021; Annis et al., 2020; Tavares da Costa et al., 2019). Accordingly, the quality evaluation from
85
the perspective of fine-scale morphology and geomorphometric derivatives in the context of landslide science has
86
remained an interesting aspect to elaborate on.
87
This study has aimed to optimize inputs used for representing morphological data in landslide susceptibility
88
assessment and to understand their interactions by: identifying the most suitable DEM for accurately representing
89
fine-scale slope morphology; proposing a new metric for analyzing optimal SU parameters for landslide
90
susceptibility mapping, integrating landslide inventory data with landslide area and numerosity; extending and
91
applying the methodology to test landslide susceptibility at a regional scale in the Marche Region of Central Italy.
92
Study Area
93
In this study, we have selected two distinct study areas. The first Area of Interest (AOIa) has encompassed the
94
entire Marche region, located in central-eastern Italy (Figure 1, AOIa). From the morphological point of view, this
95
region is characterized by three different types of landforms that extend in the north-south direction. In the western
96
part, the region has been crossed by the Apennines which can reach, in the area, a peak of 2476m a.s.l. at Monte
97
Vettore. Then, the reliefs degrade to more rounded hills in the central part of the region till the flat eastern coastal
98
strip. From a geological point of view, the Apennines, a Neogene fold-and-thrust belt that developed following
99
the closure of the Mesozoic Tethys Ocean, have been characterized by calcareous units, calcareous-marly units
100
and arenaceous, pelitic-arenaceous and marly arenaceous units (aged from Jurassic to Neogene). The h illy and
101
coastal areas have mainly been characterized by Neocene/quaternary clayey formations. Several small rivers
102
traverse the region from the west to the east side. In particular, the basins of Misa, Esino, Cesano, and Metauro
103
rivers were affected by an exceptional thunderstorm in September 2022 which triggered floods and landslides
104
(Corti et al., 2024). One of the highest rainfall intensity of the 2022 event was registered in a sub-portion of the
105
Marche region, that has been selected as the second study area (AOIb) for this study (Figure 1, AOIb) not only
106
because the consequences of the exceptional rainfall event but also because, morphologically, it can be considered
107
a representative sample of the Marche mountainous region. Moreover, the area has been covered by a high-
108
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
4
resolution dataset (1m pixel size) which allows us to effectively conduct the experiments as described in the
109
following text.
110
A relevant portion of the territory of Marche region (AOIa) presents slope failures. The most populated dataset of
111
landslide in the area is the inventory of the Piano stralcio per l’Assetto Idrogeologico (PAI) of Marche Region
112
(Figure 1).Click or tap here to enter text. In the area of Marche region (AOIa), the PAI inventory counts 19,296
113
delimited landslides for a total landslide area of 1394 km2 which covers 15% of the total regional surface classified
114
as flow, slide and complex landslides.
115
116
Figure 1: Study area in central Italy. On the left, is the study area AOIa, encompassing the entire Marche region which
117
has been analyzed in the second phase of the study. On the right is study area AOIb, a sub-portion of the Marche region
118
where we conducted the DEM analysis in the first phase covered by the 1 m pixel size airborne LiDAR survey. The
119
Piano Stralcio per l’Assetto Idrogeologico (PAI) landslide inventory of the Marche Region identifies 19,296 landslide
120
bodies as polygons (image background from © Google Maps 2019).
121
122
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
5
Materials and Methods
123
The methodology implemented in this study has aimed to assess the quality of freely available DEMs, framing
124
their use for landslide susceptibility assessment. DEMs have been essential because they allow the derivation of
125
landslide predisposing factors and generate a morphology-based terrain subdivision: SUs. Thus, these two uses of
126
a DEM in landslide susceptibility assessment have been investigated.
127
The analysis has been conducted in two sequential phases (Figure 2): the first phase the differences in DEM
128
derivatives have been assessed by comparing global and a national DEM to a local high-resolution reference
129
elevation data in AOIb. In the second phase we have evaluated 18 SUs partitions on the base of internal/external
130
homogeneity, landslide extension and landslide number using the best performing open-source DEM, which has
131
been identified in the first phase of this study.
132
133
Figure 2: Workflow of the two-phases analysis. Phase 1 DEM assessment: comparison of global and national DEM to
134
a local high-resolution reference elevation model (w ith reference to AOIb). Phase 2 Slope Units delineation: selection
135
of the optimal parameters for SUs delineation (with reference to AOIa).
136
Phase 1: DEM assessment
137
In this phase, the accuracy of three global DEM (ALOS World 3D-30m, Copernicus-GLO-30, FABDEM) and one
138
national DEM (TINITALY) has been evaluated by a comparison with a local airborne LiDAR in the study area
139
AOIb.
140
ALOS World 3D-30m (ALOS World 3D - 30m. V3.2, 2024) has been released by Japan Aerospace Exploration
141
Agency (JAXA) in 2015, at a horizontal resolution of 1 arc-second, approximately 30 meters as a DSM (Caglar
142
et al., 2018). This product, surveyed from 2006 to 2011, uses the 5-meter mesh of "World 3D Topographic Data"
143
and is provided in two resampled versions by JAXA (mean resampling kernel is used in this study), with elevation
144
expressed according to the Earth Gravitational Model 1996 (EGM96).
145
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
6
Copernicus-GLO-30 (COP; ESA & Sinergise, 2021), has been obtained from the WorldDEM at 1 arc second as a
146
DSM, a product of the radar data acquisition of 12 meters TanDEM-X mission from 2011 to 2015. Forest And
147
Buildings removed Copernicus DEM (FABDEM; Hawker et al., 2022), has been made available as a corrected
148
global Digital Terrain Model (DTM) available at 1 arc-second grid spacing (60°S-80°N) derived from Copernicus
149
GLO-30. Machine learning techniques have been devised to improve mean absolute vertical error in built-up and
150
forested areas in comparison to COP (Hawker et al., 2022). Both FABDEM and COP elevations have been referred
151
to EGM 2008 geoid.
152
TINITALY’s 1.0 (Tarquini et al., 2007), and version 1.1 (Tarquini et al., 2023), has covered the whole of Italian
153
territory, as a DTM available at 10m pixel size. Heterogenous data, mainly based on Technical Regional
154
Cartography with elevations derived by means of photogrammetric method, has been used to build a national scale
155
model. In particular, the Technical Regional Cartography (CTR) map scaled at 1:10000 with 10m interval for
156
contour lines is used for Marche region in the compilation of TINITALY. A Triangular Irregular Network (TIN)
157
structure has been employed in constructing the DEM to tackle varying data density and redundancy. Merging
158
various types of input data is followed by significant investigation to ensure the seamless production of a high
159
resolution and considerably the most accurate representation for Italy, with a root mean square error ranging from
160
0.1 to 6 meters (Tarquini et al., 2007).
161
The reference DEM (as called hereafter) has been a Digital Terrain Model (DTM) acquired in 2012 using airborne
162
LiDAR, with a pixel size of 1x1m, and a reported vertical and planimetric accuracy of 15 cm and 30 cm,
163
respectively (Ministero dell'Ambiente e della Sicurezza Energetica, https://gn.mase.gov.it/portale/pst-dati-lidar).
164
This reference DTM has been aggregated via averaging the pixel size to 30m.
165
COP, FABDEM, ALOS and TINITALY have been projected in WGS84 UTM 33N at a pixel size of 30 meters
166
using bilinear interpolation for alignment with the reference DEM. The inclusion of COP and FABDEM, along
167
with ALOS as a global DEM and TINITALY as a national-scale elevation model for comparison, has been invoked
168
by several studies (Bielski et al., 2024; Guth & Geoffroy, 2021; Meadows et al., 2024; Osama et al., 2023;
169
Trevisani et al., 2023). All the DEMs, except TINITALY (geoid model not publicly available), have been
170
transformed to a common geoid model, EGM2008 respectively for alignment and comparison with the reference
171
grid. TINITALY is based on the Italian geodetic network (IGM95) where the measured ground points have been
172
described by the Italian geoid called ITALGEO 2005 (Albertella et al., 2008; Barzaghi et al., 2007). Barzaghi and
173
Carrion (2009) have concluded that the difference between ITALGEO05 (regional geoid model) and EGM2008
174
(global geoid model) is negligible for many applications, and both are capable to represent the region of Italy.
175
Therefore, no geoid transformation for TINITALY has been required.
176
To perform the quality assessment of selected DEMs, elevation differences have been considered for compatibility
177
with precedent studies. Indeed, studies focusing on DEMs comparison (Polidori and Hage, 2020) are generally
178
based on elevation differences, using standard statistical metrics such as standard deviation and Root Mean Square
179
Error (RMSE), and in some cases slope and aspect have been considered (Meadows et al., 2024; Zhang et al.,
180
2019). However, as suggested in many studies (Bielski et al., 2024; Crema et al., 2020; Florinsky et al., 2019;
181
Gesch, 2018; Guth & Geoffroy, 2021; Kakavas et al., 2020; Liu et al., 2019; Purinton & Bookhagen, 2017;
182
Trevisani et al., 2023), statistical metrics of elevation differences alone fail to fully capture the quality of DEMs,
183
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
7
including the capability to represent fine-scale morphology and the presence of artifacts. Therefore, for this reason
184
and because the focus of the work has been to investigate mainly the accuracy of the DEMs geomorphometric
185
derivatives, along with the differences in elevation, a straightforward and simple approach to take the local spatial
186
variability of surfaces into account based on a geostatistical-based methodology (Isaaks and Srivastava, 1989), as
187
discussed by Trevisani et al. (2023b), has been proposed.
188
The approach has been based on the derivation of a residual DEM, also known as Topographic Position Index
189
(TPI; Guisan et al., 1999; Hiller and Smith, 2008; Wilson and Gallant, 2000), and the calculation of roughness
190
indices. The residual DEM, derived by detrending the original surface, has permitted to highlight the capability
191
of DEMs to reproduce local fine-scale morphology. Moreover, the residual DEM has been used as input for the
192
calculation of roughness indices such as the standard deviation of residual DEM (Grohmann et al., 2011) or even
193
geostatistical based estimators such as the variogram (Eq 1, with p = 2), the madogram (Eq. 1, with p = 1) and
194
(Eq. 2) represents the more robust Median Absolute Differences (MAD; Trevisani and Cavalli, 2016; Trevisani
195
and Rocca, 2015). The generalization of the variogram have been described as in Eq. (1) and MAD as Eq. (2);
196
, (1)
197
where,
198
199
, (2)
200
where is the separation vector (lag) between two locations , is the value of the variable of interest in
201
the location (e.g., residual elevation), and is the number of point pairs with a separation vector found in
202
the search window considered. Accordingly, the variogram is the half of the mean squared differences and
203
the MAD is the median of the absolute differences . It should be highlighted that there are roughness indices
204
such as MADk2 and the Radial Roughness Index (RRI) that have been calculated directly from the DEM, without
205
detrending (Trevisani et al., 2023c, a).
206
A simple short-range omnidirectional roughness index, such as MAD calculated for lag distances of 2 pixels and
207
circular kernel of 3 pixels, permits to analyze fine-grain roughness (see Trevisani et al., 2023a; Trevisani and
208
Rocca, 2015 for a full discussion). The MAD omnidirectional roughness index essentially provides a measure of
209
omnidirectional spatial variability (median differences in residual elevation) by comparing all pixel values
210
separated by a distance of |h| pixels in the considered moving window. An alternative roughness index which does
211
not require the definition of calculation parameters is the RRI (Trevisani et al., 2023c), that has been derived to
212
improve the popular Topographic Ruggedness Index (TRI; Riley et al., 1999).
213
All the comparisons have been done using a pixel size of 30x30m. This value was assumed because it is closer to
214
the size of global 1 arc second DEMs, except for TINITALY which is released with a pixel size of 10x10m.
215
TINITALY has been upscaled by mean-pixel aggregation to 30x30m pixel size. The 30m DEM (TINITALY30m)
216
has also been compared with the 10m pixel size version (TINITALY10m) in AOIb to assess the effect of upscaling
217
on the analysis. Given that slope, roughness indices and residual DEM are scale-dependent geomorphometric
218
derivatives, a normalization has been done to compare the results of the differences between the derivatives at
219
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
8
different resolutions of TINITALY and the reference DEM. Accordigly, a normalized difference has been adopted
220
for each derivative D:
221
.
222
Finally, an additional analysis has been conducted. Since the goal of the research proposes attribution to landslide
223
studies, the DEM-derived slope difference distribution in the landslide areas delineated by the PAI inventory is
224
also included. To avoid overestimation of landslide areas, the overlapping polygons, primarily representing
225
reactivations, have been merged.
226
To further assist in evaluating the quality of DEMs in the frame of landslide susceptibility assessment, the SUs
227
have been generated using various DEMs (global and national). This has allowed for a comparison of the SUs
228
produced from the reference EM with those derived from the global DEMs under evaluation, highlighting any
229
differences in terrain partitioning and geometry. The software r.slopeunits (Alvioli et al., 2016) has been used to
230
generate the SU maps, starting from the SU parameters proposed by Alvioli et al. (2016) for AOIb. After a few
231
corrections and optimizations, the parameters have been set as: flow accumulation threshold to 5×105 m2,
232
minimum SU area as 80,000 m2, circular variance as 0.4 and clean size of 60,000 m2 with the cleaning method
233
(flag -m) that removes SU smaller than the clean size as well as removes odd-shaped polygons and SUs with
234
width as small as two grid cells (Alvioli et al., 2016). To quantify the similarity between SUs derived from
235
reference DEM and from each DEM under observation, the Jaccard Index (Jaccard, 1901) has been utilized to
236
estimate Intersection-over-Union (IoU) ratio between the reference (in this case SU derived from reference DEM)
237
and the predicted (in this case the DEM under test). The Jaccard Index can measure the segmentation of the SU
238
in reference to the overlapping of the defined shape and similarity of terrain-representation. Ranging from 0,
239
signifying no similarity to 1 that signifies identical sets, this index considers the combined size which is inclusive
240
of the intersection. Hence, the higher the index value, the better delineation of terrain as per the considered
241
reference.
242
Phase 2: Slope Units delineation
243
This phase of the work has been focused on the identification of the most representative and freely available DEM
244
to subdivide the study area in SUs for landslide modelling. Therefore, 18 SUs partitions have been generated with
245
r.slopeunits software and then compared with landslide areas and landslide counts mapped in the AOIa to find the
246
optimal ones. The optimal DEM obtained from the first phase has been used to test SU delineation in the study
247
area with a range of parameters. As proposed by Alvioli et al. (2016), an aspect segmentation metric has been used
248
to analyze the optimal parameters for the Marche region, altering two parameters: the minimum surface area of
249
SU and the minimum circular variance for terrain, and fixing the parameters flow accumulation and clean size.
250
The aspect segmentation metric has been based on the concept of partitioning terrain by grouping pixels sharing
251
similar aspect properties. This has been transferred to SU delineation, with the assumption, given the partitioning
252
has been evaluated by the internal homogeneity and external heterogeneity of SU. The aspect segmentation metric
253
can be written as:
254
, (3)
255
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
9
where (SU homogeneity) is the local aspect variance and is the autocorrelation which represents the external
256
heterogeneity of the adjacent SUs and evaluates the morphometric delineation of the SUs, explained by the
257
minimum surface area of a SU () and the minimum circular variance () (see for more details, Alvioli et al.,
258
2016). The first term of value is estimated based on the homogeneity of pixels grouped in a single SU, thus a
259
higher value represents a better segmentation. In the same way, on the base of the second term of Eq. 3, the greater
260
the difference between the average aspect value of each SU and each of the relative adjacent SU, the higher is the
261
value. Overall, from a geometrical point of view the optimal a and c combination is the one that maximizes the
262
metric value.
263
Differently from Alvioli et al. (2016) where the Area Under the Curve (AUC) derived from landslide susceptibility
264
assessment has been also considered in selecting the optimal SU parameters, this study proposes to compare the
265
landslides extension ) and landslide density ) per SU. The former sums the percentage of the landslide area
266
included inside the SU where the failure has been triggered (from the initiation point). The latter is the inverse of
267
the average number of landslides in each SU. andcan be expressed as;
268
, (4)
269
, (5)
270
where Li , in Eq. 4, is the total landslide area of all the events triggered in the ith SU, li is the cumulative landslides
271
area inside the ith SU which excludes the extension of landslide that occupies adjacent SUs, N, in Eq. 5, is the
272
number of unstable SUs, di is the number of landslides triggered in the ith SU.
273
, (6)
274
where is the final metric which combines , and . The optimal combination of a and c for SU delineation in
275
the study area selected is the one that maximizes the metric in Eq. 6. SU parameters for the experiment on entire
276
Marche region have been tested with; flow accumulation threshold to 10×105 m2, clean size of 20,000 m2 with the
277
cleaning method (flag -m). Minimum area (a) has been tested with 40, 80, 150, 200, 300 and 500x103 m2 with
278
corresponding circular variance (c) of 0.1, 0.4 and 0.7 for each a, making 18 combinations.
279
The Susceptibility Zoning plugin (SZ-plugin), integrated with QGIS and developed by Titti et al. (2022), has been
280
used to calculate the aspect segmentation metric (F) and to map the landslide susceptibility in the Marche region
281
(AOIa). This analysis has utilized the DEM selected in Phase 1 and assessed four Slope Unit (SU) delineations,
282
ranked from highest to lowest performance, as mapping units for evaluating landslide susceptibility. The analysis
283
has been conducted using a Generalized Additive Model (Loche et al., 2023). The covariates selection includes:
284
lithology, from national dataset (http://portalesgi.isprambiente.it/), landcover (2018 CORINE,
285
https://land.copernicus.eu/en) as categorical covariates. The continuous covariates have been generated using the
286
Spatial Reduction Tool (Titti et al., 2022) from the phase 1-selected DEM as derivatives; slope angle, planar and
287
profile curvature as ordinal covariates and northness, eastness as linear covariates. The collinearity between the
288
predisposing factors has been evaluated by the Pearson’s coefficient. The results have been validated with a 10-
289
fold spatial cross-validation which clusters the dataset with a k-means approach (Elia et al., 2023). The overall
290
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
10
prediction capacity has been estimated with ROC-based AUC (Fawcett, 2006), F1 score (Singhal, 2001) and
291
Coen’s Kappa score (K; Kraemer, 2015).
292
Results
293
The differences between elevation, residual DEMs, roughness indices and slope variations within the four selected
294
open-access DEMs and the reference DEM have been shown in Figure 3. The boxplots report the distribution of
295
the differences higlighting the median, the first and the third quartile excluding the outlayers. Moreover, since the
296
differences report positive and negative values, the absolute mean difference has been calculated. Therefore, the
297
lower the variance and the absolute mean difference, the better is the output considered.
298
Overall, TINITALY resampled at 30m (TINITALY30m) has showcased the best performance across all metrics,
299
with a smaller distribution of differences and lower absolute mean difference. ALOS, on the other hand, has
300
displayed the largest difference among all DEMs across all metrics. Between COP and FABDEM, COP has shown
301
a larger distribution of elevation differences, and as expected, COP has had a stronger tendency to overestimate
302
elevation with respect to FABDEM (Figure 3). However, for slope (Figure 3B) and isotropic roughness (Figure
303
3C), FABDEM has displayed more spread in differences.
304
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
11
305
Figure 3:Boxplots visualizing the differences among the DEMs at 30 meters, using different metrics with the absolute
306
mean calculated; A) Elevation, B) Slope, C) Isotropic Roughness Index, D) Radial Roughness Index and E) Residual
307
DEM.
308
Figure 4 exhibits the differences of the selected derivatives between TINITALY30m and TINITALY10m. Apart
309
the elevation, TINITALY at 10m is quantifying a larger distributions in normalized differences for the terrain
310
indices. The absolute mean difference confirms the trend.
311
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
12
312
Figure 4: Boxplots showing the differences in TINITALY at 10m and 30m with respect to the reference LiDAR at the
313
respective resolution, for different indices with the absolute mean calculated; A) Elevation, B) Slope, C) Isotropic
314
Roughness Index, D) Radial Roughness Index and E) Residual DEM.
315
Since the main topic of our analysis is to support landslide susceptibility mapping, we have investigated the
316
performance of the selected DEMs to derive slope, which is considered one of the most relevant landslide
317
predisposing factors, in the area where landslide bodies have been mapped. Figure 5 shows the slope-difference
318
within the mapped polygons of the PAI landslide inventory. TINITALY30m is seen to have the smallest differences
319
in terms of absolute mean and the distribution among all the other DEMs (Figure 5A). Similarly, in Figure 5B,
320
the distributions of the normalized differences of TINITALY 10m and 30m clearly highlight the larger differences
321
distribution of the 10m DEM.
322
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
13
323
Figure 5: A) Slope differences for 30m DEMs as compared to the reference DEM in PAI landslide polygons. B)
324
Normalized difference in slope with reference DEM for 10m and 30m TINITALY in PAI landslide polygons.
325
The last part of the DEMs comparison would investigate the effect on the SUs delineation of different DEMs.
326
Table 1 reports the Jaccard index tested comparing the SUs delineated with DEMs at 30m and SU generated with
327
the reference DEM. The highest similarity index is for TINITALY30m.
328
Table 1: Jaccard Index represented as Intersection-over-union for SUs generated from the DEMs under test with the
329
reference LiDAR DEM SUs.
330
DEM
IoU
ALOS
0.866
FABDEM
0.896
COP
0.887
TINITALY30m
0.912
331
The second phase of the analysis has been focused on the optimal SUs delineation to assess landslide susceptibility
332
in AOIa. Since in the previous analysis TINITALY30m has been found as the most accurate DEM to represent the
333
morphology of the mountainous area of the Marche region, we have generated 18 SU combinations based on
334
TINITALY30m to find the optimal SUs partition of AOIa. Figure 6 shows the visual differences in delineation for
335
some of the parameter combinations. Smaller values of circular variance and minimum area result in smaller
336
dimensions of SUs which can restrict heterogeneity between adjacent SUs while, ideally, SUs should maintain
337
external heterogeneity for better terrain representation.
338
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
14
339
Figure 6: SU combinations. 9 out of the 18 combinations are shown to highlight differences as the values of two
340
parameters change, i.e., minimum area and circular variance.
341
Figure 7 reports the behavior of the F, A and D metrics and the final S metric based on the 18 combinations of a
342
and c. Considering that each of the metric represents a goodness for the final SU partition, higher the F, A and D,
343
better is the SU partition. Excluding F, which shows an almost irregular pattern with the maximum at c equal to
344
0.1 and a equal to 40x103m2 (Figure 7.1). A and D have a mutually opposite almost linear pattern which reach a
345
maximum pairing: in A where c is equal to 0.7 and a is each of the values assigned (Figure 7.2), in D with c equal
346
to 0.1 and a equal to 40x103m2 (Figure 7.3). A shows a better performance increasing the mapping unit extension
347
of the study area, whereas D shows better performance with smaller partitions.
348
The product of the normalized metrics results in the S value which is maximized in the range of a between 300x103
349
m2 and 200x103 m2 and by a value of 0.1 for c (Figure 7.4). Therefore, in between the tested combination, c equal
350
to 0.1 and a equal to 300x103 m2 produce the optimal SU partition for landslide susceptibility mapping in the
351
Marche region with a SU extension of 0.40 km2 on average (dataset freely available on Ahmed & Titti 2024). On
352
the contrary the worst-case partition is the one which combines c equal to 150x103 m2 and a equal to 0.7 with a
353
SU extension of 0.84 km2 on average.
354
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
15
355
Figure 7: Behavior of the F, A and D metrics and the final S metric with respect to parameters a and c: (1) shows
356
the F value of SU aspect segmentation metric, (2) visualizes the landslide extension inside a SU (A), (3) shows the
357
landslide density (D) and (4) depicts the results of the final combined metric S.
358
Consequently, the susceptibility assessment with the S -optimal and S-worst case SUs partition has been carried
359
out. The maps resulting from the susceptibility analysis and the relative confusion matrixes based on the S-optimal
360
and the S-worst case SUs delineation of TINITALY30m dataset are represented in Figure 8, while the quality
361
metrics generated from the 10-fold spatial cross validation by ROC analysis are reported in Figure 9. The
362
confusion matrix of the S-optimal delineation (Figure 8B) and of the S-worst case delineation (Figure 8D) report
363
37% of TP (True Positive), 6% of TN (True Negative), 31% of FP (False Positive) and 26% of FN (False Negative)
364
and 45% of TP, 6% of TN, 24% of FP and 25% of FN respectively and performance metrics equal to 0.68 of AUC,
365
0.6 of F1 score, 0.23 of Cohen’s Kappa index and 0.74 of AUC, 0.67 of F1 score, 0.29 of Cohen’s Kappa index,
366
respectively (Figure 9).
367
In addition, two more landslide susceptibility analysis have been carried out using SUs partitions with intermediate
368
S values: c equal to 200x103 m2 and a equal to 0.4, c equal to 40x103 m2 and a equal to 0.1, to investigate the
369
relation between AUC and the number, or extension, of the slope units (see Discussion section).
370
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
16
371
Figure 8: Landslide susceptibility mapping with TINITALY 30m using: A) the selected optimal SU delineation
372
(a=300x103 m2, c=0.1) with the relative confusion matrix (B) (TN 6% of all and 13% of unstable units); C) the selected
373
worst case SU delineation (a=150x103 m2, c=0.7) with the relative confusion matrix (D) (TN 6% of all and 12% of
374
unstable units). Image background from © Google Maps 2019.
375
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
17
376
Figure 9: ROC curve with AUC, F1 score and Kappa coefficient values for 10-fold cross validation. A) the optimal SUs
377
delineation (a=300x103 m2, c=0.1); B) the worst-case SUs delineation (a=150x103 m2, c=0.7).
378
Discussion
379
Based on the results of the quantitative comparison between ALOS, COP, FABDEM, TINITALY10m and
380
TINITALY30m, the latter has performed better than the other DEMs as per the indices used in this study (Figure
381
3). These comparisons are insightful for morphological differences for instance, in regard to roughness indices
382
(Figure 3D), all DEMs tend to oversmooth with respect to the reference DEM. This can be indicative of the spatial
383
support being larger than 30m in reality, meaning that the spatial data density is much lower than the given
384
resolution. It is also interesting to realize the difference between COP and FABDEM. FABDEM being a product
385
of COP (DSM), as a DTM, in essence it should be closer to the LiDAR representation of the terrain with vegetation
386
and buildings removed, but it produces a less accurate output. The efforts of generating a DTM from COP have
387
been motivated in the application of flood modelling trying to optimize the terrain representation, especially in
388
areas of relatively low elevation. However, the algorithm has not been devised for optimizing geomorphometric
389
derivatives such as slope (Hawker et al., 2022). This can be particularly relevant when modelling slope instability.
390
Thus, FABDEM in the region considered does not improve the terrain representation as compared to COP (Bielski
391
et al., 2024). This behavior is visible in Figure 3 where FABDEM shows larger difference distributions than COP
392
for slope, residual DEM and both roughness indices. For instance, in regard to roughness indices (Figure 3D), all
393
DEMs tend to oversmooth with respect to the reference DEM which can be indicative of the spatial support being
394
larger than 30m in reality, meaning that the spatial data density is much lower than the given resolution.
395
ALOS consistently features high differences in all computed metrics against the counter global DEMs which
396
could be explained with the analysis of Caglar et al. (2018). They concluded that ALOS contains a significant
397
number of anomalies in elevation values, possibly attributed to unfiltered sensor noise and processing algorithms
398
which are often not easily identifiable. Nonetheless, ALOS is still ranking well above other global products like
399
SRTM and NASADEM according to quantitative assessments on DEM derived parameters and is still comparable
400
with COP and FABDEM (Bielski et al., 2024; Guth et al., 2024).
401
The numerical comparisons resulting in Figure 3 can be supported by the graphical representation of the slope
402
differences in Figure 10. Although the spatial distribution of differences varies, larger differences are most
403
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
18
noticeable in the ALOS DEM, followed by COP and FABDEM, compared to TINITALY30m, which exhibits
404
fewer differences in slope compared to the reference DEM.
405
406
Figure 10: Difference in slope (degrees) between the four tested DEMs (30m) and the reference LiDAR DEM, by
407
subtracting the LiDAR value from the test DEM value.
408
TINITALY was originally published with a pixel size of 10x10m. Since the pixel sizes of the open global DEMs
409
selected to be compared with the reference DEM in the AOIb area are around 30x30m we have decided to conduct
410
the entire analysis using the same grid-cell size of 30m. Therefore, the original TINITALY10m has been resampled
411
to 30x30m cell size. Despite this, the accuracy of TINITALY10m has been also investigated. Therefore, we have
412
compared the performance of TINITALY30m and of TINITALY10m using normalized differences instead of
413
simple differences. Although this was not the primary aim of the study, the tests indicate that TINITALY at 30m
414
pixel size outperforms the 10m pixel size (Figure 4). These differences in performance, apart from the expected
415
lower uncertainty related to the larger spatial support, may be attributed to the interpolation approach used for
416
TINITALY10m. In areas with low sampling density, noticeable artifacts appear, which can significantly affect the
417
calculation of geomorphometric derivatives. Resampling from the original 10m pixel size to a coarser one (30m)
418
can partially filter out these artifacts. Thus, higher resolution does not necessarily guarantee better results if it is
419
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
19
not supported by high-quality elevation data or if it contains a high number of artifacts (Chen et al., 2020;
420
Mahalingam and Olsen, 2016). Additionally, the use of contour lines as input data along with triangulator for
421
interpolation may result in spurious spikes at regular intervals within elevation zones and in areas with triangular
422
slope-faces (Zingaro et al., 2021). Considering the acquisition dates of DEMs in comparison to the LiDAR,
423
COP30 and ALOS have been surveyed closer to the time of the LiDAR than TINITALY but even so,
424
TINITALY30m has shown better results when compared with the LiDAR. Comparing slope differences in
425
landslide areas across the selected global open-access DEMs, as well as TINITALY10m and TINITALY30m, yield
426
similar results. The graphs in Figure 5 present similar distribution of relative differences in Figure 3 and Figure
427
4.Comparing slope differences in landslide areas across the selected global open-access DEMs, as well as
428
TINITALY10m and TINITALY30m, yield similar results. The graphs in Figure 5 present similar distribution of
429
relative differences in Figure 3 and Figure 4.
430
The similarity between the geometry of delineated SUs with the same parameters, as compared with the ones
431
delineated from the reference DEM, indicates a higher value of the Jaccard Index for TINITALY30m. This means
432
that the SUs delineated using TINITALY30m most closely resemble those from the reference LiDAR DEM. The
433
remaining of the global DEMs also produce SUs with a high similarity index.
434
In the end of Phase 1, we can conclude that for the Marche region, the use 30m resampled TINITALY DEM is
435
recommended for SU definition, therefore the rest of the analysis proposed for Phase 2 has been based on
436
TINITALY30m.
437
Extending the analysis of SU delineation from AOIb, we have used multiple SU parameters for a more detailed
438
analysis in AOIa with landslide polygons. A landslide can be described as a downslope movement of rock mass,
439
earth or debris (Cruden, 1991). Understandably, slope-facing direction and slope angle can be considered as
440
driving factors for slope failures and can be used to dissect the terrain into units which can morphologically
441
describe landslide prone areas. Landslide susceptibility evaluates the probability of occurrence of a landslide
442
according to a set of variables. Susceptibility depends upon a set of variables whose values are associated in a
443
unitary manner to each mapping unit. Therefore, the mapping unit represents a portion of territory that each
444
variable describes numerically by a single value as if it was a point object. Consequently, the smaller the dimension
445
of the map unit, the more representative the single variable is. However, a spatial event such as a landslide, which
446
is a non-point event, does not represent a homogeneous object according to the variables chosen to predict it (i.e.,
447
the degree of slope is not homogeneous throughout the landslide area). Thus, to evaluate the probability of
448
occurrence of this event, it is necessary to identify unique values for each chosen predictor calculated within a
449
portion of territory that coincides as much as possible with the landslide. It is also comprehensible that including
450
stable areas, the portion of territory that most closely resembles the landslide area is the slope-aspect which can
451
be represented by the SU. Therefore, to satisfy both the needs described above, the mapping unit should be as
452
concise as possible to describe the shape of the landslide area.
453
The methodology adopted to evaluate the SU subdivision has been designed to address the forementioned
454
requirements by integrating new metrics, specifically tailored for landslide studies considering the relevance of
455
terrain units with landslide inventories. In addition to the aspect segmentation metric (F) proposed by Alvioli et
456
al. (2016), the landslide extension coefficient (A) and the landslide density coefficient (D) have also been included.
457
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
20
In a way, the F metric can define the shape of the SU on the base of the spatial aspect distribution (Figure 11A
458
and Figure 11B), while a balance between A and D can define the extension of the SU.
459
According to A, the optimal SUs are the ones that contain the entire landslide, with no landslide area falling in
460
adjacent SUs. The landslide coefficient A may not fully capture the extent of landslide area especially when
461
dealing with landslides characterized by high mobility, as in the case of flow-like landslide which can reach
462
considerable distances where the run-out may move out from the homogenous slope-aspect. Nevertheless, the
463
frequency distribution of the landslide classes in the landslide inventory will balance the A value, therefore the
464
run-out of flow-like landslides may have an impact on the SU dimensions if their presence is significative in the
465
inventory. Otherwise, part of the unstable area may fall in the adjacent SUs. Consequently, the larger the SU is,
466
the higher is the probability of including the entire landslide, as is visible in Figure 11C and Figure 11D where an
467
example of the lowest and highest performing SU partition according to A is represented. In contrast to A, the D
468
metric would avoid the overestimation of the SU dimension which should be limited, ideally, to a single landslide
469
(see the example in Figure 11E and Figure 11F). A correct use of D metric requires that reactivated landslides
470
should be excluded and considered as unique events, to avoid doubling the number of polygons in the same spatial
471
unit.
472
The variability of the SU extension with respect to the parameters a and c can also be described through the
473
number of unstable units in relation to the total number of SUs. Figure 12 shows how as D increases and A
474
decreases, the unstable units increase. At the same time as D increases and A decreases, the SU extension is
475
reduced and therefore SU count increases.
476
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
21
477
Figure 11: Selection of SUs partition of a sub-portion of the study area AOIa. A) and B) The SUs partitions with the
478
lowest and the highest value of F respectively; C) and D) The SUs partitions with the lowest and the highest value of A
479
respectively; E) and F) A random selection of SUs partition with F and A values in between the highest and lowest.
480
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
22
481
Figure 12: Evolution of the portion of unstable SUs in the study area with varying values of a and c.
482
All metrics unified in S maximizes their effect, as shown in an example in Figure 13 where the comparable
483
differences explain the concept of the ration between the number and extension of landslides contained in the SUs.
484
While it is difficult to minimize SU area as well as contain the landslide area, it is to be considered that the spatial
485
and areal accuracy of landslide inventories can significantly affect the output since the best terrain partition is
486
interpreted based on the dimensions and number of landslide polygons. In this case study, the PAI of Marche
487
region has been used to test the methodology, and while the landslide inventory plays a crucial role, it has to be
488
mentioned that the dataset may come with limitations. The inventory has not been systematically updated for the
489
mapped landslide areas and the dataset has been updated by reports from scientific literature, local authorities and
490
projects of the municipalities (Costanzo and Irigaray, 2020). Nonetheless, the methodology remains compatible
491
with landslide polygons and SUs supporting the selection of an optimal terrain partitioning.
492
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
23
493
Figure 13: SUs partitions of a sub-portion of the Marche study area (AOIa) compared to landslides distribution from
494
the PAI. A) the SU partition (a: 150x103 m2 and c: 0.7) with lowest value of S, B) the SU partition (a: 300 x103m2 and c:
495
0.1) with highest value of S.
496
Two susceptibility analyses have been carried out selecting the S-optimal and S-worst case SUs partitions. Since,
497
the goal of this study is not to assess landslide susceptibility of the Marche region, but to investigate the potential
498
effect of a thought-out SUs delineation for landslide susceptibility evaluated with largely used metrics such as
499
AUC, F1-score and Cohen’s Kappa score, the predisposing factors selected for the susceptibility analysis are not
500
entirely representative of the geo-environmental conditions. In particular, not all predisposing factors (e.g., land
501
use, vegetation indices and others) have been considered (see also Titti et al 2024). Therefore, the cross-validation
502
results (Figure 9A) of the susceptibility map (Figure 8A) calculated with the optimal SU subdivision are not
503
performing high in the metrics considered (AUC = 0.68, F1 score = 0.6, K = 0.23 on average). Nevertheless, it is
504
interesting to highlight the trend of the relation between the mapping unit extension and the AUC value along with
505
other metrics.
506
AUC is calculated as the integral of the ROC curve. The ROC curve depends on the balance between unstable
507
units and stable units in the training dataset, thus, the higher is the ratio between the number of unstable SUs and
508
the total number of SUs, higher is the AUC because higher is the learning capacity of the model to recognize True
509
Positive mapping units increasing the True Positive Rate value of the ROC curve. In the 18 combinations selected,
510
to investigate the highest-performing a and c values for SUs delineation, we haven’t changed the landslide number
511
but the extension of the SUs whose trend is visible through the number of SUs pattern in Figure 14. Considering
512
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
24
all the combinations of a and c performed in our experiment, the higher the extension of the mapping units, the
513
higher the proportion between the number of unstable units and the number of all the mapping units and higher
514
the AUC (Figure 14). Same considerations can be done for the F1 score, and the Cohen’s Kappa index whose
515
behaviors follow similar trend of the AUC.
516
517
Figure 14: Trend of the number of unstable and total SUs comparison (of 18 combinations for a and c) and the behavior
518
of the metrics resulting from the landslide susceptibility analysis. The parameters (a, c) are labelled along the
519
performance metrics to represent the respective trend.
520
Therefore, at least in the experiments made for this study, the metrics selected are not suitable for comparing
521
susceptibility maps directly because the training datasets are differently balanced. Nevertheless, a comparison
522
between the S-optimal and S-worst case susceptibility maps, as shown in Figure 8A and Figure 8C respectively,
523
can still be made. Graphically, the maps exhibit a similar spatial pattern of landslide probability of occurrence.
524
This is further supported by the fact that the number of True Negative units relative to unstable units is nearly the
525
same, at 13% and 12% for the S-optimal and the S-worst case, respectively. The primary distinction lies in the
526
susceptibility value, which is on average lower in the S-optimal delineation than in the S-worst case This
527
difference is attributed to the overestimation of unstable units in the S-worst case due to the imbalance between
528
stable and unstable units.
529
Conclusions
530
This study encompasses DEM utilization from the viewpoint of fine-scale morphology and terrain sub-division
531
into mapping units in the frame of regional predictive landslide modelling. The aim is to compare freely available
532
global and national DEMs from which morphological landslide predisposing factors and optimized terrain
533
partition in slope units are derived to map landslide susceptibility. Therefore, the investigation initially identified
534
the optimal DEM among the available ones and then selected the optimal SUs partition in the alternative
535
combinations generated.
536
The global DEMs (ALOS, COP, FABDEM) and TINITALY resampled at 30m have shown considerable
537
differences with respect to the reference DEM (an airborne LiDAR resampled at 30m pixel size) in the selected
538
geomorphometric derivatives in AOIb. Concerning the SUs delineation, the TINITALY30m has shown the best
539
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
25
performance thus, it has been selected to generate 18-parameter SUs subdivisions in AOIa. To define the optimal
540
SUs delineation, a novel method has been proposed, which evaluates the SUs alternatives on the base of internal
541
aspect homogeneity/external heterogeneity, landslides numerosity and landslides extension. According to the S
542
metric (Eq. 6), the SUs partition generated with c equal to 0.1 and a equal to 300x103 m2 results in the optimal
543
subdivision, contrasting with c equal to 0.7 and a equal to 150x103 m2 as the worst case one.
544
Ultimately, to understand the effect of the terrain partition on the landslide susceptibility model, we have
545
performed the S-optimal and the S-worst case landslide susceptibility. It is understood that the performance
546
metrics (AUC, F1, K) of the landslide susceptibility models do not necessarily equate with the S metric
547
performance. Indeed, AUC, F1 and K depict opposite trends as compared with the S metric.
548
Though only TINITALY30m has been used in extending the analysis for SU experiments, COP30, as the second-
549
best performing DEM for fine-scale morphology, can also be considered in future studies. A holistic comparison
550
could help evaluate its effectiveness in landslide susceptibility studies. Moreover, since the result of the S-method
551
depends on the landslide inventory, further research would pave the way for space-time inventories performing
552
multi-temporal SUs delineations to reach the best terrain delineation for slope failure prediction. Developing
553
space-time landslide inventories and adapting SUs delineation for dynamic, evolving terrains could significantly
554
enhance the predictive capability of landslide models. Ultimately, continued innovation in DEM selection, SU
555
partitioning methods, and landslide inventory development will contribute to more effective landslide risk
556
management strategies and mitigation efforts.
557
Data availability
558
The optimal SUs partition of the Marche study area (AOIa) is freely available at Ahmed and Titti (2024).
559
Author contributions
560
MA: Conceptualization, Methodology, Formal analysis, Writing - Original Draft; GT: Conceptualization,
561
Methodology, Formal analysis, Writing - Original Draft, Funding acquisition; ST: Methodology, Formal analysis,
562
Writing - Review & Editing; LB: Writing - Review & Editing, Supervision; MF: Writing - Review & Editing,
563
Supervision.
564
Competing interests
565
The authors declare that they have no conflict of interest.
566
Acknowledgements
567
This study was carried out within the RETURN Extended Partnership and received funding from the European
568
Union Next480 Generation EU (National Recovery and Resilience Plan – NRRP, Mission 4, Component 2,
569
Investment 1.3 – D.D. 1243 2/8/2022, PE0000005).
570
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
References
572
Abrams, M., Bailey, B., Tsu, H., and Hato, M.: The ASTER Global DEM, Photogramm Eng
573
Remote Sensing, 76, 344–348, 2010.
574
Ahmed, M. and Titti, G.: Slope Units delineation for Marche region in Italy [Dataset],
575
https://doi.org/https://doi.org/10.5281/zenodo.13769104, 2024.
576
Ahmed, M., Tanyas, H., Huser, R., Dahal, A., Titti, G., Borgatti, L., Francioni, M., and
577
Lombardo, L.: Dynamic rainfall-induced landslide susceptibility: A step towards a unified
578
forecasting system, International Journal of Applied Earth Observation and Geoinformation,
579
125, 103593, https://doi.org/10.1016/J.JAG.2023.103593, 2023.
580
Albertella, A., Barzaghi, R., Carrion, D., and Maggi, A.: The joint use of gravity data and
581
GPS/levelling undulations in geoid estimation procedures, Bollettino di Geodesia e Scienze
582
Affini, 67, 49–59, 2008.
583
Alvioli, M., Marchesini, I., Reichenbach, P., Rossi, M., Ardizzone, F., Fiorucci, F., and Guzzetti,
584
F.: Automatic delineation of geomorphological slope units with r.slopeunits v1.0 and their
585
optimization for landslide susceptibility modeling, Geosci Model Dev, 9, 3975–3991,
586
https://doi.org/10.5194/gmd-9-3975-2016, 2016.
587
Barzaghi, R. and Carrion, D.: Testing EGM2008 in the Central Mediterranean area, in: External
588
Quality Evaluation Reports of EGM08, International Association of Geodesy and International
589
Gravity Field Service, 133–143, 2009.
590
Barzaghi, R., Borghi, A., Carrion, D., and Sona, G.: Refining the estimate of the Italian quasi-
591
geoid, Bollettino di Geodesia e Scienze Affini, 50, 145–159, 2007.
592
Becek, K.: Assessing global digital elevation models using the runway method: The advanced
593
spaceborne thermal emission and reflection radiometer versus the shuttle radar topography
594
mission case, IEEE Transactions on Geoscience and Remote Sensing, 52, 4823–4831,
595
https://doi.org/10.1109/TGRS.2013.2285187, 2014.
596
Bielski, C., Lopez-Vazquez, C., Grohmann, C. H., Guth, P. L., Hawker, L., Gesch, D., Trevisani,
597
S., Herrera-Cruz, V., Riazanoff, S., Corseaux, A., Reuter, H. I., and Strobl, P.: Novel approach
598
for ranking DEMs: Copernicus DEM improves one arc second open global topography, IEEE
599
Transactions on Geoscience and Remote Sensing, 62, 1–22,
600
https://doi.org/10.1109/TGRS.2024.3368015, 2024.
601
Van den Bout, B., Lombardo, L., Chiyang, M., van Westen, C., and Jetten, V.: Physically-based
602
catchment-scale prediction of slope failure volume and geometry, Eng Geol, 284, 105942,
603
https://doi.org/10.1016/J.ENGGEO.2020.105942, 2021.
604
Brenning, A.: Spatial prediction models for landslide hazards: review, comparison and
605
evaluation, Natural Hazards and Earth System Sciences, 5, 853–862,
606
https://doi.org/10.5194/NHESS-5-853-2005, 2005.
607
Brock, J., Schratz, P., Petschko, H., Muenchow, J., Micu, M., and Brenning, A.: The
608
performance of landslide susceptibility models critically depends on the quality of digital
609
elevation models, Geomatics, Natural Hazards and Risk, 11, 1075–1092,
610
https://doi.org/10.1080/19475705.2020.1776403, 2020.
611
26
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
Caglar, B., Becek, K., Mekik, C., and Ozendi, M.: On the vertical accuracy of the ALOS world
612
3D-30m digital elevation model, Remote Sensing Letters, 9, 607–615,
613
https://doi.org/10.1080/2150704X.2018.1453174, 2018.
614
Carrara, A., Cardinali, M., Detti, R., Guzzetti, F., Pasqui, V., and Reichenbach, P.: GIS
615
techniques and statistical models in evaluating landslide hazard, Earth Surf Process Landf, 16,
616
427–445, https://doi.org/10.1002/ESP.3290160505, 1991.
617
Catani, F., Lagomarsino, D., Segoni, S., and Tofani, V.: Landslide susceptibility estimation by
618
random forests technique: Sensitivity and scaling issues, Natural Hazards and Earth System
619
Sciences, 13, 2815–2831, https://doi.org/10.5194/NHESS-13-2815-2013, 2013.
620
Chaplot, V., Darboux, F., Bourennane, H., Leguédois, S., Silvera, N., and Phachomphon, K.:
621
Accuracy of interpolation techniques for the derivation of digital elevation models in relation to
622
landform types and data density, Geomorphology, 77, 126–141,
623
https://doi.org/10.1016/J.GEOMORPH.2005.12.010, 2006.
624
Chen, Z., Ye, F., Fu, W., Ke, Y., and Hong, H.: The influence of DEM spatial resolution on
625
landslide susceptibility mapping in the Baxie River basin, NW China, Natural Hazards, 101,
626
853–877, https://doi.org/10.1007/S11069-020-03899-9/FIGURES/12, 2020.
627
Claessens, L., Heuvelink, G. B. M., Schoorl, J. M., and Veldkamp, A.: DEM resolution effects
628
on shallow landslide hazard and soil redistribution modelling, Earth Surf Process Landf, 30,
629
461–477, https://doi.org/10.1002/ESP.1155, 2005.
630
Corti, M., Francioni, M., Abbate, A., Papini, M., and Longoni, L.: ANALYSIS AND
631
MODELLING OF THE SEPTEMBER 2022 FLOODING EVENT IN THE MISA BASIN,
632
Italian Journal of Engineering Geology and Environment, 69–76,
633
https://doi.org/10.4408/IJEGE.2024-01.S-08, 2024.
634
Costanzo, D. and Irigaray, C.: Comparing Forward Conditional Analysis and Forward Logistic
635
Regression Methods in a Landslide Susceptibility Assessment: A Case Study in Sicily,
636
Hydrology, 7, 37, https://doi.org/10.3390/HYDROLOGY7030037, 2020.
637
Crema, S., Llena, M., Calsamiglia, A., Estrany, J., Marchi, L., Vericat, D., and Cavalli, M.: Can
638
inpainting improve digital terrain analysis? Comparing techniques for void filling, surface
639
reconstruction and geomorphometric analyses, Earth Surf Process Landf, 45, 736–755,
640
https://doi.org/10.1002/ESP.4739, 2020.
641
Cruden, D. M.: A simple definition of a landslide, Bulletin of the International Association of
642
Engineering Geology - Bulletin de l’Association Internationale de Géologie de l’Ingénieur, 43,
643
27–29, https://doi.org/10.1007/BF02590167, 1991.
644
Elia, L., Castellaro, S., Dahal, A., and Lombardo, L.: Assessing multi-hazard susceptibility to
645
cryospheric hazards: Lesson learnt from an Alaskan example, Science of The Total Environment,
646
898, 165289, https://doi.org/10.1016/J.SCITOTENV.2023.165289, 2023.
647
Copernicus Global Digital Elevation Model:
648
Fawcett, T.: An introduction to ROC analysis, Pattern Recognit Lett, 27, 861–874,
649
https://doi.org/10.1016/J.PATREC.2005.10.010, 2006.
650
Fenton, G. A., McLean, A., Nadim, F., and Griffiths, D. V.: Landslide hazard assessment using
651
digital elevation models, Canadian Geotechnical Journal, 50, 620–631,
652
https://doi.org/10.1139/CGJ-2011-0342, 2013.
653
27
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
Florinsky, I. V.: Accuracy of local topographic variables derived from digital elevation models,
654
International Journal of Geographical Information Science, 12, 47–62,
655
https://doi.org/10.1080/136588198242003, 1998.
656
Florinsky, I. V., Skrypitsyna, T. N., Trevisani, S., and Romaikin, S. V.: Statistical and visual
657
quality assessment of nearly-global and continental digital elevation models of Trentino, Italy,
658
Remote Sensing Letters, 10, 726–735, https://doi.org/10.1080/2150704X.2019.1602790, 2019.
659
Gesch, D. B.: Best practices for elevation-based assessments of sea-level rise and coastal
660
flooding exposure, Front Earth Sci (Lausanne), 6,
661
https://doi.org/10.3389/FEART.2018.00230/BIBTEX, 2018.
662
Gesch, D. B., Evans, G. A., Oimoen, M. J., and Arundel, S.: The National Elevation Dataset, in:
663
Digital Elevation MoFDigital Elevation Model Technologies and Applications: The DEM Users
664
Manual, American Society for Photogrammetry and Remote Sensing, 83–110, 2018.
665
Grohmann, C. H.: Evaluation of TanDEM-X DEMs on selected Brazilian sites: Comparison with
666
SRTM, ASTER GDEM and ALOS AW3D30, Remote Sens Environ, 212, 121–133,
667
https://doi.org/10.1016/J.RSE.2018.04.043, 2018.
668
Grohmann, C. H., Smith, M. J., and Riccomini, C.: Multiscale analysis of topographic surface
669
roughness in the Midland Valley, Scotland, IEEE Transactions on Geoscience and Remote
670
Sensing, 49, 1200–1213, https://doi.org/10.1109/TGRS.2010.2053546, 2011.
671
Guisan, A., Weiss, S. B., and Weiss, A. D.: GLM versus CCA spatial modeling of plant species
672
distribution, Plant Ecol, 143, 107–122, https://doi.org/10.1023/A:1009841519580/METRICS,
673
1999.
674
Guth, P. L. and Geoffroy, T. M.: LiDAR point cloud and ICESat-2 evaluation of 1 second global
675
digital elevation models: Copernicus wins, Transactions in GIS, 25, 2245–2261,
676
https://doi.org/10.1111/TGIS.12825, 2021.
677
Guth, P. L., Trevisani, S., Grohmann, C. H., Lindsay, J., Gesch, D., Hawker, L., and Bielski, C.:
678
Ranking of 10 Global One-Arc-Second DEMs Reveals Limitations in Terrain Morphology
679
Representation, Remote Sens (Basel), 16, 3273, https://doi.org/10.3390/RS16173273, 2024.
680
Hawker, L., Neal, J., and Bates, P.: Accuracy assessment of the TanDEM-X 90 Digital Elevation
681
Model for selected floodplain sites, Remote Sens Environ, 232, 111319,
682
https://doi.org/10.1016/J.RSE.2019.111319, 2019.
683
Hawker, L., Uhe, P., Paulo, L., Sosa, J., Savage, J., Sampson, C., and Neal, J.: A 30 m global
684
map of elevation with forests and buildings removed, Environmental Research Letters, 17,
685
https://doi.org/10.1088/1748-9326/AC4D4F, 2022.
686
Hiller, J. K. and Smith, M.: Residual relief separation: digital elevation model enhancement for
687
geomorphological mapping, Earth Surf Process Landf, 33, 2266–2276,
688
https://doi.org/10.1002/ESP.1659, 2008.
689
Huang, F., Teng, Z., Guo, Z., Catani, F., and Huang, J.: Uncertainties of landslide susceptibility
690
prediction: Influences of different spatial resolutions, machine learning models and proportions
691
of training and testing dataset, Rock Mechanics Bulletin, 2,
692
https://doi.org/10.1016/J.ROCKMB.2023.100028, 2023.
693
Isaaks, E. H. and Srivastava, R. Mohan.: Applied geostatistics, Oxford University Press,
694
London, 1989.
695
28
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
Jaccard, P.: Etude comparative de la distribution florale dans une portion des Alpes et du Jura,
696
142nd ed., edited by: Bulletin de la Société vaudoise des sciences naturelles, Impr. Corbaz,
697
1901.
698
ALOS World 3D - 30m. V3.2: Distributed by OpenTopography.
699
https://doi.org/10.5069/G94M92HB. Accessed: 2024-04-16, last access: 18 March 2024.
700
Jarvis, A., Guevara, E., Reuter, H. I., and Nelson, A. D.: Hole-filled SRTM for the globe :
701
version 4, https://doi.org/10.2/JQUERY.MIN.JS, 2008.
702
Kakavas, M., Kyriou, A., and Nikolakopoulos, K. G.: Assessment of freely available DSMs for
703
landslide-rockfall studies, in: Proceedings of SPIE, https://doi.org/10.1117/12.2573604, 2020.
704
Kamiński, M.: The Impact of Quality of Digital Elevation Models on the Result of Landslide
705
Susceptibility Modeling Using the Method of Weights of Evidence, Geosciences (Basel), 10,
706
https://doi.org/10.3390/GEOSCIENCES10120488, 2020.
707
Karakas, G., Kocaman, S., and Gokceoglu, C.: On the effect of dem quality for landslide
708
susceptibility mapping, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial
709
Information Sciences, V-3–2022, 525–531, https://doi.org/10.5194/ISPRS-ANNALS-V-3-2022-
710
525-2022, 2022.
711
Kraemer, H. C.: Kappa Coefficient, Wiley StatsRef: Statistics Reference Online, 1–4,
712
https://doi.org/10.1002/9781118445112.STAT00365.PUB2, 2015.
713
Liu, K., Song, C., Ke, L., Jiang, L., Pan, Y., and Ma, R.: Global open-access DEM performances
714
in Earth’s most rugged region High Mountain Asia: A multi-level assessment, Geomorphology,
715
338, 16–26, https://doi.org/10.1016/J.GEOMORPH.2019.04.012, 2019.
716
Loche, M., Alvioli, M., Marchesini, I., and Lombardo, L.: Landslide Susceptibility within the
717
binomial Generalized Additive Model, in: European Geosciences Union (EGU23),
718
https://doi.org/10.13140/RG.2.2.14089.62565, 2023.
719
Lombardo, L. and Tanyas, H.: Chrono-validation of near-real-time landslide susceptibility
720
models via plug-in statistical simulations, Eng Geol, 278,
721
https://doi.org/10.1016/J.ENGGEO.2020.105818, 2020.
722
Lombardo, L., Bakka, H., Tanyas, H., van Westen, C., Mai, P. M., and Huser, R.: Geostatistical
723
Modeling to Capture Seismic-Shaking Patterns From Earthquake-Induced Landslides, J
724
Geophys Res Earth Surf, 124, 1958–1980, https://doi.org/10.1029/2019JF005056, 2019.
725
Lombardo, L., Opitz, T., Ardizzone, F., Guzzetti, F., and Huser, R.: Space-time landslide
726
predictive modelling, https://doi.org/10.1016/j.earscirev.2020.103318, October 2020.
727
Mahalingam, R. and Olsen, M. J.: Evaluation of the influence of source and spatial resolution of
728
DEMs on derivative products used in landslide mapping, Geomatics, Natural Hazards and Risk,
729
7, 1835–1855, https://doi.org/10.1080/19475705.2015.1115431, 2016.
730
Meadows, M., Jones, S., and Reinke, K.: Vertical accuracy assessment of freely available global
731
DEMs (FABDEM, Copernicus DEM, NASADEM, AW3D30 and SRTM) in flood-prone
732
environments, Int J Digit Earth, 17, https://doi.org/10.1080/17538947.2024.2308734, 2024.
733
Muralikrishnan, S., Pillai, A., Narender, B., Reddy, S., Venkataraman, V. R., and Dadhwal, V. K.:
734
Validation of Indian National DEM from Cartosat-1 Data, Journal of the Indian Society of
735
Remote Sensing, 41, 1–13, https://doi.org/10.1007/S12524-012-0212-9/FIGURES/13, 2013.
736
29
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
Osama, N., Shao, Z., and Freeshah, M.: The FABDEM Outperforms the Global DEMs in
737
Representing Bare Terrain Heights, Photogramm Eng Remote Sensing, 89, 613–624,
738
https://doi.org/10.14358/PERS.23-00026R2, 2023.
739
Pawluszek, K. and Borkowski, A.: Impact of DEM-derived factors and analytical hierarchy
740
process on landslide susceptibility mapping in the region of Rożnów Lake, Poland, Natural
741
Hazards, 86, 919–952, https://doi.org/10.1007/S11069-016-2725-Y/FIGURES/26, 2017.
742
Pirasteh, S. and Li, J.: Landslides investigations from geoinformatics perspective: quality,
743
challenges, and recommendations, Geomatics, Natural Hazards and Risk, 8, 448–465,
744
https://doi.org/10.1080/19475705.2016.1238850, 2017.
745
Polidori, L. and Hage, M. El: Digital Elevation Model Quality Assessment Methods: A Critical
746
Review, Remote Sens (Basel), 12, https://doi.org/10.3390/RS12213522, 2020.
747
Pulighe, G. and Fava, F.: DEM extraction from archive aerial photos: accuracy assessment in
748
areas of complex topography, Eur J Remote Sens, 46, 363–378,
749
https://doi.org/10.5721/EuJRS20134621, 2013.
750
Purinton, B. and Bookhagen, B.: Validation of digital elevation models (DEMs) and comparison
751
of geomorphic metrics on the southern Central Andean Plateau, Earth Surface Dynamics, 5,
752
211–237, https://doi.org/10.5194/ESURF-5-211-2017, 2017.
753
Qin, C. Z., Bao, L. L., Zhu, A. X., Wang, R. X., and Hu, X. M.: Uncertainty due to DEM error in
754
landslide susceptibility mapping, International Journal of Geographical Information Science, 27,
755
1364–1380, https://doi.org/10.1080/13658816.2013.770515, 2013.
756
Qiu, H., Zhu, Y., Zhou, W., Sun, H., He, J., and Liu, Z.: Influence of DEM resolution on
757
landslide simulation performance based on the Scoops3D model, Geomatics, Natural Hazards
758
and Risk, 13, 1663–1681, https://doi.org/10.1080/19475705.2022.2097451, 2022.
759
Reichenbach, P., Rossi, M., Malamud, B. D., Mihir, M., and Guzzetti, F.: A review of
760
statistically-based landslide susceptibility models, Earth Sci Rev, 180, 60–91,
761
https://doi.org/10.1016/J.EARSCIREV.2018.03.001, 2018.
762
Riley, S. J., Degloria, S. D., and Elliot, S. D.: A Terrain Ruggedness Index that Quantifies
763
Topographic Heterogeneity, Intermountain Journal of Sciences, 5, 23–27, 1999.
764
Saleem, N., Enamul Huq, M., Twumasi, N. Y. D., Javed, A., and Sajjad, A.: Parameters Derived
765
from and/or Used with Digital Elevation Models (DEMs) for Landslide Susceptibility Mapping
766
and Landslide Risk Assessment: A Review, ISPRS Int J Geoinf, 8, 545,
767
https://doi.org/10.3390/IJGI8120545, 2019.
768
Schlögel, R., Marchesini, I., Alvioli, M., Reichenbach, P., Rossi, M., and Malet, J. P.: Optimizing
769
landslide susceptibility zonation: Effects of DEM spatial resolution and slope unit delineation on
770
logistic regression models, Geomorphology, 301, 10–20,
771
https://doi.org/10.1016/J.GEOMORPH.2017.10.018, 2018.
772
Singhal, A.: Modern Information Retrieval: A Brief Overview, IEEE Data Eng. Bull., 24, 35–43,
773
2001.
774
Steger, S., Moreno, M., Crespi, A., Zellner, P. J., Gariano, S. L., Brunetti, M. T., Melillo, M.,
775
Peruccacci, S., Marra, F., Kohrs, R., Goetz, J., Mair, V., and Pittore, M.: Deciphering seasonal
776
effects of triggering and preparatory precipitation for improved shallow landslide prediction
777
using generalized additive mixed models, Natural Hazards and Earth System Sciences, 23,
778
1483–1506, https://doi.org/10.5194/NHESS-23-1483-2023, 2023.
779
30
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
Strobl, P. A., Bielski, C., Guth, P. L., Grohmann, C. H., Muller, J. P., López-Vázquez, C., Gesch,
780
D. B., Amatulli, G., Riazanoff, S., and Carabajal, C.: The digital elevation model
781
intercomparison experiment demix, a community-based approach at global dem benchmarking,
782
International Archives of the Photogrammetry, Remote Sensing and Spatial Information
783
Sciences - ISPRS Archives, 395–400, https://doi.org/10.5194/ISPRS-ARCHIVES-XLIII-B4-
784
2021-395-2021, 2021.
785
Takaku, J., Tadono, T., and Tsutsui, K.: Generation of high resolution global DSM from ALOS
786
PRISM, International Archives of the Photogrammetry, Remote Sensing and Spatial Information
787
Sciences - ISPRS Archives, 40, 243–248, https://doi.org/10.5194/isprsarchives-XL-4-243-2014,
788
2014.
789
Tarquini, S., Isola, I., Favalli, M., Mazzarini, F., Bisson, M., Pareschi, M. T., and Boschi, E.:
790
TINITALY/01: a new Triangular Irregular Network of Italy, Annals of Geophysics, 50, 407–425,
791
https://doi.org/10.4401/ag-4424, 2007.
792
Tarquini, S., Isola, I., Favalli, M., Battistini, A., and Dotta, G.: TINITALY, a digital elevation
793
model of Italy with a 10 meters cell size (Version 1.1) [Data set],
794
https://doi.org/https://doi.org/10.13127/tinitaly/1.1, 2023.
795
Titti, G., Borgatti, L., Zou, Q., Cui, P., and Pasuto, A.: Landslide susceptibility in the Belt and
796
Road Countries: continental step of a multi-scale approach, Environ Earth Sci, 80, 1–18,
797
https://doi.org/10.1007/S12665-021-09910-1/FIGURES/11, 2021a.
798
Titti, G., van Westen, C., Borgatti, L., Pasuto, A., and Lombardo, L.: When Enough Is Really
799
Enough? On the Minimum Number of Landslides to Build Reliable Susceptibility Models,
800
Geosciences (Basel), 11, 469, https://doi.org/10.3390/GEOSCIENCES11110469, 2021b.
801
Titti, G., Sarretta, A., Lombardo, L., Crema, S., Pasuto, A., and Borgatti, L.: Mapping
802
Susceptibility With Open-Source Tools: A New Plugin for QGIS, Front Earth Sci (Lausanne),
803
10, 842425, https://doi.org/10.3389/FEART.2022.842425/BIBTEX, 2022.
804
Trevisani, S. and Cavalli, M.: Topography-based flow-directional roughness: Potential and
805
challenges, Earth Surface Dynamics, 4, 343–358, https://doi.org/10.5194/ESURF-4-343-2016,
806
2016.
807
Trevisani, S. and Rocca, M.: MAD: robust image texture analysis for applications in high
808
resolution geomorphometry, Comput Geosci, 81, 78–92,
809
https://doi.org/10.1016/J.CAGEO.2015.04.003, 2015.
810
Trevisani, S., Teza, G., and Guth, P.: A simplified geostatistical approach for characterizing key
811
aspects of short-range roughness, Catena (Amst), 223, 106927,
812
https://doi.org/10.1016/J.CATENA.2023.106927, 2023a.
813
Trevisani, S., Skrypitsyna, T. N., and Florinsky, I. V.: Global digital elevation models for terrain
814
morphology analysis in mountain environments: insights on Copernicus GLO-30 and ALOS
815
AW3D30 for a large Alpine area, Environ Earth Sci, 82, https://doi.org/10.1007/s12665-023-
816
10882-7, 2023b.
817
Trevisani, S., Teza, G., and Guth, P. L.: Hacking the topographic ruggedness index,
818
Geomorphology, 439, 108838, https://doi.org/10.1016/J.GEOMORPH.2023.108838, 2023c.
819
Trigila, A., Iadanza, C., and Spizzichino, D.: Quality assessment of the Italian Landslide
820
Inventory using GIS processing, Landslides, 7, 455–470, https://doi.org/10.1007/S10346-010-
821
0213-0/FIGURES/12, 2010.
822
31
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
Van Westen, C. J., Rengers, N., Terlien, M. T. J., and Soeters, R.: Prediction of the occurrence of
823
slope instability phenomena through GIS-based hazard zonation, Geologische Rundschau, 86,
824
404–414, https://doi.org/10.1007/S005310050149/METRICS, 1997.
825
Wilson, J. P. and Gallant, J. C. (Eds.): Terrain analysis : principles and applications, John Wiley
826
& Sons, Inc, 2000.
827
Zhang, K., Gann, D., Ross, M., Robertson, Q., Sarmiento, J., Santana, S., Rhome, J., and Fritz,
828
C.: Accuracy assessment of ASTER, SRTM, ALOS, and TDX DEMs for Hispaniola and
829
implications for mapping vulnerability to coastal flooding, Remote Sens Environ, 225, 290–306,
830
https://doi.org/10.1016/J.RSE.2019.02.028, 2019.
831
Zingaro, M., La Salandra, M., Colacicco, R., Roseto, R., Petio, P., and Capolongo, D.:
832
Suitability assessment of global, continental and national digital elevation models for
833
geomorphological analyses in Italy, Transactions in GIS, 25, 2283–2308,
834
https://doi.org/10.1111/TGIS.12845, 2021.
835
836
837
32
https://doi.org/10.5194/nhess-2024-211
Preprint. Discussion started: 25 November 2024
c
Author(s) 2024. CC BY 4.0 License.
