Analyzing Multitemporal Datasets to Monitor Topographic Changes

in Rio Cucco Italy

Jad Ghantous

, Vincenzo Di Pietra

, Elena Belcore

and Nives Grasso

DIATI, Politecnico di Torino, C.so Duca Degli Abruzzi, Turin, Italy

Keywords: Natural Hazards, Digital Terrain Model, Point Cloud, Semantic Segmentation, Deep Learning.

Abstract: Rio Cucco is an Italian catchment located in Malborghetto of Friuli Venezia Giulia. It is considered an area

of interest regarding its hydrological and morphological properties. The area has historically been affected by

natural hazards such as rockfall and landslides, mainly related to extreme rainfall events like the 2003 storm

that affected the Fella river or the Vaia storm of October 2018. These events highlight the importance of

understanding the morphological and topographic modification of the area also in relation to the realization

of protection and hydraulic works. The changes in Rio Cucco were documented by comparing open-source

historical data with ad-hoc UAV surveys focusing the analysis on 3D products like point clouds and the digital

terrain models. The source of the recent data was an Aerial LiDAR- based survey conducted by our team in

June 2024 while the historical data was taken from the FVG region’s geoportal and referred to 2017. After

comparing the different datasets with traditional techniques like nearest neighbour Euclidean distance or DEM

of Difference, changes were evident pointing to potential rockfalls between the year 2024 and 2017. A deep

learning model was explored and in development for the semantic segmentation of the area.

1 INTRODUCTION

Natural hazards are influenced by climatic conditions.

Climate change affects the frequency and intensity of

hazards such as landslides by modifying the

hydrological regimes (Schneiderbauer et al., 2021). In

a recent study by Semnani et al. (2025), landslide

susceptibility was modeled across California based on

different climate scenarios. They found that there was

an overall increase in susceptibility in respect to time

until the year 2100 in areas such as Sierra Nevada and

the areas near the coast of California. Several areas

have experienced landslides over the years. A

landslide in Mount Meager Canada occurred in 2010

triggered by three decades of temperature increases

that degraded the permafrost (Huggel et al., 2012). In

the case of Italy, Monte Rosa was affected by a

landslide in 2007 as a result of permafrost degradation

as well (Huggel et al., 2012). Another example, is the

Esino River basin in central Italy which was studied

based on the relationship between climate change and

https://orcid.org/

0009-0001-3817-3578

https://orcid.org/

0000-0001-7501-1183

https://orcid.org/

0000-0002-3592-9384

https://orcid.org/0000-0002-9548-6765

landslides (Sangelantoni et al., 2018). The frequent

landslides prompted the creation of several research

projects to study and monitor the relationship between

the areas at risk of natural hazards and their

geomorphological properties. Among these projects,

PRIN funded MORPHEUS (GeoMORPHomEtry

through Scales for a resilient landscape), which this

case study was part of, aims to involve several

research bodies to collaborate on the relationship

between natural hazards and the sediment connectivity

of the area. Sediment connectivity refers to the degree

of linkage that controls sediment fluxes through a

landscape, especially between the sediment source

and the downstream area (Cavalli et al., 2013)

Sediment connectivity is especially important in

mountainous catchments. The presence of natural and

anthropic structures (roads, terraces, dams) greatly

affects the connectivity pathways. This as a result

impacts the estimation of natural hazards. Therefore,

morphological variations derived from multi-temporal

topographic information are fundamental to

Ghantous, J., Di Pietra, V., Belcore, E. and Grasso, N.

Analyzing Multitemporal Datasets to Monitor Topographic Changes in Rio Cucco Italy.

DOI: 10.5220/0013479100003935

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2025), pages 103-110

ISBN: 978-989-758-741-2; ISSN: 2184-500X

103

implement functional sediment connectivity analysis.

In this context, a UAV survey represents the main

source of high temporal resolution topographic data

allowing the generation of several 3D products from

LiDAR or Photogrammetric processing. Among these

products, Digital Elevation Models are fundamental to

derive structural connectivity parameters while DEMs

of Difference are used to estimate the impact of

topographic changes. Therefore, a multi-temporal

analysis is crucial to successfully achieve the goal of

the Morpheus project. In the case of this project, the

regions particularly studied in the framework of

sediment connectivity include Liguria, Veneto and

Friuli Venezia Giulia. The selected study areas

encompass several catchments located in contrasting

landscapes of northern Italy and featuring different

sediment transport processes. Among the many, Rio

Cucco basin (0.65 Km

) saw an extreme event in 2003

that triggered an unusually large debris flow in the

area.

In order to test the multi-source, multi-scale and

multi-temporal coherence of spatial heterogeneous

data, fundamental for assessing the impact of

topographic changes, a UAV- based survey of Rio

Cucco was conducted by our team in June 2024. The

presence of a natural pine forest over much of the

basin surface directed us toward the choice of a

LiDAR sensor for dense point cloud acquisition and

DEM production. After post-processing and analysis

of these datasets, a comparison was needed with the

same type of data provided by EAGLE FVG (official

geo portal of Friuli Venezia Giulia). The aim from this

comparison is to document the temporal changes

occurring in Rio Cucco and to harmonize different

datasets referred to the same study area. To this

purpose, the data will be analysed and validated in

terms of point clouds density, georeferencing,

accuracy, and uncertainties by exploiting data co-

registration algorithms. Furthermore,

to assure the

optimal output for the geomorphometric approach, a

land cover classification procedure was tested relying

on a simple deep learning model.

2 STUDY AREA

The Rio Cucco basin is a torrential watercourse

located in the municipality of Malborghetto-Valbruna

(Udine, Italy) where it serves as a right tributary of the

Fella River. Its watershed comprises two sub-basins,

fed by smaller channels and two main branches that

converge into an alluvial fan near the village of Cucco.

The basin's geology features dolomites and dolomitic

limestones from the Anisian Ladinian period

(Dolomia dello Sciliar), along with nodular

limestones, marls, and calcarenites in its upper areas.

The southern slopes of Monte Cucco are marked by

vertical rocky cliffs, frequent rockfalls, and debris

accumulations, particularly in regions dominated by

friable calcareous and dolomitic rocks prone to

thermoclastism and cryoclastism.

The alluvial fan, spanning roughly 0.2 km², hosts

a natural pine forest of black pine and Scots pine in its

apex and residential areas, meadows, and a state road

bridge at its lower end. The predominantly south-

facing aspect, combined with arid conditions caused

by the area's permeability and frequent material

deposition, inhibits soil formation. Annual

precipitation averages 1,500 mm, with summer and

autumn peaks corresponding to increased debris flow

activity. Significant flood events, such as those in June

1996 and August 2003, caused extensive sediment

mobilization and damage, with the latter involving

325 mm of rainfall over 12 hours and the transport of

approximately 80,000 m³ of sediment.

After the 2003 flood, more than 300 million euros

were invested throughout the Valcanale and Canal del

Ferro to secure, restore existing hydraulic works as

well as build new structures on the most affected

basins (FVG Region, 2007). Two branches of the

basin were extensively reworked with the construction

of cross works and berms, which converge to a

downstream storage basin of 100,000 m

. The latter,

in particular, is protected by two cyclopean boulder

embankments, joined by a filtering weir that conveys

the solid-liquid material to a reinforced concrete gutter

near the State Road crossing, which is followed by

another channelling work in cemented boulders to the

point of discharge into the Fella River. The basin plays

a vital role in mitigating flood risks, directing

sediment-laden flows into reinforced concrete

channels towards the Fella River. Figure 1 shows the

location of the basin with respect to the surrounding

area.

Figure 1: Rio Cucco Basin.

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

104

3 METHODOLOGIES

In the case of Rio Cucco, the area was studied by

comparing FVG 2017 data with the survey that was

conducted in the period between 11 and 14 June 2024.

The EAGLE FVG survey occurred in 2017 thus

all the data including the point clouds and digital

terrain model refer to that survey. There were two

aspects that were considered for the comparison. The

point clouds and digital terrain models were used in

the main comparison criteria. Moreover, after the

comparison between the datasets, a deep learning

models was explored to

assure the optimal output for

the geomorphometric approach, main goal of the

MORPHEUS Project.

3.1 Point Clouds

A DJI Matrice 300 RTK drone equipped with a DJI

Zenmuse L1 Lidar sensor was used to acquire the

point cloud in the June 2024 survey. Consequently,

DJI Terra software was used to process the data from

the drone and create the point clouds of the area (DJI,

n.d.). The point cloud that was output from the

software was a merged file that contained all the point

clouds of the area surveyed and individual point cloud

segments that make up the merged point cloud. Both

the merged and the individual point clouds were

saved. Table 1 shows the different parameters used in

the production of the point clouds with DJI Terra.

The output merged point cloud .las file was 32.6

GB in size which made the file demanding from a

computational point of view. Spatial subsampling

was used in CloudCompare to reduce the density and

size of the point clouds to a workable size (Girard

eau-Montaut, n.d.) .

The point clouds were subsampled at spaces of:

• 0.50 (size: 0.4 GB)

• 0.25 (size: 1.9 GB)

• 0.20 (size: 2.3 GB)

• 0.15 (size: 3.8 GB)

• 0.10 (size: 7.1 GB)

Spatial subsampling has an inverse relationship with

cloud density. A larger spatial subsampling leads to

higher workability but less data. The aim was to find

a workable point cloud with a point cloud density

greater than 16 points/m

Table 1: DJI Terra Software Parameters.

DJI Terra

Point Cloud O

timization Parameters

Point Cloud Density (by

Percentage)

High

Point Cloud Effective

Distance Range

3 – 300 m

Optimise Point Cloud

Accurac

Yes

Smooth Point Clou

Yes

Point Cloud Output Parameters

Ground Point

Classification

Yes

Ground Point

Classification Parameters

Ground Point Type: Steep

Slope

Maximum Diagonal of

the Building: 25 m

Iteration angle: 9°

Iteration distance: 0.6

DEM Parameters By GSD ｜0.5 m

Point Cloud Format PNTS ｜ LAS

Merged Output Yes

LiDAR Point Cloud

Block Count

Output Coordinate

System

WGS 84 / UTM zone 33N

| Default

The EAGLE FVG point clouds refer to the

acquisition conducted in the area in 2017 and had

RDN 2008 UTM 33N as the default reference system.

The comparison conducted between the recent data in

2024 and the EAGLE FVG 2017 data was based on

two main parameters. The surface density of the point

clouds and the Cloud to Cloud distance were used as

the comparison parameters of this study taking the

same reference system into account.

3.2 Digital Terrain Models

The digital terrain model (DTM) produced using DJI

Terra was based on its classification of ground points

and non-ground points according to its own

algorithm. The parameters used for the creation of the

DTM were mentioned in Table 1(regarding the slope

steepness, maximum diagonal of the building etc.).

The output DTM was a digital terrain model with

removal of canopy and a resolution of 0.5 m as set in

the DJI Terra. If no classification of ground points

was selected, the result would be a digital surface

model that has no separation between ground points

and canopy. It is to be noted that the parameters were

tweaked to produce the most connected DTM.

Keeping the default values they resulted in a heavily

tiled DTM that had empty spaces between different

tiles. Moreover, the subsampled point cloud with the

0.25 spatial spacing was input into Agisoft Metashape

Analyzing Multitemporal Datasets to Monitor Topographic Changes in Rio Cucco Italy

105

to produce a DTM with the same resolution of 0.5 m

(Agisoft LLC, n.d.).Using Agisoft Metashape to

create a digital terrain model from the produced point

cloud serves as an alternative to DJI Terra to see how

different algorithms behave in digital terrain model

generation.

After the comparison between the digital terrain

models produced using DJI Terra & Agisoft

Metashape, the next step was comparing the digital

terrain models to EAGLE FVG digital terrain models

for the year 2017. The files on the EAGLE FVG

website were in ascii format and yielded several tiles

that were merged and clipped to the shape of the Rio

Cucco basin. As previously conducted with the

comparison of the point clouds, the 2024 DTM was

converted to the reference system of the EAGLE

FVG 2017 DTM (RDN 2008 UTM 33N) to create a

Difference of Digital Elevation between the two

digital terrain models.

3.3 Land Cover Classification Using a

Deep Learning Model

An important task for this project is conducting land

cover classification of the Rio Cucco basin so that

land cover and land use can be quantified and

monitored for the area. Land cover classification is

particularly useful to monitor natural hazards. There

have been studies regarding the usage of land cover

classification in risk management. UAV very high

resolution data was used in the classification of the

Niger River (Belcore et al., 2022). The authors used a

multitude of geomatic techniques to process the data

and enrich feature extraction based on spectral,

textural and elevation data acquired from the survey.

This allowed for the creation of different classes

necessary for the food map creation to an area that

had very little coverage. Covering another work by

these authors, mountainous areas in Italy studied in

land cover classification using Google Earth Engine.

That study addressed the challenges of accurately

classifying mountainous areas due to the variability

of reflectance and shadows (Belcore et al., 2020). A

land cover classification is required for Rio Cucco

and as a requirement of the project MORPHEUS,

artificial intelligence techniques need to be explored

for automatic classification of the study area. Deep

learning, specifically through the use of convolutional

neural networks (CNNs), offers a powerful approach

to semantic segmentation of satellite imagery (Zhang

et al., 2019). This involves classifying each pixel in

an image into a specific category or class (e.g.,

building, road, forest). Consequently, an open-source

semantic segmentation model was used to conserve

time as well as explore the feasibility of deep learning

applications for this work.. The model DeepLabV3

ResNet50 model was developed as an open-source

library on Python. As this model is not particularly

only made for satellite images and remote sensing, the

model needed training. The model was trained with a

Kaggle dataset of satellite images that contains the

images alongside their masks (Kaggle, n.d.). This

dataset itself is compiled from three main datasets

labeled: Semantic segmentation of aerial imagery

(Roia Foundation), Land Cover Classification -

Bhuvan Satellite Data (Indian Space Research

Organization) , Urban Segmentation (International

Society for Photogrammetry and Remote Sensing).

The dataset contained classes shown in Table 2.

These classes were used as they are and in their same

color in the training and testing of the data to see the

effectiveness of the model.

Table 2: Dataset Default Classes.

Class Colo

Buildin

Dark Pur

Land / Unpaved area Light Purple

Roa

Light Blue

etation Yellow

Wate

Oran

Unlabele

Gra

Table 3 summarizes the different parameters used in

the deep learning model training.

Table 3: Deep Learning Model Parameters

Paramete

Value Descri

tion

Batch Size 8

Number of images

processed in parallel

during training

Image

Dimensions

256 x 256

Resized dimensions of

ut ima

Number of

Classes

Number of distinct

classes

Model

Architecture

DeepLabV3

(ResNet50)

The backbone used for

feature extraction and

mentation

Pretrained

Weights

COCO_WI

TH_VOC_L

ABELS

Weights pretrained on

the COCO dataset

Dataset Size 203 images

Total number of image-

mask

airs

Training

Epochs

400 epochs

Total number of epochs

used in training

Input

Channels

3 (RGB)

Number of channels in

the input image

Output

Channels

Number of

segmentation classes

redicted by the model

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

106

After training the model with that dataset for more

than 400 iterations, a new orthophoto was input to test

the model. The orthophoto for the area was compiled

by converting the subsampled 0.25 point cloud in

Agisoft Metashape into create an orthophoto. With

respect to the 2017 orthophoto, the data was

downloaded from EAGLE FVG. Both datasets were

resampled to be 0.5 m in resolution. The main

objective of these orthophotos was to create testing

for the deep learning model being developed for this

project.

4 RESULTS

Comparing the different datasets for Rio Cucco, it is

evident that were changes in the area from 2017 to

2024.The changes were particularly evident when the

difference of digital elevation (DoD) was calculated.

There were areas where there was possible sediment

deposition leading to an elevation increase at the

location of the deposition and an elevation decrease

where the sediments disconnected. This indicates that

a rockfall possibly occurred between 2017 and 2024.

4.1 Point Clouds

4.1.1 Surface Density

Based on several iterations to choose the best

subsample with respect to its usability and its

resolution. It was found that the 0.15 spatial

subsample model was the most preferable serving a

high ratio between resolution and workability. This

point cloud had more than 95 % of its density greater

than 16 points/m

but it was light enough to run

without causing any issues. Since the point clouds

produced during the June 2024 survey were larger

than the basin. The selected point cloud would be then

trimmed based on the boundaries of the basin and

used to directly compare with the EAGLE FVG 2017

points cloud based on the same area of interest in the

basin.

Table 4: Surface Density of Rio Cucco Point Cloud

Subsamples 2024.

Surface Densit

oints/m

)

Clou

µ σ Min Max

% of pts >

ts/m

0.5 10.8 6.1 0.32 21.2 25.0

0.25 54.9 31.6 0.32 109.4 85.6

0.2 94.2 54.4 0.32 188.0 91.4

0.15 171.9 99.4 0.32 343.4 95.3

0.1 390.3 226.0 0.32 780.3 97.7

Based on the results from Table 5, both point clouds

show very similar surface density with both having

more than 95 % of their points being significantly

higher than the required 16 points/m

Table 5: Surface Density of Clipped Rio Cucco Point Cloud

FVG 2017 and 0.15 Subsample 2024.

Surface Density (points/m

)

Clou

µ σ Min Max

% of pts >

16 pts/m

2017 190.1 110.0 0.32 379.9 95.7

2024 171.4 99.2 0.32 342.4 95.3

4.1.2 Cloud to Cloud Distance

Next, to accurately check the differences between the

point cloud of 2017 and the one created in June 2024,

the boundaries between both point clouds should be

identical and the reference system should also be

identical. The reference system that should be worked

on is RDN2008 UTM 33N, which is the same

reference system used by EAGLE FVG.

Furthermore, the Cloud-to-Cloud distance was

performed before and after ICP by placing the

EAGLE FVG point cloud as the reference option and

the Rio Cucco 2024 0.15 Subsample point cloud as

the compared option in each scenario. The Cloud to

Cloud Distance helps in identifying what the

differences between both point clouds were during

the different time periods of the area.

Table 6 shows the differences in the Cloud to

Cloud distance between the same point clouds from

June 2024 saved in different formats and the 2017

EAGLE FVG point cloud.

Table 6: Cloud to Cloud Distance Between the 2017

EAGLE FVG Point Clouds and 0.15 Subsample 2024.

Cloud 2 Cloud Distance

(

)

Clou

µ σ Min Max

% of pts

< 1

Before ICP

RDN 3.7 2.1 0.00 7.4 13.7

After ICP

(

85 % Final Overla

RDN 3.6 2.1 0.00 7.2 14.1

As seen from Table 6, performing an ICP slightly

reduces the nearest neighbor distance between the

two point clouds. Observing that table, it is possible

to notice a decrease in the mean distance of about 10

cm with the same standard deviation, which means

that both EAGLE FVG data and the authors data are

been produced with the same accuracy in

georeferencing. It is noticeable that the mean distance

between the point clouds is more than three meters.

Analyzing Multitemporal Datasets to Monitor Topographic Changes in Rio Cucco Italy

107

This would be better explained in the next section

when the difference of digital elevation is calculated.

4.2 Digital Terrain Models

Regarding the comparisons between the DJI Terra

DTM, the Agisoft Metashape DTM, and the EAGLE

FVG, the analysis showed that it was favourable to

use the DJI Terra DTM to compare with the EAGLE

FVG DTM. Regarding the comparisons between the

DJI Terra DTM, the Agisoft Metashape DTM, and

the EAGLE FVG, the analysis showed that it was

favourable to use the DJI Terra DTM to compare with

the EAGLE FVG DTM. This was due to having little

difference between the Agisoft Metashape and DJI

Terra DTM, but the Agisoft Metashape DTM

interpolated areas that were not acquired in the point

cloud. Also, the DJI Terra DTM had an advantage of

containing the data from all the points instead of a

subsampled point cloud.

Consequently, a Difference of Digital Elevation

(DoD) was performed between the DJI Terra DTM of

the year 2024 and the EAGLE FVG DTM of the year

2017.

Figure 2: Difference of Digital Elevation (DoD) Between

DJI Terra 2024 and EAGLE FVG 2017 DTMs.

Clearly, there were differences between the two

digital terrain models particularly highlighted in the

pinnacles of the terrain and tops of the rocks in that

inclined area. When highlighting the

+ 3

and

–

areas we see that areas contoured in blue show an

elevation decrease whereas areas with a red contour

signify an elevation increase. The areas that remained

in teal blue and showed no contour experienced

minimal change. Rockfall can be tied to the high

difference between 2024 and 2017. Based on figure,

the areas that lost elevation ended up depositing the

elevation lower in the basin which is typical of

rockfall situations. There were some areas shown in

yellow that were trees not cleaned properly with the

DJI Terra software and was considered as ground

points while clearly, they were not.

4.3 Land Cover Classification Using a

Deep Learning Model

Upon completion of the model training, the model

began to recognize patterns and identify the classes

found in the control images based on their classified

masks. After this step, the Rio Cucco orthophotos

were introduced. Both datasets which refer to the Rio

Cucco 2024 orthophoto and the Rio Cucco 2017

orthophoto were tested to check the effectiveness of

the model at identifying the classes in an

unintroduced image. The semantic segmentation with

this deep learning model in its current form produced

mixed results in terms of its classification as it detects

vegetation, and bare land well but struggles with

accurately detecting roads and buildings in the new

image. Based on Figure 3, it is evident that the model

works very well at predicting the classes based on the

images that it has gotten used to (its original dataset).

In Figure 4 however, it can be observed that the model

recognizes the shapes and is able to detect well the

land (represented in light purple) while mistaking

vegetation as water (orange instead of purple). The

model produced mixed results with buildings as it

was able to recognize their location but did not have

proper edge detection for them. The testing of this

model yielded mixed results in terms of its

segmentation abilities.

Figure 3: Deep Learning Model Training Results.

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

108

Figure 4: Deep Learning Model Results with Rio Cucco

Orthophoto.

5 DISCUSSIONS

With respect to the Rio Cucco basin, comparing past

data with present data is quintessential to understand

the differences both spatially and temporally. Based

on the findings of the data when comparing the data

of the survey in 2024 and the FVG 2017 data taken

from EAGLE FVG, the differences that existed were

evident especially when conducting the DoD. The

DoD pointed to an increase of elevation in certain

areas and a decrease in elevation in others. The

increase of elevation was found further down the

slope and the decrease of elevation was found higher

in the slopes. Several explanations could be tied to

why this variation of elevation occurred. The

variation could have been tied to possible rockfall

events that happened and lead to sediment deposition

by removing debris from higher up the slope to be

deposited further down the slope. However, the

variations could have also been due to errors in the

model itself. To clarify which case most likely caused

this variation between the two DTM, an examination

of the drone images for the June 2024 survey was

done. After careful examination, rock debris was

noticed downstream in the areas consistent with the

blue and red colour contours. Fig.5 shows an image

of rocks deposited near a slopped area in the basin.

Based on the location of the rock deposition, it can be

inferred that the elevation differences in the DTM

were most likely due to a rockfall event.

Based on the history of the area and the nature of

its geology, it is expected that rockfall events will

continue to occur in the area especially because of the

calcareous and dolomitic set of rocks that are subject

to thermoclastism and cryoclastism. With respect to

the deep learning model, the testing showed mixed

results in terms of its power at classifying

unintroduced images. Other steps to increase the

effectiveness of the model would include data

augmentation techniques and introducing the digital

Figure 5: Rockfall in Sloped Areas in Rio Cucco (2024).

terrain models or the digital surface models in the

training phase so that the model can segment the area

with respect to certain elevation patterns

differentiating between water, road, buildings,

vegetation and other classes.

6 CONCLUSIONS

In summary, the Rio Cucco basin is subject to several

changes over the years and needs to be monitored in

terms of natural hazard studies. The geological

parameters as well as the morphology makes the Rio

Cucco basin particularly liable to the effects of

climate change. The nature of its rock formations and

slopes which are prone to weathering due to abrupt

changes in temperature highlights the importance of

consistent monitoring. With the advancement of

geomatic techniques , artificial intelligence (AI)

models, and internet of things (IoT) sensors, perhaps

a predictive approach could be pursued to model

where and when natural hazards in the area would

most likely occur. Consequently, preventative

measures could be taken to reduce the impact of these

events that are likely to become more common with

the emergence of climate change.

ACKNOWLEDGEMENTS

This study was carried out within the

«GeoMORPHomEtry throUgh Scales for a resilient

landscape» project – funded by European Union –

Next Generation EU within the PRIN 2022 program

(D.D. 104 - 02/02/2022 Ministero dell’Università e

della Ricerca). This manuscript reflects only the

authors’ views and opinions and the Ministry cannot

be considered responsible for them.

Analyzing Multitemporal Datasets to Monitor Topographic Changes in Rio Cucco Italy

109

A special thanks is to be extended to our team

members at CNR-IRPI Padova who aided us with the

information regarding the Rio Cucco basin as well as

helping us with the June 2024 survey.

REFERENCES

Agisoft LLC. (n.d.). Agisoft Metashape.

https://www.agisoft.com

Belcore, E., Piras, M., & Pezzoli, A. (2022). Land Cover

Classification from Very High-Resolution UAS Data

for Flood Risk Mapping. Sensors, 22(15), 5622.

https://doi.org/10.3390/s22155622

Belcore, E., Piras, M., & Wozniak, E. (2020). Specific

Alpine Environment Land Cover Classification

Methodology: Google Earth Engine Processing for

Sentinel-2 Data. The International Archives of the

Photogrammetry, Remote Sensing and Spatial

Information Sciences, XLIII-B3-2020, 663–670.

https://doi.org/10.5194/isprs-archives-XLIII-B3-2020-

663-2020

Cavalli, M., Trevisani, S., Comiti, F., & Marchi, L. (2013).

Geomorphometric assessment of spatial sediment

connectivity in small Alpine catchments.

Geomorphology, 188, 31–41. https://doi.org/10.1016/

j.geomorph.2012.05.007

DJI. (n.d.). DJI Terra - Make the World Your Digital Asset

- DJI. https://enterprise.dji.com/it/dji-terra

Girardeau-Montaut, D. (n.d.). CloudCompare - Open

Source project. https://www.danielgm.net/cc/

Huggel, C., Clague, J. J., & Korup, O. (2012). Is climate

change responsible for changing landslide activity in

high mountains? Earth Surface Processes and

Landforms, 37(1), 77–91. https://doi.org/10.1002/

esp.2223

Indian Space Research Organization. (2023, May 31). Land

cover Classification : Bhuvan Satellite Data. Kaggle.

https://www.kaggle.com/datasets/khushiipatni/satellite

-image-and-mask

International Society for Photogrammetry and Remote

Sensing. (2024, March 20). Urban Segmentation -

ISPRS. Kaggle. https://www.kaggle.com/datasets/

aletbm/urban-segmentation-isprs

Kaggle. (n.d.). Kaggle: your machine learning and data

science community. https://www.kaggle.com/

Roia Foundation. (2020, May 29). Semantic segmentation

of aerial imagery. Kaggle. https://www.kaggle.com/

datasets/humansintheloop/semantic-segmentation-of-

aerial-imagery

Sangelantoni, L., Gioia, E., & Marincioni, F. (2018). Impact

of climate change on landslides frequency: The Esino

river basin case study (Central Italy). Natural Hazards,

93(2), 849–884. https://doi.org/10.1007/s11069-018-

3328-6

Schneiderbauer, S., Fontanella Pisa, P., Delves, J. L.,

Pedoth, L., Rufat, S., Erschbamer, M., Thaler, T.,

Carnelli, F., & Granados-Chahin, S. (2021). Risk

perception of climate change and natural hazards in

global mountain regions: A critical review. Science of

The Total Environment, 784, 146957.

https://doi.org/10.1016/j.scitotenv.2021.146957

Semnani, S. J., Han, Y., Bonfils, C. J., & White, J. A. (2025).

Assessing the impact of climate change on rainfall-

triggered landslides: A case study in California.

Landslides. https://doi.org/10.1007/s10346-024-

02428-0

Zhang, C., Sargent, I., Pan, X., Li, H., Gardiner, A., Hare,

J., & Atkinson, P. M. (2019). Joint Deep Learning for

land cover and land use classification. Remote Sensing

of Environment, 221, 173–187. https://doi.org/10.1016/

j.rse.2018.11.014

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

110