Remote Sensing as a Sentinel for Safeguarding European Critical Infrastructure in the Face of Natural Disasters

Abstract

Critical infrastructure, such as transport networks, energy facilities, and urban installations, is increasingly vulnerable to natural hazards and climate change. Remote sensing technologies, namely satellite imagery, offer solutions for monitoring, evaluating, and enhancing the resilience of these vital assets. This paper explores how applications based on synthetic aperture radar (SAR) and optical satellite imagery contribute to the protection of critical infrastructure by enabling near real-time monitoring and early detection of natural hazards for actionable insights across various European critical infrastructure sectors. Case studies demonstrate the integration of remote sensing data into geographic information systems (GISs) for promoting situational awareness, risk assessment, and predictive modeling of natural disasters. These include floods, landslides, wildfires, and earthquakes. Accordingly, this study underlines the role of remote sensing in supporting long-term infrastructure planning and climate adaptation strategies. The presented work supports the goals of the European Union (EU-HORIZON)-sponsored ATLANTIS project, which focuses on strengthening the resilience of critical EU infrastructures by providing authorities and civil protection services with effective tools for managing natural hazards.

1. Introduction

Critical infrastructure resilience (CIR) is an essential concept in contemporary risk management and urban planning, particularly as societies face increasing threats from natural disasters [1]. Critical infrastructures include transportation networks, energy grids, water supply systems, and communication networks, and they are foundational to the functioning of modern societies. Their resilience is defined as the capacity to prepare for, respond to, recover from, and adapt to adverse events while maintaining essential functions [2,3].

The historical evolution of critical infrastructure resilience can be traced back to the early 2000s when the concept began to gain traction in response to significant events that exposed vulnerabilities in infrastructure systems [4]. Major events such as 2005 Hurricane Katrina in the USA up until the 2010 earthquake and tsunami leading to the Fukushima nuclear accident in Japan were pivotal moments that highlighted the fragility of critical infrastructures and the cascading effects of their failures [5]. These events prompted governments and organizations to reevaluate their approaches to infrastructure resilience, leading to the establishment of various frameworks and guidelines aimed at enhancing resilience. The U.S. Department of Homeland Security (DHS) established the National Infrastructure Protection Plan (NIPP) in 2006, which emphasized the importance of resilience in protecting critical infrastructure [6]. Similarly, in 2006, the European Union introduced the European Programme for Critical Infrastructure Protection (EPCIP) to enhance the resilience of critical infrastructures across member states [7]. These initiatives marked a shift towards a more integrated and systematic approach to resilience, recognizing the interdependencies among various infrastructure systems.

The theoretical foundations of CIR draw from various disciplines, including engineering, sociology, and environmental science. Resilience theory, originally rooted in ecological studies, states that systems can absorb disturbances while maintaining their core functions [8]. In the context of critical infrastructure, resilience encompasses several dimensions [1], including the following:

• Technical Resilience: The ability of physical systems to withstand disruptions through robust design and engineering practices. It refers to the capacity of systems to maintain and quickly restore functionality in the face of disruptions, including natural hazards, cyberattacks, or technical failures. It encompasses the robustness, redundancy, adaptability, and recovery capabilities of technical components and architectures, enabling systems to absorb shocks, respond effectively, and evolve based on emerging threats and operational insights.
• Organizational Resilience: The capacity of organizations to adapt and respond effectively to crises, including the development of contingency plans and training programs.
• Economic Resilience: The ability of infrastructure systems to recover economically from disruptions, including the assessment of costs associated with failures and recovery efforts.
• Social Resilience: The role of communities in enhancing resilience through social networks, public awareness, and engagement in resilience planning [9].

Recent advances in CIR research have led to the development of innovative methodologies and frameworks aimed at enhancing resilience [10]. These advancements emphasize the need for a holistic understanding of resilience that encompasses both physical and digital dimensions [11]. The emergence of geospatial frameworks has allowed for a more nuanced analysis of infrastructure resilience. These frameworks enable the identification of vulnerabilities specific to geographic contexts, facilitating targeted resilience strategies [12]. According to Der Sarkissian et al. [1], the systemic and spatial interdependencies of CIR require the use of spatial data and geographic information systems (GISs). GIS offers a robust framework to address the complexity of CIR and their interdependence by enabling the spatial modeling of infrastructure as an interconnected system of systems. Additionally, GIS establishes a shared semantic platform for visualization, simulation, analysis, modeling, data integration, and communication of infrastructure-related data, all of which are essential for enhancing CI resilience strategies. By supporting simulation and modeling, GIS enables the identification of spatial interrelationships and risk-prone interdependencies across infrastructure systems. This capability is particularly valuable for GIS-based risk assessment models, which support scenario planning, “what-if” simulations, and enhanced situational awareness in the face of natural hazards.

The integration of climate change considerations into resilience planning has gained prominence [1]. Research has highlighted the importance of designing climate-resilient infrastructures that can withstand extreme weather events, thereby reducing the risk of service disruptions [13]. For instance, studies have shown that retrofitting existing infrastructures to enhance their resilience to flooding and heatwaves can significantly mitigate risks [14]. Recent studies have emphasized the interdependence of critical infrastructures, where disruptions in one sector can have cascading effects on others. This understanding necessitates a holistic approach to resilience that considers the interactions between various infrastructure systems [15].

For European CIR, the expected annual damage to critical infrastructure in Europe is currently EUR 3.4 billion per year and is expected to reach EUR 19.6 billion by the 2050s and EUR 37.0 billion by the 2080s, solely due to climate change effects. According to Lomba-Fernàndez [15], the climate-related hazards with the greatest potential to disrupt critical infrastructure in Europe include the following: (i) heatwaves, (ii) cold waves, (iii) river floods, (iv) coastal floods, (v) droughts, (vi) wildfires, and (vii) windstorms.

Despite significant progress in the field, several gaps remain in the literature on critical infrastructure resilience. There is a lack of comprehensive frameworks that integrate resilience strategies across different sectors. While many studies focus on individual infrastructures, there is a need for research that examines the interdependencies and collective resilience of multiple systems [1]. The governance of critical infrastructures presents challenges, particularly for coordinating efforts between the public and private sectors. The availability and quality of data on critical infrastructures also remain a challenge. Often, many studies rely on existing or incomplete data, which can hinder the development of effective resilience strategies. Improved data collection and sharing practices are therefore necessary to inform resilience planning.

To bridge these gaps, particularly those regarding the integration of resilience strategies across multiple interconnected sectors, this study proposes a holistic framework that combines synthetic aperture radar (SAR), optical satellite imagery, and GIS-based risk assessment to monitor and enhance resilience across diverse infrastructure systems, such as transport, energy, and communication networks. This includes an analysis of climate-related hazards, such as floods and wildfires, and their impacts on critical infrastructures, using climate projections and case studies. By addressing these key challenges, this study contributes to more robust, integrated, and sustainable resilience frameworks for informing policy and practice in the face of evolving risks. The technologies developed in this paper correspond to the objectives of the European Horizon ATLANTIS project: https://www.atlantis-horizon.eu, accessed on 29 May 2025. ATLANTIS focuses on enhancing the resilience of key EU critical infrastructures, providing authorities and civil protection with accurate and ready-to-use tools for preventing, responding to, and recovering from natural hazards.

Accordingly, this paper is organized as follows: Section 2 details the materials and methods used in this study, namely, the study area and the four case studies, the utilized datasets, and the methodologies to assess landslides, wildfires, and flood risks, as well as earthquake damage assessment. Section 3 presents the results of the various approaches described in Section 2. Section 4 discusses the limitations of this study and its method, as well as their replicability. Finally, Section 5 concludes this article and states the ways forward.

2. Materials and Methods

2.1. Study Areas and Case Studies

The methodologies presented in this paper were applied and validated in several real-world case studies. A total of four cases were identified (

Table 1. Retained case studies and context.

Event and Date	Context
Fréjus (French Italian borders): Tunnel landslide on 27 August 2023	On 27 August 2023, a massive landslide occurred between Saint-Michel-de-Maurienne and Freney (France), 6 km near the entrance of the Frejus Road Tunnel. An estimated total of 700 m3 of rock spilled onto the A43 motorway and RD1006 road, causing the shutdown of car traffic along both routes for several days. Moreover, the landslide led to the shutdown of the Frejus Rail Tunnel, which is not expected to be re-opened until spring of 2025. Due to its large volume and extension, the landslide can be detected and analyzed through remote sensing techniques. Therefore, it is well suited as a validation case study for the developed landslide delineation models.
Piedmont (Italy): Wildfires on October 2017	The Piedmont region (Italy) experienced several wildfires in the second half of October 2017, just days after the alert of maximum danger for forest fires was declared. From October 18th to the 26th, both regional and national resources were involved in managing the emergency, including state aviation fleet crews, field teams of volunteers and firefighters, and air carriers. Italy’s Department of Civil Protection activated Copernicus Emergency Management Service–Rapid Mapping (EMSR)253 (https://mapping.emergency.copernicus.eu/activations/EMSR253/, accessed on 29 May 2025) event on October 28th to manually produce delineation maps from satellite images, showing the extent of the fire over six areas of interest. These maps were used to validate the developed deep learning model for wildfire delineation and severity estimation.
Slovenia: Floods of August 2023	The Slovenian Environment Agency (ARSO) issued a red weather alert on 4 August 2023 at 02:00 due to severe weather conditions marked by heavy rainfall affecting western, northern, eastern, and central regions of the country. This resulted in significant flooding and landslides, rendering some areas inaccessible and leading to the evacuation of residents along the Drava, Sava, and Sora rivers. The heavy rainfall resulted in damage to critical infrastructure such as roads, bridges, culverts, and the energy network. This event was used as a validation case study for the flood delineation model and flood risk map algorithm to assess their effectiveness and reliability in a real operational context. In particular, Copernicus EMSR680 (https://mapping.emergency.copernicus.eu/activations/EMSR680/, accessed on 29 May 2025) activation was used for testing, as it contains manually produced delineation maps of flooded areas.
Amatrice (Italy): Earthquake on 24 August 2016	On 24 August 2016, a 6.2 magnitude earthquake occurred with an epicenter in Amatrice in the Umbria region (Italy). The earthquake resulted in significant loss of life and injury, with 299 fatalities and approximately 400 injuries. Over half of the buildings in Amatrice were damaged or destroyed, leaving an estimated 4454 individuals displaced. This event was selected due to the significant ground deformation it caused and the availability of Sentinel-1 data for the area and time of the event.

), each faced with one of the following hazards: landslides, wildfires, floods, and earthquakes. These case studies were chosen given the impact on CIR, in addition to the magnitude, extent, and impacts of the disasters. Figure 1 presents the retained case studies.

2.2. Datasets

The utilized dataset in this study consists of satellite-derived data, hydrological repositories, and climatic models (

Table 2. Study utilized dataset.

Data Source	Dataset	Type	Source/Provider	Spatial Resolution	Purpose
Earth observation data	Sentinel-2	Optical Satellite Imagery	European Space Agency (Copernicus)	10 m/px	Hazard delineation, land cover classification
	Sentinel-1	SAR Satellite Imagery	European Space Agency (Copernicus)	10 m/px	Flood and land subsidence monitoring
	Advanced Land Observing Satellite (ALOS) PALSAR (Phased Array type L-band Synthetic Aperture Radar) DEM	Digital Elevation Model (DEM)	Japan Aerospace Exploration Agency	30 m/px	Terrain analysis, landslide and flood modeling
	NASA SRTM DEM	Digital Elevation Model (DEM)	NASA	30 m/px	Topographic analysis, hydrological modeling
	ESA World Cover	Land Cover Classification	European Space Agency	10 m/px	Land cover assessment, wildfire susceptibility
Hydro-climatic data	World Wide Fund for Nature (WWF) HydroSHEDS Flow Accumulation	Hydrological Data	WWF Conservation Science Program the U.S. Geological Survey, the International Centre for Tropical Agriculture, The Nature Conservancy, and the Center for Environmental Systems Research of the University of Kassel, Germany.	15 arc-seconds	River flow analysis, flood hazard assessment
	WWF Free Flowing Rivers	Hydrological Data	WWF Conservation Science Program in partnership with the U.S. Geological Survey, the International Centre for Tropical Agriculture, The Nature Conservancy, and the Center for Environmental Systems Research of the University of Kassel, Germany.	15 arc-seconds	River network analysis, environmental studies
	JRC Historical Water Data	Hydrological Data	Joint Research Centre (JRC)	Variable	Flood history analysis, hazard risk assessment
	Coupled Model Intercomparison Project Phase 6 (CMIP6) Climate Projections	Climate Model Data	Intergovernmental Panel on Climate Change	Downscaled to 1 km/px	Future climate impact analysis on hazards
Other data	Historical Hazard Events Database	Multi-hazard Data Collection	Open-access databases	Variable (vectorial format)	Training/validation of hazard models

).

2.3. Methodology

In this section, the methodologies developed for studying the four hazards (i.e., landslides, floods, wildfires, and land subsidence/earthquakes) are detailed.

2.3.1. Landslides

Two technologies for the delineation of landslides have been developed.

Landslide Delineation Through AI-Based Change Detection from Sentinel-2

The first approach focuses on the delineation of landslides through Artificial Intelligence (AI)-based change detection from Sentinel-2 optical satellite imagery. Given a pair of two Sentinel-2 images, taken before and after a landslide-triggering event, a machine learning model outputs a binary segmentation mask, delineating landslides which are visible in the “post” image but not in the “pre” image. The workflow is designed to overcome the limitations of existing works:

Most previous studies focus on a single ecoregion, i.e., a geographical area exhibiting homogeneous properties in terms of lithology, morphology, landforms, soil composition, vegetation type, etc. When considering a supervised machine learning approach, the limited geographical scale of the training data inherently hampers the model’s ability to generalize to new visually different areas [16]. In this study, several open-access landslide inventories from the literature, with a focus on heterogeneity of ecoregions, landslide sizes, and landslide-triggering causes, are collected.

Most works leverage commercial (very) high-resolution data (e.g., [17]). While high-resolution imagery can capture fine-grained details that might be useful for delineating even small landslides and improving segmentation performances, acquiring this data is relatively expensive. Moreover, high-resolution satellites typically cover less ground area with each flyover (due to their narrow swath width) [18]. Both factors make the use of high-resolution satellite imagery for emergency response use cases limited, especially when the region of interest is large. Using medium-resolution non-commercial satellite data, such as Sentinel-2 optical imagery, can be a solution, but it is still not as widely explored in the literature [19], hence its inclusion in this study.

Existing landslide segmentation models typically do not work with bi-temporal pairs of images (before and after landslides were triggered) but rather with a single image acquired after the landslide event. This leads to the detection of both newly activated landslides associated with the event of interest and old landslides which may have happened in the past. This ambiguity is not desirable in emergency response use cases, in which it is critical to rapidly map the impact area of newly activated events. In response to this gap, this study proposes addressing landslide delineation as a change detection task. Moreover, leveraging both pre- and post-event images allows the better delineation of landslides due to the additional contextual information, avoiding false detection (e.g., sediment deposits in the river channels, objects with high reflectance).

To overcome visual homogeneity limitations, several open-access landslide inventories from several ecoregions of the world were harmonized and combined. Through this step, a global and diverse landslide database was constructed. This database includes manually validated landslides of different sizes, ranging from very small (in the range of 100 m²) to very large (in the range of 100,000 m²), and landslides triggered by different catastrophic events (earthquakes and heavy rainfall). Data is stored in a georeferenced vectorial format (i.e., polygons).

For each region of interest in the database, pre- and post-event Sentinel-2 L2A images are acquired (ranging between 3 months before and 1 month after the occurrence of the landslide-triggering event), hence constructing a dataset of multiple bitemporal pre–post-image pairs. Sentinel-2 L2A images are surface reflectance products that are corrected for atmospheric effects and terrain distortion and georeferenced to standard coordinate systems. As Sentinel-2 Level 2 images are atmospherically corrected, no georeferencing error is introduced. Accordingly, there are no risks of false-positive landslide artifacts.

By rasterizing polygons, a ground truth segmentation mask is obtained for each pre-post image pair. Due to the extended presence of clouds in the images, a cloud detector trained on the CloudSen12 dataset [20] is used to generate a cloud mask. Specifically, the detector segments each image into four categories: clear sky, thick clouds, thin clouds, and cloud shadow.

The chosen model architecture is a UNet segmentation network [21] with a ResNet-18 backbone [21]. To train the model, Sentinel-2 image pairs are tiled; then, tiles with a low number of landslide pixels (<200) are discarded to keep only informative patches and decrease the training time. Finally, pixels are normalized to [0,1], as this range is more easily managed by a deep learning model. The model was trained for 100 epochs on the Sentinel-2 tile pairs. The loss function is set as the sum of a cross-entropy loss and a dice loss. To overcome class imbalance between background/”no landslide” and “landslide” classes, the weight for the “landslide” class in the cross-entropy loss is set to 5. Pixels classified as thick clouds in any of the two generated cloud masks are always treated as “no landslide”, whereas thin clouds are to be ignored in (i.e., excluded from) the training process. To obtain the final segmentation mask, the output logits are rectified through the sigmoid function and set at a threshold of 0.5. To enhance the visual heterogeneity of the dataset and increase the generalization capability of the model, several data augmentation techniques were applied to the image tiles, such as random horizontal/vertical flips, random rotations, histogram matching, brightness, and contrast jittering.

A key limitation of this approach is the 10 m/px resolution of Sentinel-2 imagery, which restricts the detection and precise delineation of small landslides (e.g., <100 m²), as these events may span only a few pixels.

We evaluated the model on the pre–post-pairs of one of the landslide inventories we have processed, which was left out of the training. The model achieved an F1 score of 34.8%, a precision of 58.2% and a recall of 24.9%. The achieved metrics are relatively low. In spite of the efforts put into dataset creation and data augmentation, we argue that such underwhelming performances are due to the sub-optimal quality of the landslide inventories available in the literature, as well as the low resolution of Sentinel-2 imagery compared to the size of a typical landslide. Nonetheless, precision is sufficiently high, indicating that the model can effectively segment landslides with a reasonably low amount of false positives.

Landslide Delineation Through Sentinel-1 SAR Backscatter Differences

The second method enables the mapping of landslides and associated damage using Sentinel-1 SAR images by comparing pre- and post-event images. This technique utilizes Sentinel-1 Synthetic Aperture Radar (SAR) data to detect changes in ground deformation and structural integrity caused by landslides. By analyzing radar backscatter, significant changes before and after the event can be identified, offering crucial insights for disaster management and response. Monitoring ground deformation and structural changes is vital for effective disaster management, and Sentinel-1 SAR data provides a robust solution due to its ability to penetrate cloud cover and deliver high-resolution imagery.

After an event occurs and the affected area is identified, Sentinel-1 images capturing the area of interest both before and after the disaster are selected. The chosen dates should provide a clear temporal window to observe significant changes. The data is filtered to include only images captured in the Interferometric Wide (IW) swath mode, which is optimized for land observations and minimizes noise. The data is further categorized based on the satellite’s orbit direction, separating them into ascending (south-to-north) and descending (north-to-south) passes. Additional precision is achieved by sorting the data into different collections based on polarization (Vertical Transmit/Vertical Receive Polarisation VV or Vertical Transmit/Horizontal Receive Polarisation VH) and orbit pass, ensuring that the analysis considers the specific characteristics of the radar signal.

To identify changes in backscatter, the significance of changes in radar backscatter is calculated between the pre-event and post-event images. Radar backscatter measures the amount of radar signal reflected back to the satellite, with changes indicating ground deformation or structural changes. A smoothing function is applied to reduce radar speckles and enhance the clarity of the detected changes, improving the visual quality of the results. A threshold is then applied to the smoothed significance image to classify areas into significant change, low change, and no change, helping to identify the most affected areas and clearly delineating impacted regions.

Landslide detection effectively identifies the extent of the landslide, highlighting areas of significant rock movement and debris flow and indicating high levels of displacement. This analysis reveals two main areas of movement: dislodged objects and the area where the displaced mass settled.

2.3.2. Wildfires

Real-time monitoring of wildfires is a complex task due to their rapid spread and potentially very large extent, especially in densely vegetated areas. An effective tool for wildfire monitoring should therefore provide near-real-time information not only on extent and ignition points but also on severities in terms of disruption of natural ecosystems, human settlements, etc. The timely knowledge of this information can help in concentrating and planning the correct strategies to minimize as many negative impacts as possible. Such information can be derived from satellite images, which are generally available in a short time after the start of the wildfire event. A deep learning model for the monitoring of wildfires was developed. This model is designed to perform both the delineation of wildfires from Sentinel-2 images and the estimation of their severity.

The proposed approach tries to derive such information from satellite images, which are available a short time after the start of the fire event. Specifically, an AI model has been developed to identify, delineate, and estimate the severity of a wildfire from Sentinel-2 images given that a burned area is present in the image.

Given a Sentinel-2 image, the model provides two outputs: first, a binary map containing the delineation of the burned area; second, a map containing the estimation of the severity of the fire based on an index of 5 levels (

Table 3. Taxonomy of wildfire severity according to Copernicus EMS.

Index	Severity
0	No fire
1	No visible damage
2	Possibly damaged
3	Damaged
4	Destroyed

) with respect to the taxonomy used in Copernicus EMS [22].

The model’s architecture is divided into two steps: the first one produces the delineation mask, while the second one produces the severity mask, as shown in Figure 2. The two models have been trained one after the other.

To train the delineation model, which is based on a UPerNet-50 architecture [23], a comprehensive dataset covering several fire events was first created. For this purpose, a list of 433 Copernicus EMS wildfire events was collected. The geographical distribution is shown in Figure 3. For each event, pre- and post-event Sentinel-2 images were acquired, in addition to land cover data derived from the ESA World Cover dataset [24]. A cloud detector for each Sentinel-2 image was applied to exclude cloud pixels from the training process.

A multitask learning technique was used to obtain more robust delineation results. The delineation model consists of a single encoder and a single decoder with two classification heads ($h_{D}$ and $h_{LC}$), each one responsible for a specific task: delineation of burned area and prediction of land cover, respectively. The input of the model is a pair of S2 images (pre- and post-event). The ground truth masks are obtained from the manually validated delineation map from the Copernicus EMS database for the delineation head, and the ESA World Cover map is used for the land cover prediction head (Figure 4).

Using a multitask learning framework and a single image encoder, the delineation model learns robust shared features, capturing common patterns between the two tasks and ultimately obtaining better delineation results with respect to training only on wildfire delineation maps. At inference time, the land cover classification head is dropped. To train the severity estimation model (based on a UNet architecture), the satellite imagery and the produced delineation map are both passed as inputs. The model classifies each input pixel into a severity index. The severity grading map from the Copernicus EMS wildfire database is used as ground truth. We evaluated the model on the test split which was left out of training. The model obtained an F1 score of 91.86% in the delineation task and an average root mean squared error (RMSE) of 0.876 in the severity grading task.

2.3.3. Floods

Two technologies have been developed for floods: one for the delineation of floods and one to produce future risk maps. For the latter, i.e., production of future risk maps, climate data was included. Further details are given in the corresponding section (Flood Risk Map Through a Multi-Sourced Analysis and Flood Risk Projection sub-sections).

Flood Delineation Through Deep Learning Model

The first approach is an AI-based delineation model which leverages SAR images from Sentinel-1. SAR imagery is particularly suited for the task of flood delineation, as it is not affected by the presence of clouds, which would obstruct optical images instead. Moreover, the relatively high spatial resolution (10 m/px) and high revisit time (three days at the Equator) make Sentinel-1 ideal for flood delineation.

The Sen1Floods11 dataset [25] is used as a ground truth delineation data source. Sen1Floods11 is a dataset created specifically for the training and validation of deep learning flood detection algorithms from Sentinel-1 and Sentinel-2 data. It is composed of 11 distinct flood events selected from the news, covering frequently flooded regions around the world. The dataset features 4831 Sentinel-1 and Sentinel-2 tiles of size 512 × 512 pixels and a resolution of 10 m/px. The largest part of the dataset is composed of 4370 tiles, automatically labeled by simple flood delineation algorithms, which are meant to be used as weakly supervised training data. Due to the excessive roughness of the automatic delineation masks, 446 high-quality hand-labeled delineation maps were used as ground truth for training the model.

The chosen model architecture is ResUNet for binary segmentation [26]. ResUNet is a hybrid model that combines the strengths of residual networks (ResNet) and the UNet segmentation model. The ResNet18 backbone provides the model with the ability to learn hierarchical features, while the UNet structure enables precise spatial localization, which is crucial for delineating flood extents. The model was trained on the Sentinel-1 tiles of the Sen1Floods11 dataset (VV and VH polarization channels) using the hand-labeled flood delineation maps as ground truth. Training is performed by minimizing a weighted cross-entropy loss to make up for the relative scarcity of pixels for the “flood” class. To avoid overfitting the small volume of training images and to improve the generation capability of the delineation model, several data augmentation operations were applied, specifically the following: random rotations and horizontal and vertical flips, each with a probability of 0.5.

As a post-processing step, a set of pixels with known permanent water bodies (such as lakes and rivers), which was set to “no flood”, was defined using the hydrography layer of OpenStreetMap. We performed a validation of our model on a small selection of Copernicus EMS activations of European flood events (namely, 5 AOIs taken from EMS257 and EMS358 activations). The model achieved an average F1 score of 44.28%.

Flood Risk Map Through a Multi-Sourced Analysis

The second approach identifies areas at risk of flooding by first producing a historical analysis of the area for flood water occurrence at 30 m spatial resolution. A map of the location and temporal distribution of surface water from 1984 to 2025 is produced using LANDSAT imagery and provides statistics on the extent of and change in those water surfaces through time. This data is generated using a multitude of scenes from Landsat 4, 5, 7, 8, and 9. Each pixel is individually classified into water/non-water using the Landsat quality band bit mask.

Four Distinct Images Are Produced from This Analysis:

• Water occurrence: The frequency with which water was present. Range between 0 and 100%.
• Water max extent: Binary image containing 1, indicating that water is detected: 0 for no water and 1 for water.
• Water seasonality: Number of months water is present. Range between 0 and 12.
• Water recurrence: The frequency with which water returns from year to year. Range between 0 and 100%.

Along with the historical analysis already established, an algorithm to derive the pixel-by-pixel flood risk severity of flooding in the study area is developed. This algorithm produces an image at 50 m resolution with 5 levels of severity: 1 being the areas that will be the least affected and 5 being the areas that will be the most affected. This analysis is only useful for determining areas that will be affected first by flooding based on return period scenarios of 10, 20, 50, 75, 100, 200, and 500 years.

This algorithm uses elevation, slope, precipitation, and runoff projections; distance from river; flow accumulation points; soil characteristics; and land cover classification. Elevation and slope were derived from NASA SRTM Digital Elevation 30 m; precipitation and runoff projections for the 2030, 2050, and 2070 years at Shared Socioeconomic Pathways SSP 4.5 and 8.5 were derived from CMIP6 and downscaled to the required resolution; distance from rivers was derived from WWF HydroSHEDS Free Flowing Rivers Network v1; flow accumulation was derived from WWF HydroSHEDS Flow Accumulation in 15 arc-seconds; soil characteristics were derived from OpenLandMap.

These inputs were compounded, and they were assigned severity indicators ranging between 1 and 5, where 1 is considered to be a very low chance of flooding and 5 is considered to be a very high chance of flooding. These results were used in an equation where each indicator was given a weight out of 100 (using historical flooding, we determined these weights y using a multiple regression method), and a final severity number was calculated. This severity number indicates which areas will be flooded first in the case of certain flooding scenarios:

Flood Risk formula: (Runoff × 7.5) + (Precipitation to Runoff × 7.5) + (Precipitation > 20 mm × 5) + (Precipitation > 50 mm × 5) + (Land Cover × 5) + (Flow Direction & Accumulation × 15) + (Soil Characteristics × 10) + (Max Extent × 17.5) + (Distance × 17.5)

(1)

Weights were determined based on trial-and-error testing until the best fit with historical floods in the area was obtained. Finally, the following formula is obtained:

$$\mathrm{F}\mathrm{l}\mathrm{o}\mathrm{o}\mathrm{d} \mathrm{R}\mathrm{i}\mathrm{s}\mathrm{k}=\mathrm{R}\times \mathrm{a}+{\displaystyle \frac{\mathrm{P}}{\mathrm{R}}}\times \mathrm{b}+\mathrm{P}20\times \mathrm{c}+\mathrm{P}50\times \mathrm{d}+\mathrm{L}\mathrm{U}\mathrm{L}\mathrm{C}\times \mathrm{e}+\mathrm{F}\&\mathrm{A}\times \mathrm{f}+\mathrm{S}\times \mathrm{g}+\mathrm{W}\times \mathrm{h}+\mathrm{D}\times \mathrm{i}$$

(2)

where

• R: Maximum runoff per pixel in mm/year (CMIP6 indicator);
• P/R: Ratio of precipitation to runoff (CMIP6 indicator);
• P20: Maximum number of days where precipitation per pixel is higher than 20 mm/day;
• P50: Maximum number of days where precipitation per pixel is higher than 50 mm/day;
• LULC: Level4 landcover map;
• F&A: Water accumulation points for each pixel (this is the water collection area and the degree to which the area outflows into each pixel);
• S: Soil physical characteristics;
• W: Historical water presence;
• D: Distance from rivers in Km.

The slope of each pixel in degrees is calculated and is used as a final restrictor of flooding (where, above a certain slope, no flooding can occur).

2.3.4. Land Subsidence

Land subsidence is a geological hazard occurring globally, and it is caused by natural processes (e.g., soil consolidation) or anthropogenic activities (e.g., underground construction, over-extraction of groundwater). It can lead to surface cracking, differential settlement, structural damage, and other detrimental effects on infrastructure. Monitoring land subsidence is therefore critical for infrastructure stakeholders to prevent damage, mitigate associated risks, and detect gradual, long-term deformations.

Multi-interferogram techniques, which leverage long time series of interferometric synthetic aperture radar (InSAR) data, offer a robust method for monitoring land subsidence [27]. These techniques effectively filter noise and atmospheric effects, providing accurate measurements of ground displacement. Compared to traditional geodetic monitoring methods, they have demonstrated significant cost-effectiveness and high precision in numerous studies.

SAR is an active remote sensing technology that transmits pulses of energy and records their echoes reflected from the Earth’s surface. The Differential Interferometric SAR (DInSAR) technique utilizes two SAR images acquired at different times with consistent geometry. The phase difference between the two images is used to generate an interferogram, enabling precise measurement of surface displacement.

Atmospheric variations between image acquisitions can introduce errors (decorrelation) in deformation estimates. To mitigate these effects, multi-temporal InSAR techniques analyse extended time series of SAR imagery, which allows for the detection of even slow continuous deformations such as land subsidence.

The Small Baseline Subset (SBAS) technique was used to estimate the vertical movement of the ground. This technique employs distributed targets (e.g., open fields) as reference points. Interferograms are generated using image pairs with small temporal and spatial baselines, reducing the impact of decorrelation. The SBAS processing workflow consisted of the following steps:

• Data Acquisition: Sentinel-1 imagery is freely available and can be downloaded. For optimal results, it is recommended to use a large dataset (at least 20 images) to enable selection of the best image pairs and minimize decorrelation effects.
• Pre-Processing: The metadata of the imagery can be refined by incorporating precise orbit files. A Digital Elevation Model (DEM) of the study area is also required for accurate processing. Specifically, we used NASA SRTM DEM.
• Connection Graph Generation: Spatial and temporal baselines of the images are calculated, and optimal image pairs with minimal baselines are selected for interferogram generation.
• Interferogram Generation: Interferograms are computed for the selected image pairs, forming the basis for deformation analysis.
• First Inversion: Singular Value Decomposition (SVD) is employed to calculate the average surface deformation velocity and residual topography.
• Second Inversion: The temporal evolution of surface deformation is estimated based on the generated interferograms.
• Geometric Correction: The processed data, initially in range-azimuth geometry, are transformed into cartographic coordinates for ease of interpretation by end-users.

2.3.5. Earthquake Damage Detection

Earthquake-induced damage was detected through a change detection methodology that utilized Sentinel-1 Synthetic Aperture Radar (SAR) Ground-Range-Detected (GRD) data, which are processed within the Google Earth Engine (GEE) platform. Google Earth Engine 1.5.12. is developed by Google LLC, 1600 Amphiteatre Parkway, Mountain View, California 94043, United States (Mountain View, CA, USA). This approach relied on backscatter differences to identify areas of significant structural change, with specific measures implemented to mitigate temporal decorrelation and ensure coherence stability. Sentinel-1 GRD data in the Interferometric Wide (IW) mode with a 10 m/px resolution were selected, including both VV (vertical-vertical) and VH (vertical-horizontal) polarizations to capture complementary backscatter information. Data were separated into ascending (south-to-north) and descending (north-to-south) orbit passes to account for viewing geometry variations. Multiple images were collected before and after the earthquake event for each orbit, and polarization was used to create robust temporal composites. For the Amatrice case study (24 August 2016), pre-event images spanned approximately three months prior (e.g., July 2016), while post-event images were acquired immediately after the event (e.g., August–September 2016). This multi-image approach minimized the impact of temporal decorrelation caused by unrelated changes, such as vegetation growth or seasonal variations, by averaging backscatter signals over time.

Mean backscatter images were computed for pre- and post-event periods for each orbit and polarization (VV ascending, VH ascending, VV descending, and VH descending). The variance of each image collection relative to its mean was calculated to enable a statistical assessment of backscatter differences. Significance values from all four datasets (VV/VH, ascending/descending) were averaged to produce a composite significance map, which reduced noise and improved robustness. A focal mean smoothing filter with a 10 m radius was applied to the significance map to mitigate radar speckle. A high significance threshold (e.g., 5) was established to isolate pixels with strong backscatter changes, indicative of severe structural damage, such as collapsed buildings.

Comprehensive ground-based damage assessments were unavailable for the Amatrice earthquake. Consequently, detected damage areas were compared with the seismic intensity shake map provided by the Italian National Institute of Geophysics and Volcanology (INGV). This indirect validation approach ensured alignment with known seismic impacts, as described in Section 3.1.3.

2.3.6. Recap of This Study’s Methods

The flood risk detection method provides a comprehensive assessment by combining multiple influencing factors into a single, integrative equation. A key feature is the use of machine learning and kriging to enhance precipitation data resolution, enabling the identification of localized rainfall variations that significantly impact flood risk. This flexible framework allows users to input higher-resolution or localized data, improving the accuracy of flood projections and supporting more effective risk management.

Similarly, the land subsidence monitoring approach is highly adaptable and can be applied to various regions. By incorporating pre-, co-, and post-event analysis using freely available Sentinel-1 SAR data, it offers a robust framework for tracking subsidence over time. The use of multi-temporal interferometric techniques and standardized processing workflows ensures detailed detection of ground deformation, providing valuable insights for infrastructure planning, urban development, and environmental policy. Further details on the limitations and advantages are given in Section 4.

3. Results

In this section, results obtained from the different methodologies are presented. In the end of this section, a workflow linking the different components is provided.

3.1. Monitoring and Early Hazard Detection

3.1.1. Landslides

The landslide detection methodology was applied to the Fréjus tunnel. The exact location at which the landslide occurred was determined by analyzing news on the Fréjus landslide event and looking at the topographic map on Google Maps. It must be noted that no ground truth delineation of this landslide was found in the literature, nor could it be manually produced by us due to a lack of publicly available post-event aerial surveys or high-resolution satellite images. Therefore, the delineation results produced by our models are discussed qualitatively and visually, as no quantitative metrics could be computed.

Landslide Delineation from Sentinel-2

Firstly, the deep learning model for landslide delineation was tested on the Fréjus landslide test case. As the model works with 256 × 256 pixel tiles and given that Sentinel-2 has a 10 m/px resolution, a “search area” of 2.56 km × 2.56 km for which its center is the Fréjus landslide’s location was manually defined. The model was tested by detecting the presence of landslides after 27 August 2023 in this area.

A pair of Sentinel-2 images depicting the area of the case study before and after the landslide occurrence on 27 August 2023 was then acquired. The pre-event image was taken on 3 July 2023, while the post-event scene was taken on the 6th of September 2023. The post-event image was the first cloud-free S2 image after the landslide occurrence. Choosing the first available post-event Sentinel-2 image is critical in an emergency response scenario to guarantee a relatively time-efficient (based on the satellite’s revisit period) emergency responder discovery of landslide occurrences, particularly for vast areas.

The two patches were pre-processed (band values were divided by 10,000 and clipped to [0,1]). The model was run on the patch pair, outputting a binary segmentation mask and rendering pixel values of 1 for “landslide” and 0 otherwise. The Sentinel-2 image pair, along with the predicted mask, is presented in Figure 5.

As can be seen in Figure 5, the output segmentation mask is able to correctly delineate the landslide. The landslide has a detected extension of 25 pixels, which corresponds to roughly 2500 m², given the Sentinel-2 resolution of 10 m/px.

Landslide Detection from Sentinel-1

The extent of the landslide was effectively identified (Figure 6) by calculating the significance of changes in radar backscatter before and after the event. These areas are depicted in shades ranging from yellow to red on the aerial view, with red indicating higher levels of rock movement and yellow indicating lower severity.

Two distinct areas of movement are visible. The first, located to the south and marked by dark red, represents the region where rock formations were dislodged. The second area, to the north, shows where the rocks settled around the Fréjus train line and tunnel entrance, consistent with images and reports from the area. This result encompasses the four landslides that occurred in the region during the full landslide timeline. The affected area, including both dislodged and settled rocks, covers approximately 2000 square meters. This result is compatible with the output of the Sentinel-2-based segmentation model presented in the previous section.

While the model successfully detected the Fréjus landslide, its small size relative to the 10 m/px resolution of Sentinel-2 limited the delineation’s precision, as the event occupied only a few pixels. The validation relied on the seismic intensity shake map from the Italian National Institute of Geophysics and Volcanology (INGV), as comprehensive ground-based damage assessments were unavailable. For the Fréjus landslide case study (August 2023), using Sentinel-1 SAR, no ground truth delineations were accessible, necessitating qualitative comparisons with visual reports and satellite imagery.

3.1.2. Wildfires

Wildfires were studied in the Piedmont case. Wildfire delineation and severity estimation models were assessed in the six areas of interest (AOIs) identified by the Copernicus Emergency Management Service–Rapid Mapping (EMSR)253 activation. The concerned areas are the municipalities of Locana, Susa, Condove, Cumiana, Casteldelfino, and Pietraporzio. A sample Sentinel-2 image taken after the wildfire is presented in Figure 7, along with a manual delineation of the wildfire trace produced for the EMSR253 activation and the delineation and severity estimation from the wildfire model.

Performance metrics were also computed to assess the agreement between the AI-based mapping and the Copernicus EMS manual mapping used as ground truth reference (

Table 4. Performance metrics achieved by the wildfire detection deep learning model on the 6 AOIs of the EMSR253 activation (2017 wildfires in Piedmont, Italy).

Areas of Interest	IoU	F1	Precision	Recall
AOI01	0.41	0.58	0.41	0.99
AOI02	0.62	0.77	0.64	0.95
AOI03	0.62	0.79	0.87	0.72
AOI04	0.50	0.67	0.97	0.51
AOI05	0.91	0.95	0.95	0.95
AOI06	0.99	1.00	0.99	1.00
Overall	0.56	0.72	0.66	0.79

). Precision, recall, intersection of union (IoU), and F1 score are key performance metrics used to evaluate how well a model detects specific features: in this case, wildfire-affected areas. Precision indicates how many of the areas predicted as burned by the model were actually correct, meaning it reflects the rate of false positives; a high precision means the model rarely labels non-burned areas as burned. Recall, on the other hand, measures how many of the truly burned areas were successfully detected by the model, with high recall meaning very few actual wildfires were missed (i.e., low false negatives). The F1 score combines these two metrics into a single number by taking their harmonic mean, providing a balanced measure that is especially useful when there is a trade-off between precision and recall. IoU complements these metrics by evaluating how well the predicted fire-affected area overlaps spatially with the ground truth, with higher IoU values indicating more accurate delineation of burned regions. As the EMSR253 activation does not include manually produced grading products, the performance of the severity estimation cannot be quantitatively estimated.

The presented results prove the capacity of the method but also reveal its limitations. For each AOI, the model can correctly identify the main area affected by the fire and estimate its severity. In a few AOIs, the boundaries are not precise, and sometimes, the model provides false-positive artifacts (i.e., AOI01).

3.1.3. Floods

Flood Delineation

The flood delineation model was applied to some of the AOIs from the EMSR680 activation for which manual flood delineation was made available. In particular, the performances of the flood delineated model were examined on the floods that occurred in the areas of Kranj (along the Sava River), Zalec (along the Savinja River), and Murska Sobota (along the Mura River). These zones are named AOI01, AOI02, and AOI03, respectively.

The Sentinel-1 pre-event image with an acquisition date of 28 July 2023 was first downloaded. The Sentinel-1 post-event input image was taken on 5 August 2023, only one day after the floods’ occurrence. The retained image is the same one used as reference for the EMSR680 activation used for producing manual delineations. A sample pre–post-Sentinel-1 pair is shown in Figure 8, along with the manual delineation of the flooded areas produced for the EMSR680 activation and the delineation predicted by the model.

The performance metrics (explained in Section 3.1.2) used to assess the agreement between the AI-based mapping were also checked against the Copernicus EMS manual mapping used as ground truth reference (

Table 5. Performance metrics achieved by the flood detection deep learning model on 3 AOIs of the EMSR680 activation (2023 floods in Slovenia).

Areas of Interest	IoU	F1	Precision	Recall
AOI01	0.70	0.78	0.81	0.76
AOI03	0.56	0.61	0.60	0.63
AOI04	0.53	0.56	0.55	0.56
Overall	0.60	0.65	0.65	0.65

).

Flood Risk

The modeling results of flood risks are shown in Figure 9. Comparing these results with the JRC data, a strong correlation between the flood risk model and the actual flood map can be observed (Figure 10) in Slovenia, with a zoomed-in view present in Figure 11. The prediction algorithm highlights areas with a very high risk of flooding in the darkest blue and areas with moderate risks in light blue, and it masks out areas with very low or no risk. A comparative analysis between the produced flood risk model and the JRC’s flood detection reveals that the high- and very high-risk areas correspond to the actual flooded regions, marked by red contours in the flood detection image. This similarity is evident in areas such as south of Murska Sobota, north of Ptuj, north of Ljubljana, and west of Velenje.

Slovenian Carinthia was one of the hardest-hit regions, with heavy rain causing rivers to swell and resulting in major overflows, particularly in the foothills of the Julian Alps from Idrija through the Ljubljana basin to Slovenian Carinthia, where 150–200 mm of rain fell. Styria, in the northeastern part of Slovenia, also experienced significant flooding, with heavy rainfall leading to the overflow of rivers and streams, causing damage to infrastructure and residential areas. Prekmurje, in the easternmost part of Slovenia, was similarly affected by heavy rainfall, resulting in the overflow of rivers and streams and significant damage, as can be seen in Figure 9. From the same figure, the main road networks can be seen to coincide with the highly flooded area, hence underlining the impact of this risk on CIR.

For the Slovenia flood case study (August 2023), flood risk maps were validated against manually produced delineation maps from Copernicus Emergency Management Service (EMSR680) and historical flood data from the Joint Research Centre (JRC).

3.1.4. Earthquake Damage Detection

Destroyed areas were effectively identified by calculating the significance of changes in radar backscatter before and after the earthquake event (Figure 12). These areas were represented in yellow to red shades on the aerial view of the affected area, with red indicating higher levels of destruction. These areas likely represent structures that were severely damaged or destroyed by the earthquake, aligning with the reported structural damage, where over half of the buildings in Amatrice were affected (Figure 13).

4. Discussion

In this section, a discussion of our study’s findings is provided.

4.1. Limitations of the Study

In this study, a methodology for assessing and monitoring landslides, floods, wildfires, and earthquake damages was presented. While several methods exist, this paper focused on the development of a robust methodology based on readily available Earth observation data. The main objective was to develop a fast and reliable method to allow the delineation of the hazards, assessing their impacts and determining how CIRs are affected.

The advantage of the approach proposed in this paper is its flexibility and relative simplicity. To test the validity of the proposed methods, results were compared against real test cases for landslides, earthquakes, wildfires, and floods. The results confirm that the developed AI hazard delineation models work well in the context of real emergency response scenarios. However, some mismatches were observed and can be attributed to the following seven factors:

• For Landslides: One way to increase the robustness of the mapping methodology would be to use multiple pre-event images, say $N$, pairing them with the post-event image and returning $N$ masks. The masks would then be aggregated by classifying pixels as “landslide” if their value is greater than $N/2$ (i.e., if at least half of the segmentation masks classify that pixel as landslide). This experimental setting is still compatible with an emergency response scenario, as many cloud-free pre-event images are typically available in any given location. However, caution must be taken in choosing pre-event images that are too distant from the date of the landslide, as too many naturally occurring landscape transformations could mislead the model. Despite the challenges of delineating landslides from pre–post-Sentinel-2 images (low resolution in relation to the size of typical landslides, potential presence of clouds, and relatively small training set), the model was successfully validated on the Fréjus Tunnel landslide case study despite its relatively low area (<2500 m²). Still, the 10 m/px resolution of Sentinel-2 imagery is a key limitation of this approach, as it effectively hampers the delineation of landslides below 100 m².
• For Subsidence: Interferometric coherence metrics for validating deformation signals could have been incorporated, and details on schemes for the range ambiguity suppression of spaceborne SAR based on underdetermined blind source separation are also important to mention. However, this study focused primarily on the application of the SBAS technique using available Sentinel-1 data, with an emphasis on achieving reliable long-term ground deformation estimates in the context of land subsidence. Regarding the use of coherence metrics, coherence is computed and taken into account during processing: only pixels with a coherence higher than a certain threshold (that is configurable) are used to extract the subsidence values. In future work, the integration of other approaches allowing lower values of coherence and, depending on the computed coherence, classifying the outputs as more or less reliable will be carried out.
• For Wildfires: These limitations can be addressed by experimenting with different scenarios. For example, the dataset can be expanded to include wildfire events outside of Europe. This will make the model’s detection of differences in the ground’s morphological characteristics more precise. Another possible approach could be to also exploit pre-fire satellite imagery and modify the model architecture accordingly. To improve the accuracy of the prediction, we can also try to artificially increase the native resolution of Sentinel-2 data by applying super-resolution techniques.
• For Flood Delineation: The flood delineation model (Section 2.3.3) uses the Sen1Floods11 dataset, which includes only one European flood event (Spain). This risks poor generalization to regions with distinct hydrology (e.g., Slovenia’s alpine rivers in AOI03–04).

However, the choice to use Sen1Floods11 was driven by several pragmatic and methodological reasons:

• Global Diversity and Sentinel-1 Consistency: Despite limited European cases, Sen1Floods11 covers over 400 flood events from diverse hydrological and geographic settings (e.g., deltaic, tropical, arid, temperate basins). This diversity enables models trained on it to learn generalizable radar backscatter patterns related to flood inundation. Since all events use Sentinel-1 imagery under consistent pre-processing protocols, the spectral signatures captured in the dataset are broadly transferable to other regions, including Europe.
• Lack of Open High-Quality Alternatives: Currently, Sen1Floods11 are one of the only openly available, pixel-labeled, large-scale SAR flood datasets suitable for supervised deep learning. While we acknowledge the European underrepresentation, no comparable labeled dataset with significant alpine or Slovenian examples currently exists in the public domain, which necessitated working within these limitations.
• Model Adaptation and Regional Post-Processing: To address the generalization challenge, post-classification refinements such as topographic masking, seasonally aware land cover filtering, and elevation constraints—tailored to the alpine characteristics of AOI03-04—were applied. These region-specific enhancements help bridge the gap between the training data context and the target environment, improving the accuracy of flood delineation in steep and narrow river valleys.

Accordingly, expanding datasets to include more alpine and European flood events in future work is a must for ensuring better accuracy.

5.

For Earthquake Detection: Technologies such as Synthetic Aperture Radar (SAR), Interferometric SAR (InSAR), optical satellite imagery, and LiDAR can detect ground deformation, surface ruptures, landslides, and damage to infrastructure with high spatial resolution and sub-centimeter accuracy. Remote sensing also aids in monitoring interseismic strain accumulation over months or years, which can indicate areas of increased seismic potential. However, despite its strengths, remote sensing has limited capabilities for short-term or real-time earthquake prediction due to the episodic nature of satellite passes, delays in data processing, and the absence of consistent, detectable precursors observable from space. Therefore, immediate earthquake detection and early warning remain reliant on dense, ground-based seismic networks and accelerometers rather than remote sensing.

6.

Differences Between Manual Delineation Maps and AI-Based Ones: This may be attributed to several factors. First, manual delineations are produced by experts, typically using commercial very-high-resolution (VHR) satellite or drone images as reference. For instance, the Piedmont wildfires (EMSR253) which we tested the model’s accuracy against were manually delineated using Airbus Pléiades-1A/B (0.5 m/px) and Airbus SPOT7/8 (2.0 m/px). Conversely, the developed model operates with Sentinel-1 and Sentinel-2 imagery, which are of lower resolution (10 m/px). Consequently, the automatic detection and delineation of hazards at that resolution are challenging tasks, especially for relatively small landslides and floods.

7.

Data Accuracy and Availability: The datasets used for training the models may be too limited in terms of geographical coverage. For example, the Sen1Floods11 dataset used for training the flood delineation model includes only one flood event in Europe (that occurred in Spain). Consequently, the models’ applicability in other regions of the world may be limited. However, image augmentation techniques mitigate this issue to some extent.

Despite these limitations, the models presented in this study can be considered valuable tools in the aftermath of a natural disaster, as they provide a relatively accurate outlook on the location, extent, and magnitude of damage as soon as a satellite image over the area of interest becomes available. Enhanced situational awareness granted by AI-based technologies is especially useful in the context of critical infrastructures: the predicted delineation maps are georeferenced and can thus be overlayed onto critical infrastructure maps in any version of any GIS software to help assess how much the disaster impacted them and to subsequently coordinate responses.

4.2. Replicability of the Methods

The flood risk detection approach allows a comprehensive assessment of flood risk by considering a wide range of factors. The use of machine learning and kriging for increasing the resolution of precipitation data is a crucial part of the methodology, enabling the capture of fine-scale variations in precipitation that can significantly influence flood risk. However, the main strength of the methodology lies in the integration of all inputs into a single equation, providing a holistic measure of flood risk. This approach enables the generation of detailed and accurate flood risk projections, which are crucial for effective flood risk management. Moreover, users can substitute their own inputs into the equation, where finer-scale data is applicable, which will ensure more accurate results.

The presented method for land subsidence monitoring demonstrates strong replicability and can be applied to a wide range of regions beyond the initial study areas, offering valuable insights into land subsidence dynamics in diverse environmental and infrastructural contexts. By leveraging the temporal approach and encompassing pre-event, peri-event, and post-event analyses, this methodology provides a comprehensive framework for understanding subsidence processes over time. This approach is particularly useful for assessing long-term trends, evaluating the impact of anthropogenic activities, and monitoring the effectiveness of mitigation strategies. The use of freely available and globally accessible Sentinel-1 SAR data ensures broad applicability and contributes to the adaptation of this approach in other contexts. Furthermore, the combination of multi-temporal interferometric techniques and robust processing workflows allows for detailed characterization of deformation patterns, offering actionable information for urban planners, infrastructure managers, and policymakers. This adaptability underscores the importance of addressing land subsidence challenges and supporting sustainable management across diverse scenarios

5. Conclusions

This study explores the effectiveness of remote sensing technologies, including synthetic aperture radar (SAR) and optical satellite imagery, for safeguarding European critical infrastructure against natural hazards through four case studies: the Fréjus landslide, Piedmont wildfires, Slovenia floods, and the Amatrice earthquake. Integrated with geographic information systems (GISs), automatic methods for hazard delineation and risk mapping enable timely monitoring, early hazard detection, and risk assessment across transport, energy, and communication sectors. These methods contribute to the risk-based design of CIR, enhancing classical infrastructure planning and prioritizing reinforcement measures while also facilitating coordinated emergency response post-disaster.

Based on the case study findings, three key recommendations are proposed to enhance critical infrastructure resilience (CIR). First, adopt multi-sensor remote sensing (SAR, optical, and digital elevation models) to ensure comprehensive monitoring of natural hazards, leveraging SAR’s cloud-penetrating capabilities and optical imagery’s visual detail, as shown in the case studies. Second, compile and share open-access catalogues of natural hazard events, such as those from the Copernicus Emergency Management Service, to support AI-based hazard delineation and improve model accuracy. Third, provide local authorities and infrastructure managers with training and user-friendly GIS tools to strengthen disaster response and recovery, building on the methodologies developed.

These findings and recommendations align with the ATLANTIS HORIZON project (https://www.atlantis-horizon.eu), acessed on 29 May 2025, which aims to enhance the resilience of EU critical infrastructures against natural hazards. The validated technologies offer authorities and civil protection services practical tools for preventing, responding to, and recovering from disasters, contributing to robust CIR. Future research should refine multi-temporal EO techniques, integrate real-time data with Internet of Things (IoT) sensors, and expand datasets to improve hazard mapping precision in complex environments.