Data and Methods

Generative AI architecture

The core architecture for the Generative AI-powered demonstrator leverages a combination of key components, including AWS Bedrock, LLaMA 3, LangChain, Django, and datasets from OGC-compliant services. AWS Bedrock serves as the backbone of the system, providing the necessary infrastructure to host and deploy the LLaMA 3 language model (see Figure 2). This ensures the system can efficiently handle large datasets, perform model inference, and generate real-time responses. LLaMA 3 functions as the core Generative AI model, enhancing the demonstrator’s ability to deliver context-aware and data-driven insights for coastal resilience applications.

Figure 2: Generative AI-powered demonstrator architecture

Coastal Vulnerability Index (CVI) calculations

In detail, we followed a systematic approach to calculate the coastal vulnerability across Europe and Figure 3 presents a structured workflow for estimating the CVI by integrating multiple geospatial and environmental parameters as well as different spatial queries and AI-driven insights. Thus, a structured workflow for calculating the CVI through a multi-step geospatial analysis framework was developed and the process is divided into five key stages (A to E).

Figure 3: A workflow presenting all CVI calculation processes (Adapted by Theocharidis et al. 2024 [2]).

Stage one (Process A) incorporates the study area selection, the bounding box extraction, and the coastline retrieval using the Overpass API of OpenStreetMap (OSM). Process B involves generating transects along the extracted coastline in .geojson format and assigning spatial data to each transect for CVI computation of Process C. All transects were overlayed with raster files (e.g., elevation, slope, landcover) at the intersecting cells and for all vectorized data (i.e. grid points of Copernicus Data), the closest grid point or feature centroid was joined to the transect and the attributes of the corresponding data.

The aforementioned criteria are sourced from different APIs, including OpenEO, Copernicus Marine Toolbox, and Deltares, ensuring comprehensive data integration. These datasets help in identifying key parameters across three major factor categories:

• Geological Factors: These factors describe the physical characteristics of the coastal region, influencing its susceptibility to erosion and geomorphological changes. The key parameters include:

  • Coastal Slope: Determines how steep or gradual the coastline is, affecting wave energy dissipation.

  • Rate of Coastline Erosion: Measures how quickly the coastline is retreating due to natural or anthropogenic influences.

  • Coastal Elevation: Influences the likelihood of coastal inundation and flood risk.

• Hydro-physical Factors: These factors represent dynamic oceanographic and climatic influences on coastal vulnerability. The key parameters include:

  • Mean Tidal Range: Affects water level fluctuations and coastal inundation risks.

  • Mean Significant Wave Height: Determines wave energy impact on coastal erosion.

  • Relative Sea Level Rise: Reflects long-term changes in sea levels due to climate change and land subsidence.

• Socio-economic Factors: These factors assess human activities and land use that influence coastal resilience. The main parameter considered is:

  • Land Use/Land Cover: Indicates human settlements and changes of the coastal environment, which can either mitigate or exacerbate vulnerability.

Having identified and extracted the criteria and all spatial datasets, the CVI is computed for each transect using the formula:

CVI =

, where each variable represents a key coastal factor such as elevation, slope, erosion rate, tidal range, wave height, and land use. This formula integrates multiple environmental and socio-economic parameters to quantify coastal vulnerability. The result helps identify high-risk coastal areas for targeted resilience planning.

Table 1: CVI Criteria Rankings

	CVI Score
Factor	Very Low (1)	Low (2)	Moderate (3)	High (4)	Very High (5)	Availability
Coastal land cover	Tree/Forest	Shrubland, bare soil	Cropland, grassland	Herbaceous wetland	Urban	✔
Coastal slope (%)	>12	8 – 12	4 – 8	2 – 4	<2	✔
Rate of coastline erosion (m/year)	>2	(+1) – (+2)	(–1) – (+1)	(–1.1) – (–2)	≤(–2)	✔
Mean tidal range (m)	>6	4 – 6	2 – 4	1 – 2	<1
Mean significant wave height (m)	0 – 0.55	0.55 – 0.85	0.85 – 1.05	1.05 – 1.25	>1.25	✔
Coastal elevation (m)	≥ 20	10 – 20	5 – 10	2 – 5	0 – 2	✔
Relative sea-level rise (mm/year)	<1.8	1.8 – 2.5	2.5 – 3	3 – 3.4	>3.4	✔

The final steps integrate the Raking Method (scale 0-5) applied to rank vulnerability factors, visualizing all sub-criteria in LLaMA 3 and the CVI results for the study area selected by the user (Process D). These datasets were assigned to transects by overlaying raster data or linking vector data to the closest points of each transect. Indicators of vulnerability were then scored on a scale from 0 (low) to 5 (high) (see Table 1). The CVI for each transect was calculated using a composite formula and mapped to visualize vulnerability levels: low (green), moderate (yellow), and high (red), In the last phase (Process E), the CVI results are transferred to LLaMA 3 for accuracy analysis and uncertainty assessment, ultimately providing feedback to the user.

References

[2]

Theocharidis, C., Doukanari, M., Kalogirou, E., Christofi, D., Mettas, C., Kontoes, C., Hadjimitsis, D., Argyriou, A. V., & Eliades, M. (2024). Coastal Vulnerability Index (CVI) Assessment: Evaluating Risks Associated with Human-Made Activities along the Limassol Coastline, Cyprus. In Remote Sensing (Vol. 16, Issue 19, p. 3688). MDPI AG. https://doi.org/10.3390/rs16193688