PSI - Issue 84

Francesca Ceccato et al. / Procedia Structural Integrity 84 (2026) 599–606

601

was used as the base layer. From the DEM, primary terrain variables were derived, including slope, aspect, northness, eastness, hillshade, Topographic Position Index (TPI). Height Above Nearest Drainage (HAND) was derived from the MERIT Hydro dataset. In addition, several curvature metrics (Gaussian, mean, horizontal, vertical, maximum, and minimum curvature) were computed to characterize local convexity, concavity, and slope deformation relevant to hydro-geomorphological instability. Vegetation conditions were characterized using Sentinel-2 Level-2A imagery (COPERNICUS/S2_SR) for 2023. The red and near-infrared bands were used to compute NDVI, and a median composite was generated to minimize cloud and seasonal effects before reprojection to 30-m resolution. Land cover was derived from the Google Dynamic World V1 dataset (2023) using a median composite and mode aggregation to produce a 30-m Land Use/Land Cover (LULC) map. Soil properties, including bulk density, clay content, and sand content at 200 cm depth, were obtained from OpenLandMap and standardized to the study area and resolution.

Table 1. Data sources.

Data

Resolution Source

URL

NDVI (Normalized difference vegetation index) monthly average over a year

Google Earth Engine Google Earth Engine Google Earth Engine

https://developers.google.com/earth engine/datasets/catalog/landsat-8

30 m

NASA SRTM Digital Elevation 30m

https://developers.google.com/earthengine/datasets/catalog/USGS _SRTMGL1_003#description https://developers.google.com/earth engine/datasets/catalog/OpenLandMap_SOL_SOL_BULKDENS FINEEARTH_USDA-4A1H_M_v02#description https://www.progettoiffi.isprambiente.it/en/italian-landslide inventory-iffi/ ; https://doi.org/10.5281/zenodo.8009366 (Hengl et al., 2017) GitHub - ISRICWorldSoil/SoilGrids250m: Global spatial predictions of soil properties and classes at 250 m resolution OpenLandMap_SOL_SOL_CLAY-WFRACTION_USDA 3A1A1A_M_v02 - Earth Engine Code Editor https://code.earthengine.google.com/?scriptPath=Examples:Datase ts/OpenLandMap_SOL_SOL_SAND-WFRACTION_USDA 3A1A1A_M_v02 GOOGLE_DYNAMICWORLD_V1 - Earth Engine Code Editor https://disc.gsfc.nasa.gov/datasets/GPM_3IMERGM_07/summary

30 m

Soil Density

250m

IFFI and ITALICA catalogue

Landslide inventory in Italy

Rainfall (Precipitation) _monthly average over a year

Google Earth Engine

11132 m

ISRICWorldSoil/ SoilGrids

Depth to bedrock (up to 2.4 m) 250 m

Weight % of clay particles

250 m

GEE

Weight % of sand particles

250 m

GEE

LULC

10 m

GEE

Lithology

250 m

GLiM

Global Lithological map

Additional datasets included depth to bedrock (Hengl et al., 2017), lithology from the Global Lithological Map (GLiM), and mean precipitation for 2023 from the CHIRPS dataset. Lithology and depth-to-bedrock layers were resampled to 30 m and uploaded to GEE to ensure consistency with other variables. All topographic, hydrological, climatic, soil, vegetation, and geological layers were harmonized to a common projection and resolution and combined into a single multi-band image. This unified predictor stack, exported as a GEE asset at 30-m resolution, constitutes the complete input dataset for subsequent landslide susceptibility modelling. The model was trained and tested using a comprehensive landslide inventory compiled from the IFFI Inventory of Landslides in Italy (ISPRA) and the ITALICA dataset. Landslide susceptibility was assessed at the pixel scale using a Random Forest (RF) classifier implemented in Google Earth Engine (GEE) (Huang et al., 2020). RF was selected for its robustness in handling high-dimensional datasets (Prakash et al., 2020; Merghadi et al., 2020). The dataset was split into 70% for training and 30% for testing, including 280 slow-moving and 800 rapid landslide points. Model performance was optimized using the out-of-bag error and evaluated with the Area Under the Receiver Operating Characteristic Curve (AUC–ROC). Feature importance was also extracted to support model interpretation.