Integrating GIS and machine learning for seismic vulnerability mapping in Al Hoceima, Morocco

Abstract

Urban centers along the Alboran–Rif margin face recurring seismic losses, yet city-scale vulnerability mapping often lacks event-informed labels and operational detail. We develop a reproducible framework for Al Hoceima that couples supervised machine learning with a curated inventory of 664 georeferenced buildings damaged during the 2004 and 2016 earthquakes and entries from the Risk-UE program. Seventeen predictors capture the structural, geotechnical, physical, social, and emergency-access conditions. After correlation screening to limit redundancy, four classifiers—Random Forest, XGBoost, Support Vector Machine, and Artificial Neural Network—were trained and validated, and their outputs were aggregated into five vulnerability classes using Jenks natural breaks. All models recover a coherent geography of risk, with very-high and high classes concentrated in the central, eastern, and south-western sectors, and low or safe classes dominant across northern and peripheral belts. Random Forest delivered the strongest performance with accuracy of 0.94, F1-score of 0.943, Kappa of 0.925, and area under the ROC curve of 0.98, while XGBoost performed closely and the remaining models were moderate. Feature-importance analysis identifies population density, distance to the epicenter, and peak ground acceleration as primary drivers, followed by access to fire stations, lithological site effects, and building age. The maps provide decision-grade guidance for retrofit targeting, land-use control, and emergency access planning. Remaining limitations include incomplete building-stock attributes in informal districts, scale and temporal inconsistencies among predictors, label scarcity, and class imbalance. The framework is immediately transferable to data-constrained Mediterranean and North African cities and can be strengthened by harmonized inventories, event-derived labels from UAV and SAR, and multi-hazard integration.