Data‐Driven Zonation for Seismic Exposure Modeling via Machine Learning

Loss estimation for seismic risk studies requires an exposure model that captures the spatial distribution of buildings, accounting for their seismic vulnerability and population. A key step in developing such models is delineating homogeneous zones (HZs), urban areas where buildings share similar construction practices. Traditionally, HZs are defined through expert judgment and visual interpretation, a time‐consuming process that is difficult to scale and reproduce. This article presents a data‐driven methodology that leverages census, cadastral, and geospatial information integrated through unsupervised machine learning to automate HZ delineation. The workflow combines socioeconomic strata, building‐height descriptors, and spatial proximity, applies dimensionality reduction, and uses density‐based clustering to estimate the number of zones followed by medoid‐based partitioning for complete assignment. When applied to a major city in Colombia within the framework of the Colombian National Seismic Risk Model (MNRS), the proposed approach produces a coarser but operationally efficient stratification (13 zones versus 43 in the MNRS) and increases the average coverage per zone (598 versus 181 census blocks; 17,843 versus 5,394 buildings per zone). The case study indicates that the resulting exposure distributions closely match MNRS outputs while reducing inspections by about 68%. Statistical comparisons show no significant differences in story‐count distributions ( p = 0.082) and negligible divergence in wall‐material composition (Jensen–Shannon distance = 0.001), supporting comparable exposure characterization with substantially reduced fieldwork. Overall, the results demonstrate that data‐driven zoning can provide a scalable and reproducible alternative to expert‐defined stratification for exposure modeling in data‐limited settings.