AI Reads Building Footprints to Map Earthquake Vulnerability

Earthquake risk models are only as good as the building inventories behind them, and in much of Central America and the Caribbean those inventories barely exist. A team of Spanish researchers now reports an automated method that pulls key seismic vulnerability traits straight from building footprints digitised on aerial images, packaged in an open-source Python tool called SeismicBuildingExposure. The work, published in Advanced Modeling and Simulation in Engineering Sciences, was trained and tested on three pilot districts in San José, Guatemala City and Santo Domingo, and its error proved comparable to the spread between human surveyors doing the same measurements.

Geospatial exposure databases, known as GEDs, store one record per building with attributes that govern how it will shake in an earthquake. Building them traditionally means costly manual and field procedures, which is why many data-scarce regions simply skip the step. The new methodology starts with a semi-automatic digitalisation stage: fine-tuned instance segmentation models extract initial building geometries from high-resolution aerial imagery, and a user can then refine the outlines with minimal guidance. From those footprints the tool automatically computes building direction, relative position within an urban block, and plan irregularity indices covering slenderness, eccentricity and setbacks.

The central innovation is not the geometry itself but the translation of several international seismic codes into deterministic geometric algorithms. The team converted provisions from Eurocode 8, the American ASCE-7 standard, Mexico’s NTC-23 code for Mexico City, and the Italian GNDT Level II manual into computations that previously depended on subjective expert judgement. Because the rules are now explicit, the same footprint always yields the same classification, eliminating the variability that plagues traditional manual surveys.

Behaviour modifiers adjust the base vulnerability assigned to a building typology, acknowledging that real conditions can significantly raise or lower seismic response. Their importance is not theoretical. After the 2011 Lorca earthquake in Spain, researchers showed that properly accounting for these modifiers reduced the discrepancy between observed and estimated damage. The modifiers that can be derived from footprints include direction, plan irregularity, shape of the building plan and block position, and the study folds the last two into a single descriptor called the Footprint Shape Index.

Block position matters because aggregated buildings behave very differently depending on where they sit. Isolated units, unconnected to neighbours, tend to show a higher probability of damage than embedded ones, mainly for lack of confining effects. Corner units concentrate shear stresses and behave asymmetrically because they are exposed to lateral loads on one side only. Edge units remain susceptible to torsional effects and differential displacements, while central units usually enjoy smaller relative displacements thanks to the mass and confinement of the ensemble. The new algorithm classifies each footprint as isolated, corner, edge, confined or torque by simulating contact forces along the polygon’s sides and analysing the resultant forces, using what the authors describe as direction-aware geometric proxies rather than simple shared-boundary lengths.

Shape irregularity is treated with equal rigour. Slenderness, the ratio of a building’s length to its width, is computed either from the second moment of area of the footprint polygon or from a minimum bounding box, with the second-moment approach recommended by default. Eccentricity, the distance between the centre of mass and the centre of stiffness, feeds the eccentricity and radius ratios that Eurocode 8 caps at 0.3 and 1 respectively, while slenderness is limited at 4 and compactness at 0.95. Setback ratios, hole ratios and a newly proposed parallelity angle, which flags triangular footprints when the length-weighted mean angle of edges to a bounding box exceeds roughly 5 to 10 degrees, cover the ASCE-7 and NTC-23 provisions. The GNDT manual’s beta indices for masonry and concrete buildings are likewise automated, including the adaptation of the code’s side lengths to arbitrary real-world polygons.

The Footprint Shape Index itself is deliberately simple. It is a rule-based categorical descriptor, not a weighted composite score, assigning each building to exactly one of four mutually exclusive classes: Regular, Eccentricity, Shape or Slenderness. A building counts as regular only if it meets code thresholds on all three underlying attributes. When several criteria fail, a hierarchy applies in which slenderness takes precedence over shape and shape over eccentricity, so the index names the governing irregularity mode rather than counting violations.

Digitalisation was the bottleneck the team attacked first. They fine-tuned two transformer-based segmentation models, Mask2Former and SAM2, on a small domain-specific dataset split 80/20 into 82 training and 21 test images, with augmentation by brightness and contrast changes, cropping and flipping. Despite the modest training set, the fine-tuned Mask2Former reached performance levels comparable to other instance segmentation benchmarks, while SAM2 produced sharper, more geometrically precise boundaries and improved further after fine-tuning. Two workflows are offered: a sequential one in which Mask2Former generates a fully automated first-pass inventory that SAM2 then helps a human operator refine, and an alternative in which SAM2 works alone with point prompts, needing no fine-tuning, where high fidelity matters from the start. A benchmark against Microsoft’s Global ML Building Footprints dataset showed that the global model performed very poorly in these dense, disordered urban centres, appearing reliable only for isolated, clearly defined structures unrepresentative of the complex clusters that dominate the pilot areas.

Validation leaned on detailed manually annotated field inventories from the three pilot areas, selected for their representative construction systems and broad attribute diversity. For relative position, experts labelled every footprint in the pilot regions, and the automatic classification overlapped strongly with those labels; the deterministic algorithm needs no training data, so the whole region served as a test set. Most discrepancies traced back to differing definitions of contact between buildings, since the automatic process treats structures as touching when separated by less than a minimum distance tied to image pixel size, 15 centimetres in this study. A sensitivity analysis showed accuracy peaking with buffers of 5 to 10 centimetres and staying above 60 percent up to 50 to 100 centimetres, meaning the method works with high-resolution satellite imagery as well as aerial surveys. For shape, surveyors measured 25 randomly selected buildings per region, and the mean absolute error between the automatic method and the survey average was comparable to the variability among surveyors themselves. Shape indices also proved stable as footprint resolution degraded, with most code parameters changing little even with buffers up to 2 metres, the notable exception being slenderness, whose error grows as rounding reduces a building’s elongation.

The authors are explicit that the workflow is a first-order approximation, not a substitute for detailed structural surveys. It ignores openings, assumes continuous uniform walls, models a single storey of three metres, and treats each footprint as one structural unit, so expansion or seismic joints and internal subdivisions that are not externally visible cannot be detected, which may overestimate plan irregularity for large multi-block buildings. Vertical irregularities, soft storeys and building height fall outside the 2D scope and are left to future work, as are material identification and machine-learning classification of structural systems. For reinforced-concrete or frame buildings, the simplifications may overestimate eccentricity, so the indices should be read as robust footprint-based proxies rather than exact engineering quantities. Even so, the team argues the framework is a substantial improvement over simplified exposure models that omit these modifiers altogether, and in post-disaster contexts it could screen pre-event imagery to flag buildings with unfavourable geometry, supporting rapid triage in exactly the regions where cadastral data is scarcest.

Subject of Research: Automated estimation of seismic behaviour modifiers from building footprints for exposure databases

Article Title: A methodology for the automated estimation of footprint-derived seismic behaviour modifiers in building exposure assessment

Article References: Ureña-Pliego, M., Rodríguez-Saiz, J., Núñez-Álvarez, G., Marchamalo-Sacristán, M., & González-Rodrigo, B. (2026). A methodology for the automated estimation of footprint-derived seismic behaviour modifiers in building exposure assessment. Advanced Modeling and Simulation in Engineering Sciences, 13(1), Article 3. https://doi.org/10.1186/s40323-026-00323-y

Image Credits: AI Generated

DOI: 10.1186/s40323-026-00323-y

Keywords: seismic risk, building exposure, footprint digitalisation, instance segmentation, plan irregularity, Eurocode 8, GNDT, ASCE-7, NTC-23, open-source software, Central America, vulnerability assessment