Anomaly Detection in Historical Arcades Using Transfer Learning

This paper introduces a novel methodology for detecting anomalies in historical modular structures (e.g., arcades or bridges) using satellite data. The approach assumes that structural displacements observed by satellites are primarily driven by environmental factors. Nevertheless, each module/segment of the structure responds differently based on its geometry and soil conditions. The proposed method relies on transfer learning techniques such as domain adaptation to harmonize statistical features (specifically, the statistical moments of InSAR displacements) across different segments of the modular structure. After harmonization, the features of all segments are expected to follow consistent temporal trends driven by environmental inputs. Structural anomalies are thus identified as segments for which displacement features deviate from expected behavior. The methodology is applied to the Portico di San Luca, a historical arcade and UNESCO World Heritage site in Bologna, Italy. Anomalies detected using the proposed approach are validated against average velocity maps generated considering five years of data and preliminary inspections on site.