Deep transfer learning: new approach for predicting seismic vulnerability

Abstract

Seismic areas known for the catastrophic impact in both human and material loss, are considered as areas where identifying and predicting vulnerability is challenging specially when data is unavailable or its collection is time consuming and costly. This paper aims to predict vulnerability of Al Haouz, Morocco, particularly Amizmeez region (SSAm) recently affected by the September 8, 2023 earthquake, based on the seismic characteristic, collected from seismic area of both Agadir, Morocco (SSAg) and Turkey (SSTu). To achieve this goal a deep transfer learning domain adaptation framework was employed, mainly performing data processing and augmentation, followed by domain adaptation via the Conditional Domain Adversarial Network (CDAN). To enhance feature alignment, the Maximum Mean Discrepancy (MMD) method was applied before training the model. The framework was first validated on labeled SSAg data, achieving 82% accuracy, before being deployed to predict vulnerability in SSAm. Results were validated using both geological maps and visual damage, detected using deep learning YOLOv11 model based on satellite images, demonstrating its ability to identify vulnerability at locations with different seismic characteristics. Ultimately, the study demonstrates the effectiveness of domain adaptation in seismic risk assessment, offering a scalable solution for predicting vulnerability in regions with limited labeled data.