Detection of Partially Structural Collapse Using Long-Term Small Displacement Data from Satellite Images

Area of Science:

• Geospatial Engineering
• Artificial Intelligence
• Civil Engineering

Background:

• Satellite sensors and interferometric synthetic aperture radar (InSAR) offer potential for long-term structural health monitoring (SHM).
• Limited image availability (small data) and environmental/operational variability pose significant challenges for SAR-based SHM.

Purpose of the Study:

• To propose novel hybrid unsupervised learning methods to overcome small data and variability issues in SAR-based SHM.
• To enhance the accuracy and reliability of long-term structural monitoring using satellite data.

Main Methods:

• Data augmentation using the Markov Chain Monte Carlo algorithm.
• Feature normalization, including an artificial neural network-based approach with iterative hyperparameter selection.
• Unsupervised teacher-student learning combining undercomplete deep and overcomplete single-layer neural networks.
• Decision making utilizing Mahalanobis-squared distance.

Main Results:

• Validation using limited long-term displacement samples from TerraSAR-X SAR images.
• Demonstrated effectiveness of the proposed methods in addressing key challenges of SAR-based SHM.
• Successful handling of small data and environmental variability in structural monitoring.

Conclusions:

• The developed hybrid unsupervised learning methods are effective for SAR-based structural health monitoring.
• These methods provide a robust solution for long-term monitoring of civil structures with limited satellite data.
• The study contributes to advancing the application of AI and remote sensing in structural engineering.