Learning Ground Displacement Signals Directly from InSAR-Wrapped Interferograms.

Area of Science:

• Geosciences
• Remote Sensing
• Artificial Intelligence

Background:

• Ground displacement monitoring is crucial for early geohazard detection.
• Interferometric Synthetic Aperture Radar (InSAR) offers sub-millimeter accuracy but requires expertise and complex data handling.
• Automated systems for direct ground displacement indication from InSAR data are highly desirable.

Purpose of the Study:

• To evaluate the feasibility of using machine learning algorithms for automated ground displacement detection from InSAR data.
• To compare the performance of different machine learning models in classifying ground movement patterns.
• To assess the generalizability and robustness of trained models in diverse geographical areas.

Main Methods:

• Utilized Sentinel-1 InSAR data, including filtered-wrapped interferograms and coherence maps.
• Applied a high-pass filter to interferograms to isolate displacement signals.
• Trained and tested machine learning models (including Cosine K-nearest neighbor) using labeled pixels representing different ground movement velocities.
• Incorporated pseudo-labeling to enhance model generalizability and tested on data from Italy, Portugal, and the United States.

Main Results:

• Machine learning models successfully identified patterns associated with slow and fast ground movements.
• The Cosine K-nearest neighbor model achieved the highest test accuracy.
• Models demonstrated good performance on test sets from adjacent areas, indicating generalizability.
• The lowest test accuracy achieved was 80.1%.

Conclusions:

• Automated ground displacement detection using machine learning from InSAR data is feasible and advantageous.
• The Cosine K-nearest neighbor model shows significant potential for reliable geohazard monitoring.
• The developed approach offers a robust method for assessing ground movement, even in previously unencountered regions.