A Complex-Valued SAR Foundation Model Based on Physically Inspired Representation Learning.

Area of Science:

• Remote Sensing
• Computer Vision
• Geophysics

Background:

• Foundation models excel in remote sensing tasks due to their generalization capabilities.
• Synthetic Aperture Radar (SAR) provides crucial all-weather, all-day Earth observation data.
• Existing SAR interpretation models face challenges in information utilization and interpretability.

Purpose of the Study:

• To develop a physically interpretable remote sensing foundation model for complex-valued SAR data.
• To improve information utilization and generalization in SAR image interpretation.

Main Methods:

• A novel foundation model pre-trained by simulating polarimetric decomposition on complex-valued SAR data.
• Utilized scattering queries representing scattering bases interacting with SAR features.
• Employed polarimetric decomposition loss and power self-supervised loss for guided pre-training.

Main Results:

• Achieved state-of-the-art performance across nine diverse downstream interpretation tasks.
• Demonstrated stable feature representation extraction and strong generalization capabilities.
• Showcased effectiveness even in data-scarce scenarios.

Conclusions:

• The proposed foundation model effectively integrates physical interpretability with advanced deep learning.
• The model offers a significant advancement for SAR image analysis and Earth observation applications.
• The approach shows promise for robust remote sensing intelligence in challenging conditions.