AIS-Based Radar Error Correction Using a Vision Transformer Variant for Range and Azimuth Error Reduction.

Area of Science:

• Maritime surveillance
• Radar signal processing
• Machine learning for remote sensing

Background:

• Shore-based maritime surveillance radars exhibit systematic range and azimuth errors.
• These errors degrade the accuracy of target-tracking systems.

Purpose of the Study:

• To develop a novel method for correcting systematic radar errors using a Vision Transformer (ViT) variant.
• To leverage Automatic Identification System (AIS) data as ground truth for error correction.
• To model complex, nonlinear error patterns without relying on explicit physical radar models.

Main Methods:

• A Vision Transformer (ViT) variant was proposed to correct radar systematic errors.
• The model utilizes Automatic Identification System (AIS) data as ground truth for calibration.
• Self-attention mechanisms were employed to model nonlinear error patterns.

Main Results:

• The proposed ViT variant significantly reduced range Mean Absolute Error (MAE) by 98.5% (514.76 m to 7.77 m).
• Azimuth MAE was reduced by 89.8% (1.37° to 0.14°).
• Controlled experiments confirmed contributions from self-attention, patch embedding, and multi-task learning, enhancing robustness.

Conclusions:

• Transformer-based architectures are viable for correcting systematic radar errors in maritime surveillance.
• The proposed method offers a significant improvement in target-tracking accuracy.
• The approach demonstrates robustness, particularly in handling tail-error cases.