XShadowBright

The “bright shadow” is a distinct bright region visible in cross-polarized (HV or VH) SAR images of moving ships, absent in co-polarized (HH, VV) images. It consistently appears in cross-pol imagery but never in co-pol. When correctly velocity-matched, its position aligns with the ship’s true location. The azimuth distance (∆x) between the shifted ship image and its shadow can be used to estimate the ship’s across-track velocity (V_tr). This phenomenon is observed in both C-band (RADARSAT-2) and X-band (TerraSAR-X), unlike ship wake methods, which are more detectable in X-band. Additionally, it is independent of the incidence angle, making it effective even at low angles where sea clutter in co-pol images is significant.

Physical Mechanism Behind the Bright “Shadow”

• The bright “shadow” appears due to strong volume scattering from disturbances created by the ship’s movement on the ocean surface.

• It is hypothesized that wave breaking, whitecap, foam, and spray effects around a moving vessel cause this scattering.

• The bright shadow is azimuthally shifted from the actual ship location due to the Doppler effect induced by the ship’s motion.

Dataset Creation Process:

The XShadowBright dataset containing 1,100 samples (50% train, 20% Validation, and 30% test) is generated by applying a series of spatial and noise-based augmentations to the original images coming from TerraSAR-X and their corresponding masks. These transformations help enhance the dataset by introducing variations in positioning, rotation, scaling, and noise while maintaining spatial consistency between the input image and its mask.

1. Spatial Augmentations (Applied to Both Image and Mask)

A shared augmentation pipeline is used to ensure identical transformations for both the input image and its corresponding mask. The following transformations are applied:

• Random Horizontal Flip (50%) – Simulates different viewing perspectives.

• Random Vertical Flip (50%) – Introduces further positional variations.

• Random Rotation (±40°) – Ensures rotational diversity while aligning interpolation.

• Random Resized Crop (Scale: 75% to 120%) – Introduces scaling variations while maintaining spatial structure.

• Random Affine Transform (Rotation ±20°, Translation ±10%, Scaling ±10%, Shear ±10°) – Simulates distortions due to imaging angles and sensor variations.

These augmentations preserve spatial coherence (using same_on_batch=True) and are crucial in scenarios such as polarimetric SAR-GMTI ship detection, where variations in imaging angles and object orientations need to be accounted for.

2. Additional Image-Only Transformations

After spatial transformations, additional image-only augmentations are applied to introduce noise and blur effects, mimicking real-world SAR data distortions:

• Gaussian Blur (Kernel: 5×5, Sigma: 0.1–2.0) – Simulates motion blur or radar-induced defocus.

• Gaussian Noise (Mean=0.0, Std=0.03, Applied 50%) – Emulates SAR noise, enhancing robustness in detection models.

To prevent numerical instability, the transformed image is clamped (clamp(min=1e-6)), ensuring all pixel values remain valid.