A Cross-Domain Mamba Network with joint spatial-frequency learning for robust SAR oil spill detection.

Area of Science:

• Environmental monitoring
• Remote sensing technology
• Marine pollution control

Background:

• Marine oil spills present significant ecological risks.
• Synthetic Aperture Radar (SAR) is crucial for oil spill detection.
• Distinguishing oil spills from look-alikes and analyzing frequency-domain features remain challenges in SAR imagery.

Purpose of the Study:

• To develop a robust SAR oil spill detection method.
• To enhance the discrimination between oil spills and look-alikes.
• To improve the accuracy of oil spill detection, especially for thin films and weak boundaries.

Main Methods:

• Proposed the Cross-Domain Mamba Network (CDMNet) for SAR oil spill detection.
• Utilized a Cross-Domain Mamba Block (CDMamba) for joint spatial and frequency representation modeling.
• Incorporated a Scale-Aware Pyramid Pooling (SAPP) module for multi-scale context and boundary fidelity.
• Employed a Multi-Level Feature Fusion Module (MFFM) for integrating semantic and spatial features.

Main Results:

• CDMNet demonstrated superior performance over existing methods on SOS and M4D datasets.
• The proposed method effectively differentiates oil spills from look-alikes.
• High accuracy was achieved in localizing thin films and delineating weak boundaries.

Conclusions:

• CDMNet offers a robust and accurate solution for SAR-based oil spill detection.
• The joint analysis of spatial and frequency domains significantly improves detection capabilities.
• The network architecture effectively handles complex SAR imagery challenges.