SAR remote sensing image segmentation based on feature enhancement

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Synthetic Aperture Radar (SAR) images are vital for remote sensing, offering consistent imaging capabilities.
• SAR image analysis faces challenges including speckle noise and unclear boundaries in high-resolution data.
• Existing methods struggle to effectively address both noise and boundary ambiguity in SAR imagery.

Purpose of the Study:

• To develop an advanced SAR remote sensing image segmentation method.
• To enhance feature representation and mitigate speckle noise in SAR images.
• To improve the clarity of boundary information in SAR image segmentation.

Main Methods:

• A feature enhancement approach combining wavelet transform with an encoder-decoder network was employed.
• A cascaded encoder-decoder based post-processing refinement module was designed for boundary clarification.
• A self-distillation module was integrated into the encoder to improve semantic information learning.

Main Results:

• The proposed method effectively enhances feature expression and reduces speckle noise.
• The refinement module significantly clarifies boundary information in segmentation results.
• The self-distillation module improved the learning of semantic information by the encoder.

Conclusions:

• The developed SAR image segmentation method demonstrates superior performance.
• The approach successfully addresses key limitations in SAR image analysis.
• The findings validate the effectiveness of the proposed techniques on benchmark datasets.