Channel reconstruction and dual attention dynamic fusion for remote sensing image semantic segmentation

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Increasing spatial resolution in remote sensing imagery leads to greater data complexity.
• Challenges in semantic segmentation include wide imaging ranges, dispersed similar land objects, complex boundaries, and dense small targets.
• Existing methods struggle to effectively integrate global and local contexts for accurate segmentation.

Purpose of the Study:

• To propose a novel semantic segmentation network, CRDFNet, for remote sensing images.
• To effectively integrate global and local contexts to address segmentation challenges.
• To improve segmentation accuracy for complex shapes and small targets.

Main Methods:

• Developed a channel reconstruction and dual attention dynamic fusion network (CRDFNet).
• Introduced a channel feature aggregation module (CFAM) to enhance high-resolution details and aggregate multi-scale features.
• Designed a dual attention feature refinement module (DAFRM) for precise small target segmentation using dynamic fusion.

Main Results:

• CRDFNet effectively integrates global and local contexts, improving segmentation accuracy.
• CFAM enhances feature fusion and detail preservation for complex boundaries.
• DAFRM achieves precise segmentation of small targets.
• Experimental results on Potsdam, Vaihingen, UAVid, and MSIDBG datasets show CRDFNet outperforms existing methods in F1 score, OA, and mIoU.

Conclusions:

• CRDFNet demonstrates superior performance in remote sensing semantic segmentation.
• The proposed network effectively handles complex image characteristics like intricate boundaries and small targets.
• CRDFNet offers a significant advancement for accurate geospatial information extraction from high-resolution imagery.