RST-Net: A Semantic Segmentation Network for Remote Sensing Images Based on a Dual-Branch Encoder Structure.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• High-resolution remote sensing images require effective semantic segmentation.
• Current methods struggle with fusing global and local features, leading to lost dependencies and blurred details.
• Limited adaptability to multi-scale object segmentation is a key challenge.

Purpose of the Study:

• To propose RST-Net, a novel semantic segmentation network for high-resolution remote sensing images.
• To address the limitations of inadequate feature fusion and multi-scale object segmentation.
• To enhance the extraction and fusion of both local spatial and global contextual information.

Main Methods:

• Developed RST-Net with a dual-branch encoder: a CNN branch (ResNeXt-50) for local features and a Shunted Transformer (ST) branch for global context.
• Integrated a multi-scale feature enhancement module (MSFEM) using atrous and depthwise separable convolutions for dynamic feature aggregation.
• Incorporated a residual dynamic feature fusion (RDFF) module in skip connections to improve encoder-decoder feature interaction.

Main Results:

• RST-Net achieved high performance on the Vaihingen and Potsdam datasets.
• Achieved Mean Intersection over Union (MIoU) scores of 77.04% on Vaihingen and 79.56% on Potsdam.
• Demonstrated significant improvements in semantic segmentation accuracy and detail preservation.

Conclusions:

• RST-Net effectively overcomes limitations in fusing global and local features for remote sensing image segmentation.
• The proposed network shows strong adaptability to multi-scale object segmentation.
• RST-Net validates its effectiveness and promising performance in complex remote sensing scenarios.