RTAS-Net: A ResNet-transformer-ASPP semantic segmentation network for remote sensing images.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Remote sensing image semantic segmentation faces challenges with scale variation and complex spatial distributions.
• These issues result in semantic discontinuity and loss of fine details for small objects.

Purpose of the Study:

• To propose RTAS-Net, a U-Net-based network designed to enhance feature representation for remote sensing semantic segmentation.
• To address scale variation and improve the recognition of both large regions and small objects.

Main Methods:

• RTAS-Net integrates Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context aggregation and Swin Transformer for cross-region dependency modeling.
• A lightweight mini-ASPP and MobileViT are employed at high-resolution skip connections to reinforce fine-scale information and local texture representations.
• The network unifies global semantics and local details through coordinated cross-level pathways.

Main Results:

• RTAS-Net demonstrated consistent improvements in mean Intersection over Union (mIoU), mean F1-score (mF1), and Overall Accuracy (OA) on benchmark datasets.
• Experimental results validate the effectiveness and practical applicability of the proposed method.
• The analysis confirmed improvements in parameter scale and inference efficiency.

Conclusions:

• RTAS-Net effectively enhances feature representation for remote sensing semantic segmentation.
• The proposed network successfully addresses scale variation and improves the segmentation of objects at different scales.
• RTAS-Net offers a practical and efficient solution for remote sensing image analysis.