UNeXt: An Efficient Network for the Semantic Segmentation of High-Resolution Remote Sensing Images

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Semantic segmentation of remote sensing images is crucial for urban planning, disaster assessment, and environmental monitoring.
• Increasing spatial resolution of remote sensing data presents challenges like scale variation and information redundancy.
• Existing Transformer-based methods offer global context but suffer from high computational cost and loss of local details.

Purpose of the Study:

• To develop a real-time semantic segmentation model for high-resolution remote sensing images.
• To address the limitations of current methods, particularly the trade-off between computational complexity and detail preservation.
• To improve the efficiency and accuracy of intelligent remote sensing data interpretation.

Main Methods:

• Proposed UNeXt, a novel semantic segmentation model combining UNet, ConvNeXt, and Transformer architectures.
• Utilized a lightweight ConvNeXt-T as the encoder and a Transnext decoder integrating Transformer and Convolutional Neural Networks (CNNs).
• Introduced a SC Feature Fuse Block (SCFB) to enhance spatial and channel information utilization and reduce complexity.

Main Results:

• UNeXt achieves real-time performance with 97 fps for 512x512 inputs on a single NVIDIA GTX 4090 GPU.
• The model attained high accuracy, with mIoUs of 85.2% on Vaihingen and 82.9% on Gaofen5 (GID5) datasets.
• Demonstrated superior performance over state-of-the-art lightweight models in both speed and accuracy.

Conclusions:

• UNeXt offers an efficient and effective solution for semantic segmentation of high-resolution remote sensing images.
• The proposed architecture successfully balances global context capture with local detail preservation.
• The model shows significant potential for practical applications in intelligent remote sensing data analysis.