SCU-Net: Semantic Segmentation Network for Learning Channel Information on Remote Sensing Images

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Semantic segmentation of remote sensing images requires models balancing computational efficiency and prediction accuracy.
• Existing methods often struggle to extract fine-grained details and generalize well across diverse datasets.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) architecture, the channel upsampling network (SCU-Net), for enhanced semantic segmentation of remote sensing imagery.
• To introduce a new upsampling convolution-deconvolution module (CDeConv) to improve feature extraction and learning efficiency.

Main Methods:

• Designed a novel CDeConv module for effective upsampling in CNNs.
• Developed SCU-Net, integrating CDeConv with a channel attention mechanism for semantic segmentation.
• Evaluated SCU-Net performance on remote sensing datasets.

Main Results:

• SCU-Net-102-A achieved a mean intersection-over-union (MIOU) of 55.84%, pixel accuracy of 91.53%, and FWIU of 85.83%.
• Demonstrated superior performance compared to state-of-the-art methods in learning detailed channel information.
• Exhibited enhanced generalization capabilities on remote sensing data.

Conclusions:

• SCU-Net effectively extracts detailed information from remote sensing images, outperforming existing semantic segmentation methods.
• The proposed CDeConv module and channel attention mechanism contribute to improved accuracy and generalization.
• SCU-Net represents a significant advancement in semantic segmentation for remote sensing applications.