Urban scene segmentation model based on multi-scale shuffle features.

Area of Science:

• Remote Sensing
• Urban Planning
• Computer Vision

Background:

• Effective urban land category monitoring is vital for resource management and planning.
• Urban remote sensing faces challenges like uneven parcel distribution, feature extraction difficulties, and information loss.

Purpose of the Study:

• To propose a multi-scale feature shuffle urban scene segmentation model.
• To enhance the accuracy and completeness of urban scene segmentation for improved land management.

Main Methods:

• Utilized a deep convolutional encoder-decoder network with BlurPool to address translation invariance.
• Incorporated GSSConv and SE modules to improve information interaction and filter redundant data.
• Applied multi-scale attention to integrate boundary and global context information for clearer feature extraction.

Main Results:

• The proposed model achieved high performance on the BDCI2017 dataset.
• Achieved scores of 83.1% (OA), 71.0% (mIoU), 82.7% (mRecall), 82.7% (P), and 82.5% (Dice).
• Outperformed several established segmentation networks in key metrics.

Conclusions:

• The multi-scale feature shuffle model significantly improves urban scene segmentation accuracy and completeness.
• This advancement aids in understanding urban development and informs future planning strategies.